³ Here is an excerpt from Will Durant's essay on Immanuel Kant in The Story of Philosophy [84] focusing on what Kant called the "Transcendental Dialectic." Durant's exposition is clearer in my estimation than Kant's attempt in Critique of Pure Reason, but still disturbingly vague and confused for my taste:

Nevertheless, this certainty, this absoluteness, of the highest generalizations of logic and science, is, paradoxically, limited and relative: limited strictly to the field of actual experience, and relative strictly to our human mode of experience. For if our analysis has been correct, the world as we know it is a construction, a finished product, almost — one might say — a manufactured article, to which the mind contributes as much by its moulding forms as the thing contributes by its stimuli. (So we perceive the top of the table as round, whereas our sensation is of an ellipse.) The object as it appears to us is a phenomenon, an appearance, perhaps very different from the external object before it came within the ken of our senses; what that original object was we can never know; the "thing-in-itself" may be an object of thought or inference (a "nouraenon"), but it cannot be experienced, for in being experienced it would be changed by its passage through sense and thought. "It remains completely unknown to us what objects may be by themselves and apart from the receptivity of our senses. We know nothing but our manner of perceiving them; that manner being peculiar to us, and not necessarily shared by every being, though, no doubt, by every human being."

The moon as known to us is merely a bundle of sensations (as Hume saw), unified (as Hume did not see) by our native mental structure through the elaboration of sensations into perceptions, and of these into conceptions or ideas; in result, the moon is for us merely our ideas. Not that Kant ever doubts the existence of "matter" and the external world; but he adds that we know nothing certain about them except that they exist. Our detailed knowledge is about their appearance, their phenomena, about the sensations which we have of them. Idealism does not mean, as the man in the street thinks, that nothing exists outside the perceiving subject; but that a goodly part of every object is created by the forms of perception and understanding: we know the object as transformed into idea; what it is before being so transformed we cannot know. Science, after all, is naive; it supposes tliat it is dealing with things in themselves, in their full-blooded external and uncorrupted reality; philosophy is a little more sophisticated, and realizes that the whole material of science consists of sensations, perceptions and conceptions, rather than of things. "Kant's greatest merit" says Schopenhauer, "is the distinction of the phenomenon from the thing-in-itself."

It follows that any attempt, by either science or religion, to say just what the ultimate reality is, must fall back into mere hypothesis; "the understanding can never go beyond the limits of sensibility." — Such transcendental science loses itself in antinomies," and such transcendental theology loses itself in "paralogisms." It is the cruel function of "transcendental dialectic" to examine the validity of these attempts of reason to escape from the enclosing circle of sensation and appearance into the unknowable world of things "in themselves."

Antinomies are the insoluble dilemmas born of a science that tries to overleap experience. So, for example, knowledge attempts to decide whether the world is finite or infinite in space, thought rebels against either supposition: beyond any limit, we are driven to conceive something further, endlessly ; and yet infinity is itself inconceivable. Again, did the world have a beginning in time? We cannot conceive eternity, but then, too, we cannot conceive any point in the past without feeling at once that before that, something was. Or has that chain of causes which science studies, a beginning, a First Cause? Yes, for an endless chain is inconceivable. No, for a first cause uncaused is inconceivable as well. Is there any exit from these blind alleys of thought? There is, says Kant, if we remember that space, time and cause are modes of perception and conception, which must enter into all our experience, since they are the web and structure of experience; these dilemmas arise from supposing that space, time and cause are external things independent of perception. We shall never have any experience which we shall not interpret in terms of space and time and cause; but we shall never have any philosophy if we forget that these are not things but modes of interpretation and understanding.

So with the paralogisms of "rational" theology—which attempts to prove by theoretical reason that the soul is an incorruptible substance, that the will is free and above the law of cause and effect, and that there exists a 'necessary being' God, as the presupposition of all reality. Transcendental dialectic must remind theology that substance and cause and necessity are finite categories, modes of arrangement and classification which the mind applies to sense-experience, and reliably valid only for the phenomena that appear to such experience; we cannot apply these conceptions to the noumenal (or merely inferred and conjectural) world. Religion cannot be proved by theoretical reason. From [84]

² I'm indebted to my wife Jo for her suggestion that I would be interested in Will Durant's [84] interpretation of Immanuel Kant's Critique of Pure Reason³ .

¹ Approximately 40 years after first reading Immanuel Kant's Critique of Pure Reason, I finally understand his contributions to the history of philosophy and perhaps his confusion with or misinterpretation of David Hume regarding the role of experience and the relationship between experience on the one hand and the ability to construct theories on the other. The latter being the part that confused me at the age of 19 and led me to believe that Kant was confused as well. His idea of (Platonic) theories of unassailable truth — what he calls Transcendental Logic — is at the center of his account of how (a) our direct experience, (b) the patterns we discover in data, and (c) the theories we invent to reinterpret these patterns in the language of mathematics enable us to make predictions about events in the real world².

Human beings and modern machine learning systems are good at discovering patterns in data. Pattern matching was an important step in understanding our environment, but the data was unpredictable, the patterns inconstant and, lacking an understanding of probability, the patterns themselves were unsatisfactory for making accurate predictions and difficult to compose in order to construct broadly encompassing theories that account for more complicated patterns and statistical regularities. Humans invented logic and mathematics to provide structure to those patterns. We invented the differential calculus to explain processes that take place over time. We invented probability and statistics to account for the variability we observe in complex patterns of the sort that govern games of chance. See here for an account of Kant's Theory of Perception.

I don't think Kant understood David Hume's theory of how we come to convert raw experience into theories or perhaps I read too much into Hume. Hume and his fellow empiricists, John Locke and Francis Bacon, were wary of putting too much stock in direct experience and wanted to avoid the human tendency toward superstition. They provided the foundations for the modern scientific method to ground perception in theory. However, one can imagine what the rationalist Kant might make of Hume's view that that passion rather than reason governs human behavior. Hume argued against the existence of innate ideas, positing that all human knowledge is ultimately grounded solely in experience. Hume held that genuine knowledge must either be directly traceable to objects perceived in experience, or result from abstract reasoning about relations between ideas which are derived from experience. Their differences seem largely due to ambiguous terminology. Chronology: Bacon (1561-1626), Locke (1632-1704), Newton (1642-1727), Leibniz (1646-1716), Hume (1711-1776) and Kant (1724-1804).

⁴ A continuous-time recurrent neural network can be modeled as a system of ordinary differential equations. By the Nyquist-Shannon sampling theorem, discrete-time recurrent neural networks of the sort commonly used in machine learning can be viewed as continuous-time recurrent neural networks where the differential equations have been transformed into equivalent difference equations.

⁶ Here is a small sample of journal and encyclopedia articles on episodic memory with an emphasis on its role in autonoetic consciousness in both humans and selected animal species:

@article{GardinerPTRS_B-01,
       author = {Gardiner, J. M.},
        title = {Episodic memory and autonoetic consciousness: a first-person approach.},
      journal = {Philosophical Transactions of the Royal Society London B Biological Science},
         year = {2001},
       volume = {356},
        issue = {1413},
        pages = {1351-1361},
     abstract = {Episodic memory is identified with autonoetic consciousness, which gives rise to remembering in the sense of self-recollection in the mental re-enactment of previous events at which one was present. Autonoetic consciousness is distinguished from noetic consciousness, which gives rise to awareness of the past that is limited to feelings of familiarity or knowing. Noetic consciousness is identified not with episodic but with semantic memory, which involves general knowledge. A recently developed approach to episodic memory makes use of 'first-person' reports of remembering and knowing. Studies using this approach have revealed many independent variables that selectively affect remembering and others that selectively affect knowing. These studies can also be interpreted in terms of distinctiveness and fluency of processing. Remembering and knowing do not correspond with degrees of confidence in memory. Nor does remembering always control the memory response. There is evidence that remembering is selectively impaired in various populations, including not only amnesic patients and older adults but also adults with Asperger's syndrome. This first-person approach to episodic memory represents one way in which that most elusive aspect of consciousness, its subjectivity, can be investigated scientifically. The two kinds of conscious experiences can be manipulated experimentally in ways that are systematic, replicable and intelligible theoretically.},
}
@incollection{ClaytonandDickinsonEAB-10,
       author = {Mental Time Travel: Can Animals Recall the Past and Plan for the Future?},
        title = {N. S. Clayton and A. Dickinson},
       editor = {Breed, Michael D.  and Moore, Janice},
    booktitle = {Encyclopedia of Animal Behavior},
    publisher = {Academic Press},
         year = {2010},
        pages = {438-442},
     abstract = {According to the mental time travel hypothesis, only humans can mentally dissociate themselves from the present, traveling backward in time to recollect specific past events about what happened where and when (episodic memory) and traveling forward in time to anticipate future needs (future planning). A series of studies of the mnemonic capabilities of food-caching western scrub-jays question this assumption. In terms of the retrospective component of episodic memory, these birds remember the ‘what, where, and when’ of specific past caching episodes; they keep track of how long ago they cached different types of perishable foods that decay at different rates, and also remember whether another individual was present at the time of caching, and if so, which bird was watching when. Recent work demonstrates that the jays also make provision for a future need, caching more food in places in which they will not be given breakfast the next morning than in places where they will be receive breakfast the next morning even though there is plenty of food available to them at the time when they cache the food. Taken together these results challenge the mental time travel hypothesis by showing that some elements of both retrospective and prospective mental time travel appear not to be uniquely human.}
}
@incollection{MarinoEAB-10,
       author = {Lori Marino},
        title = {Sentience},
       editor = {Breed, Michael D.  and Moore, Janice},
    booktitle = {Encyclopedia of Animal Behavior},
    publisher = {Academic Press},
         year = {2010},
        pages = {132-138},
     abstract = {Sentience refers to the depth of awareness an individual possesses about himself or herself and others. There appear to be three related, but separable, general domains of sentience. These are self-awareness, metacognition, and theory of mind. To date, evidence shows that these three capacities are found in nonhuman animals, including primates, dolphins, dogs, rodents, and corvids. These findings are evidence of the deep psychological continuity that exists across the animal kingdom.}
}
@article{KleinQJEO-16,
       author = {Klein, S.B.},
        title = {Autonoetic consciousness: Reconsidering the role of episodic memory in future-oriented self-projection},
      journal = {Quarterly Journal of Experimental Psychology},
         year = {2016},
       volume = {69},
       number = {2},
        pages = {381-401},
     abstract = {Following the seminal work of Ingvar (1985. "Memory for the future": An essay on the temporal organization of conscious awareness. Human Neurobiology, 4, 127-136), Suddendorf (1994. The discovery of the fourth dimension: Mental time travel and human evolution. Master's thesis. University of Waikato, Hamilton, New Zealand), and Tulving (1985. Memory and consciousness. Canadian Psychology, 26, 1-12), exploration of the ability to anticipate and prepare for future contingencies that cannot be known with certainty has grown into a thriving research enterprise. A fundamental tenet of this line of inquiry is that future-oriented mental time travel, in most of its presentations, is underwritten by a property or an extension of episodic recollection. However, a careful conceptual analysis of exactly how episodic memory functions in this capacity has yet to be undertaken. In this paper I conduct such an analysis. Based on conceptual, phenomenological, and empirical considerations, I conclude that the autonoetic component of episodic memory, not episodic memory per se, is the causally determinative factor enabling an individual to project him or herself into a personal future.}
}
@article{SprengFiP-13,
       author = {Spreng, R. Nathan},
        title = {Examining the role of memory in social cognition},
      journal = {Frontiers in Psychology},
       volume = {4},
        pages = {437},
         year = {2013},
     abstract = {The function of memory is not only to recall the past, but also to form and update models of our experiences and use these models to navigate the world. Perhaps, the most complex environment for humans to navigate is the social one. Social dynamics are extraordinarily complex, unstructured, labile and difficult to predict. Successful navigation through our many social landscapes is essential to forming and maintaining the durable social bonds necessary for physical and mental health. Until recently, little research has examined the role that memory plays in social behavior and interpersonal sensitivity. There is growing evidence that recalling personally experienced events (autobiographical memory) and inferring the mental states of others (mentalizing or theory-of-mind) share an extensive functional neuroanatomy (Buckner and Carroll, 2007; Spreng et al., 2009; Spreng and Grady, 2010; Rabin et al., 2010) and may be critical for adaptive social cognition.},
}
@article{CiaramelliFiP-13,
       author = {Ciaramelli, Elisa and Bernardi, Francesco and Moscovitch, Morris},
        title = {Individualized Theory of Mind (iToM): When Memory Modulates Empathy},
      journal = {Frontiers in Psychology},
       volume = {4},
        pages = {4},
         year = {2013},
     abstract = {Functional neuroimaging studies have noted that brain regions supporting theory of mind (ToM) overlap remarkably with those underlying episodic memory, suggesting a link between the two processes. The present study shows that memory for others’ past experiences modulates significantly our appraisal of, and reaction to, what is happening to them currently. Participants read the life story of two characters; one had experienced a long series of love-related failures, the other a long series of work-related failures. In a later faux pas recognition task, participants reported more empathy for the character unlucky in love in love-related faux pas scenarios, and for the character unlucky at work in work-related faux pas scenarios. The memory-based modulation of empathy correlated with the number of details remembered from the characters’ life story. These results suggest that individuals use memory for other people’s past experiences to simulate how they feel in similar situations they are currently facing. The integration of ToM and memory processes allows adjusting mental state inferences to fit unique social targets, constructing an individualized ToM (iToM).}
}
@article{BehrendtFiP-13,
       author = {Behrendt, Ralf-Peter},
        title = {Conscious Experience and Episodic Memory: Hippocampus at the Crossroads},
      journal = {Frontiers in Psychology},
       volume = {4},
        pages = {304},
         year = {2013},
     abstract = {If an instance of conscious experience of the seemingly objective world around us could be regarded as a newly formed event memory, much as an instance of mental imagery has the content of a retrieved event memory, and if, therefore, the stream of conscious experience could be seen as evidence for ongoing formation of event memories that are linked into episodic memory sequences, then unitary conscious experience could be defined as a symbolic representation of the pattern of hippocampal neuronal firing that encodes an event memory – a theoretical stance that may shed light into the mind-body and binding problems in consciousness research. Exceedingly detailed symbols that describe patterns of activity rapidly self-organizing, at each cycle of the θ rhythm, in the hippocampus are instances of unitary conscious experience that jointly constitute the stream of consciousness. Integrating object information (derived from the ventral visual stream and orbitofrontal cortex) with contextual emotional information (from the anterior insula) and spatial environmental information (from the dorsal visual stream), the hippocampus rapidly forms event codes that have the informational content of objects embedded in an emotional and spatiotemporally extending context. Event codes, formed in the CA3-dentate network for the purpose of their memorization, are not only contextualized but also allocentric representations, similarly to conscious experiences of events and objects situated in a seemingly objective and observer-independent framework of phenomenal space and time. Conscious perception is likely to be related to more fleeting and seemingly internal forms of conscious experience, such as autobiographical memory recall, mental imagery, including goal anticipation, and to other forms of externalized conscious experience, namely dreaming and hallucinations; and evidence pointing to an important contribution of the hippocampus to these conscious phenomena will be reviewed.}
}

⁵ A reasonable working definition of autonoetic consciousness is supplied below and a small sample of relevant publications is available in this footnote⁶:

"Autonoetic consciousness is the capacity to recursively introspect on one's own subjective experience through time, that is, to perceive the continuity in one's identity from the past to the present and into the future." From [196]

"Autonoetic consciousness is distinguished from noetic consciousness, which gives rise to awareness of the past that is limited to feelings of familiarity or knowing. Noetic consciousness is identified not with episodic but with semantic memory, which involves general knowledge." From [96]

⁷ Executive functions (collectively referred to as executive function and cognitive control) are a set of cognitive processes that are necessary for the cognitive control of behavior: selecting and successfully monitoring behaviors that facilitate the attainment of chosen goals. Executive functions include basic cognitive processes such as attentional control, cognitive inhibition, inhibitory control, working memory, and cognitive flexibility. Higher order executive functions require the simultaneous use of multiple basic executive functions and include planning and fluid intelligence, i.e., reasoning and problem solving. SOURCE

⁸ Here are a couple of recent papers addressing the problem of learning compositional models of visual data:

@article{DonahueetalCoRR-14,
       author = {Jeff Donahue and Lisa Anne Hendricks and Sergio Guadarrama and Marcus Rohrbach and Subhashini Venugopalan and Kate Saenko and Trevor Darrell},
        title = {Long-term Recurrent Convolutional Networks for Visual Recognition and Description},
      journal = {CoRR},
       volume = {arXiv:1411.4389},
         year = 2014,
     abstract = {Models based on deep convolutional networks have dominated recent image interpretation tasks; we investigate whether models which are also recurrent, or "temporally deep", are effective for tasks involving sequences, visual and otherwise. We develop a novel recurrent convolutional architecture suitable for large-scale visual learning which is end-to-end trainable, and demonstrate the value of these models on benchmark video recognition tasks, image description and retrieval problems, and video narration challenges. In contrast to current models which assume a fixed spatio-temporal receptive field or simple temporal averag- ing for sequential processing, recurrent convolutional models are "doubly deep" in that they can be compositional in spatial and temporal "layers". Such models may have advantages when target concepts are complex and/or training data are limited. Learning long-term dependencies is possible when nonlinearities are incorporated into the network state updates. Long-term RNN models are appealing in that they directly can map variable-length inputs (e.g., video frames) to variable length outputs (e.g., natural lan- guage text) and can model complex temporal dynamics; yet they can be optimized with backpropagation. Our recurrent long-term models are directly connected to modern visual convnet models and can be jointly trained to simultaneously learn temporal dynamics and convolutional perceptual rep- resentations. Our results show such models have distinct advantages over state-of-the-art models for recognition or generation which are separately defined and/or optimized.}
}
@article{ChangetalCoRR-16,
       author = {Michael B. Chang and Tomer Ullman and Antonio Torralba and Joshua B. Tenenbaum},
        title = {A Compositional Object-Based Approach to Learning Physical Dynamics},
      journal = {CoRR},
       volume = {arXiv:1612.00341},
         year = {2016},
     abstract = {We present the Neural Physics Engine (NPE), a framework for learning simulators of intuitive physics that naturally generalize across variable object count and different scene configurations. We propose a factorization of a physical scene into composable object-based representations and a neural network architecture whose compositional structure factorizes object dynamics into pairwise interactions. Like a symbolic physics engine, the NPE is endowed with generic notions of objects and their interactions; realized as a neural network, it can be trained via stochastic gradient descent to adapt to specific object properties and dynamics of different worlds. We evaluate the efficacy of our approach on simple rigid body dynamics in two-dimensional worlds. By comparing to less structured architectures, we show that the NPE's compositional representation of the structure in physical interactions improves its ability to predict movement, generalize across variable object count and different scene configurations, and infer latent properties of objects such as mass.},
}
@article{HigginsetalICLR-18,
        title = {{SCAN}: Learning Hierarchical Compositional Visual Concepts},
       author = {Irina Higgins, Nicolas Sonnerat, Loic Matthey, Arka Pal, Christopher P Burgess, Matko Bošnjak, Murray Shanahan, Matthew Botvinick, Demis Hassabis, Alexander Lerchner},
      journal = {International Conference on Learning Representations},
         year = {2018},
     abstract = {The seemingly infinite diversity of the natural world arises from a relatively small set of coherent rules, such as the laws of physics or chemistry. We conjecture that these rules give rise to regularities that can be discovered through primarily unsupervised experiences and represented as abstract concepts. If such representations are compositional and hierarchical, they can be recombined into an exponentially large set of new concepts. This paper describes SCAN (Symbol-Concept Association Network), a new framework for learning such abstractions in the visual domain. SCAN learns concepts through fast symbol association, grounding them in disentangled visual primitives that are discovered in an unsupervised manner. Unlike state of the art multimodal generative model baselines, our approach requires very few pairings between symbols and images and makes no assumptions about the form of symbol representations. Once trained, SCAN is capable of multimodal bi-directional inference, generating a diverse set of image samples from symbolic descriptions and vice versa. It also allows for traversal and manipulation of the implicit hierarchy of visual concepts through symbolic instructions and learnt logical recombination operations. Such manipulations enable SCAN to break away from its training data distribution and imagine novel visual concepts through symbolically instructed recombination of previously learnt concepts.}
}
@article{JohnsonetalCoRR-16,
       author = {Justin Johnson and Bharath Hariharan and Laurens van der Maaten and Li Fei{-}Fei and C. Lawrence Zitnick and Ross B. Girshick},
        title = {{CLEVR:} {A} Diagnostic Dataset for Compositional Language and Elementary Visual Reasoning},
      journal = {CoRR},
       volume = {arXiv:1612.06890},
         year = {2016},
     abstract = {When building artificial intelligence systems that can reason and answer questions about visual data, we need diagnostic tests to analyze our progress and discover shortcomings. Existing benchmarks for visual question answering can help, but have strong biases that models can exploit to correctly answer questions without reasoning. They also conflate multiple sources of error, making it hard to pinpoint model weaknesses. We present a diagnostic dataset that tests a range of visual reasoning abilities. It contains minimal biases and has detailed annotations describing the kind of reasoning each question requires. We use this dataset to analyze a variety of modern visual reasoning systems, providing novel insights into their abilities and limitations.}
}
@misc{JohnsonGOOG-18,
       author = {Justin Johnson},
        title = {Compositional Visual Intelligence through Language},
     abstract = {The use of deep neural networks has led to fantastic recent progress on fundamental problems in computer vision. However the most well-studied tasks, such as image classification and object detection, encompass only a small fraction of the rich visual intelligence that people display. In order to make progress toward the goal of visual intelligence, we must define new tasks and construct new datasets that push the boundaries of artificial visual intelligence. An important facet of visual intelligence is composition - our understanding of the whole derives from an understanding of the parts. One avenue for studying the compositional nature of visual intelligence is through connections with natural language, which is also inherently compositional. To this end I will present three research directions toward compositional visual intelligence through language. In each we will see how incorporating compositionality forces us to rethink our tasks and datasets, but results in systems with richer visual intelligence. I will first discuss image captioning, where moving from sentence captions to dense captions and descriptive paragraphs results in richer image descriptions. I will next discuss visual question answering, where an emphasis on compositional reasoning gives rise to new datasets and models. I will then discuss text-to-image synthesis, where replacing freeform natural language with explicitly compositional scene graphs of objects and relationships allows us to generate more complex images. I will conclude by discussing future directions for improving visual intelligence through composition.}
}

⁹ The neocortex is the newest part of the cerebral cortex to evolve — the prefix "neo" meaning new. The other part of the cerebral cortex is called the allocortex. The cellular organization of the allocortex is different from the six-layered neocortex. In humans, 90% of the cerebral cortex and 76% of the entire brain is neocortex.

¹⁰ David Mayfield owns Blue Ridge Life Science an investment firm specializing in early and mid-stage neuroscience. His grandfather Dr. Frank Mayfield was a famous neurosurgeon, scientific leader and entrepreneur in medical technology whose invention of the spring aneurysm clip has saved thousands of lives. David has taken another route to improving the lives of patients with debilitating neural disorders, driven in part by tragedies within his extended family, but he is no less passionate about the science than his grandfather. Both David and I are dismayed with the way in which current financial, medical and scientific incentives are misaligned. In his life and work, he has attempted to redress the negative consequences of this misalignment, often by drawing attention to relevant science and championing new opportunities for intervention. His crusade to promulgate research on microglia and related drug development is an example of the former. Full disclosure: Dr. Mayfield was my uncle, father figure and mentor at a crucial stage in my young life following the sudden death of my father, and so David, colloquially speaking, is my nephew, or, genealogically more precise, my cousin once removed.

¹¹ There is a case to be made that funding through conventional life science VCs, or even midsize biotechs with bigger bank accounts, won't provide new experimental drugs (or their microglial targets) with the chance to succeed. The problem relates paradoxically to the experimental data which seem to show that some of these drugs are curative in such a wide range of CNS diseases and maladies, e.g., multiple sclerosis, anxiety, drug withdrawal syndromes, stroke, chronic neuropathic pain, retinal degeneration. An embarrassment of riches of sorts which is disabling for VCs and also for midsize biotechs who want their drug candidates focused on very narrow mechanisms of action, and very narrowly defined indications. But what if the embarrassment of riches were explained by the drug's impact on a pathological mechanism broadly shared by much of neurodevelopmental and neurodegenerative disease?

It turns out that doesn't matter. Even a potent and very successful biotech such as Biogen would rather have one drug which mitigates the severity of one orphan disease, than one drug which may prevent, mitigate the severity of, and possibly cure ten diseases / disorders afflicting hundreds of millions. Something to do, perhaps, with incentives, liquidity preferences, and appetite for risk built into the way VCs are funded and the way biotechs are publicly financed? One theory is that the phenomenon also relates to a confusion of the scientific methods of drug discovery with the biology of disease and its causes. Anyway, that's just a long winded way of saying that it is going to take a creative, non-conventional organization to translate the new science of microglia into therapies helping patients.

¹³ Chemokines are a family of small cytokines, or signaling proteins secreted by cells. Their name is derived from their ability to induce directed chemotaxis in nearby responsive cells; they are chemotactic cytokines. [...] Some chemokines are considered pro-inflammatory and can be induced during an immune response to recruit cells of the immune system to a site of infection, while others are considered homeostatic and are involved in controlling the migration of cells during normal processes of tissue maintenance or development. [SOURCE]

¹⁴ Cytokines are a broad and loose category of small proteins that are important in cell signaling. [...] Cytokines may include chemokines, interferons, interleukins, lymphokines, and tumour necrosis factors but generally not hormones or growth factors (despite some overlap in the terminology). Cytokines are produced by a broad range of cells, including immune cells like macrophages, B lymphocytes, T lymphocytes and mast cells, as well as endothelial cells, fibroblasts, and various stromal cells. [...] They act through receptors, and are especially important in the immune system; cytokines modulate the balance between humoral and cell-based immune responses, and they regulate the maturation, growth, and responsiveness of particular cell populations. Some cytokines enhance or inhibit the action of other cytokines in complex ways. [SOURCE]

¹⁵ Interleukins are a group of cytokines (secreted proteins and signal molecules) that were first seen to be expressed by white blood cells (leukocytes). The function of the immune system depends in a large part on interleukins, and rare deficiencies of a number of them have been described, all featuring autoimmune diseases or immune deficiency. The majority of interleukins are synthesized by helper CD4 T lymphocytes, as well as through monocytes, macrophages, and endothelial cells. They promote the development and differentiation of T and B lymphocytes, and hematopoietic cells. [...] Interleukin receptors on astrocytes in the hippocampus are also known to be involved in the development of spatial memories in mice. [SOURCE]

¹⁶ Microglia are a type of neuroglia (glial cell) located throughout the brain and spinal cord, accounting for 10–15% of all cells within the brain. As the resident macrophage cells, they act as the first and main form of active immune defense in the central nervous system (CNS). Microglia (and other neuroglia including astrocytes) are distributed in large non-overlapping regions throughout the CNS. Microglia are constantly scavenging for plaques, damaged or unnecessary neurons and synapses, and infectious agents. Microglia are extremely sensitive to even small pathological changes in the CNS. This sensitivity is achieved in part by the presence of unique potassium channels that respond to even small changes in extracellular potassium. [...] Microglia originate in the yolk sac during a remarkably restricted embryonal period and continuously renew themselves and persist throughout life. [SOURCE]

¹⁷ A trophic or growth factor is a naturally occurring substance capable of stimulating cellular growth, proliferation, healing, and cellular differentiation. Usually it is a protein or a steroid hormone. Growth factors are important for regulating a variety of cellular processes. [...] Growth factors typically act as signaling molecules between cells. Examples are cytokines and hormones that bind to specific receptors on the surface of their target cells. [...] While growth factor implies a positive effect on cell division, cytokine is a neutral term with respect to whether a molecule affects proliferation. While some cytokines can be growth factors, others have an inhibitory effect on cell growth or proliferation. Some cytokines cause target cells to undergo programmed cell death or apoptosis. [SOURCE]

¹⁸ Excerpts of a recent email message from Akram Sadek on the evolution of consciousness, and quantum computing in the brain as it relates to selection pressure to reduce the expenditure of energy:

AS: I completely agree with you and Professor [Sean] Carrol, on people's attempts to somehow conflate consciousness with quantum theory. Even considering just quantum theory itself, the R, or reduction process doesn't need a conscious observer to occur, as is popularly described sometimes. A physical measurement is all it takes, and it's just a matter of entangling a single quantum state with a great number of different states which leads to what we call 'classical' physics and a classical result (see enclosed). And from the perspective of the brain, I can't see at all how quantum mechanics could possibly explain consciousness or qualia. Quantum mechanics is really just a very simple theory at its heart. These sorts of ideas very much sully things, I agree.

On the other hand, an important feature of all biological systems is their remarkable efficiency. Food is scarce, and natural selection is an extremely powerful mechanism that ensures only the most efficient organisms will thrive. Inefficient solutions to biological problems are rapidly weeded out, in as little as a single generation. This is why organisms are able to do so much, with what amounts to very little in terms of energy and physical resources. The human brain only runs at 10W after all. My advisor at Cambridge, Simon Laughlin, discovered early in his career that neurons optimize their coding efficiency to maximize the transmission of information, whilst minimizing energy expenditure (I may have sent you the enclosed paper way back). We already know from the work on photosynthesis that biological systems will harness whatever physics they need to get the job done as efficiently as possible. If quantum mechanics operates in the brain, this is the context in which it would occur.

Theoretically, quantum information processing can occur without any energy expenditure at all, as it is a purely reversible type of computation. The actual energy expenditure that is needed arises due to the necessary error-correction with fault tolerance. If some quantum information processing scheme could have given brains an advantage in terms of using far less energy, then it wouldn't surprise me at all if it's operative. That is what interests me. Of course, if this is the case, then the brain can no longer be thought of as just a thermodynamical engine running an algorithm (which is just a mathematical object). It must be thought of as a physical object that cannot be understood outside of the physical universe it exists in. Since quantum states cannot be cloned, a brain wouldn't be able to be 'copied', like a neural network or piece of software.

¹⁹ Here is a collection of recent papers describing biologically-plausible back-propagation-like algorithms:

@article{BalduzzietalCoRR-14,
       author = {David Balduzzi and Hastagiri Vanchinathan and Joachim M. Buhmann},
        title = {Kickback cuts Backprop's red-tape: Biologically plausible credit assignment in neural networks},
      journal = {CoRR},
       volume = {arXiv:1411.6191},
         year = {2014},
     abstract = {Error backpropagation is an extremely effective algorithm for assigning credit in artificial neural networks. However, weight updates under Backprop depend on lengthy recursive computations and require separate output and error messages -- features not shared by biological neurons, that are perhaps unnecessary. In this paper, we revisit Backprop and the credit assignment problem. We first decompose Backprop into a collection of interacting learning algorithms; provide regret bounds on the performance of these sub-algorithms; and factorize Backprop's error signals. Using these results, we derive a new credit assignment algorithm for nonparametric regression, Kickback, that is significantly simpler than Backprop. Finally, we provide a sufficient condition for Kickback to follow error gradients, and show that Kickback matches Backprop's performance on real-world regression benchmarks.}
}
@article{BengioetalCoRR-16,
       author = {Yoshua Bengio and Dong{-}Hyun Lee and J{\"{o}}rg Bornschein and Zhouhan Lin},
        title = {Towards Biologically Plausible Deep Learning},
      journal = {CoRR},
       volume = {arXiv:1502.04156},
         year = {2016},
     abstract = {Neuroscientists have long criticised deep learning algorithms as incompatible with current knowledge of neurobiology. We explore more biologically plausible versions of deep representation learning, focusing here mostly on unsupervised learning but developing a learning mechanism that could account for supervised, unsupervised and reinforcement learning. The starting point is that the basic learning rule believed to govern synaptic weight updates (Spike-Timing-Dependent Plasticity) arises out of a simple update rule that makes a lot of sense from a machine learning point of view and can be interpreted as gradient descent on some objective function so long as the neuronal dynamics push firing rates towards better values of the objective function (be it supervised, unsupervised, or reward-driven). The second main idea is that this corresponds to a form of the variational EM algorithm, i.e., with approximate rather than exact posteriors, implemented by neural dynamics. Another contribution of this paper is that the gradients required for updating the hidden states in the above variational interpretation can be estimated using an approximation that only requires propagating activations forward and backward, with pairs of layers learning to form a denoising auto-encoder. Finally, we extend the theory about the probabilistic interpretation of auto-encoders to justify improved sampling schemes based on the generative interpretation of denoising auto-encoders, and we validate all these ideas on generative learning tasks.},
}   
@article{LillicrapetalCoRR-14,
       author = {Timothy P. Lillicrap and Daniel Cownden and Douglas B. Tweed and Colin J. Akerman},
        title = {Random feedback weights support learning in deep neural networks},
      journal = {CoRR},
       volume = {arXiv:1411.0247},
         year = {2014},
     abstract = {The brain processes information through many layers of neurons. This deep architecture is representationally powerful, but it complicates learning by making it hard to identify the responsible neurons when a mistake is made. In machine learning, the backpropagation algorithm assigns blame to a neuron by computing exactly how it contributed to an error. To do this, it multiplies error signals by matrices consisting of all the synaptic weights on the neuron's axon and farther downstream. This operation requires a precisely choreographed transport of synaptic weight information, which is thought to be impossible in the brain. Here we present a surprisingly simple algorithm for deep learning, which assigns blame by multiplying error signals by random synaptic weights. We show that a network can learn to extract useful information from signals sent through these random feedback connections. In essence, the network learns to learn. We demonstrate that this new mechanism performs as quickly and accurately as backpropagation on a variety of problems and describe the principles which underlie its function. Our demonstration provides a plausible basis for how a neuron can be adapted using error signals generated at distal locations in the brain, and thus dispels long-held assumptions about the algorithmic constraints on learning in neural circuits.}
}
@article{LillicrapetalNATURE-COMMUNICATIONS-16,
       author = {Lillicrap, Timothy P. and Cownden, Daniel and Tweed, Douglas B. and Akerman, Colin J.},
        title = {Random synaptic feedback weights support error backpropagation for deep learning},
    publisher = {Nature Publishing Group},
      journal = {Nature Communications},
       volume = {7},
         year = {2016},
        pages = {13276},
     abstract = {The brain processes information through multiple layers of neurons. This deep architecture is representationally powerful, but complicates learning because it is difficult to identify the responsible neurons when a mistake is made. In machine learning, the backpropagation algorithm assigns blame by multiplying error signals with all the synaptic weights on each neuron’s axon and further downstream. However, this involves a precise, symmetric backward connectivity pattern, which is thought to be impossible in the brain. Here we demonstrate that this strong architectural constraint is not required for effective error propagation. We present a surprisingly simple mechanism that assigns blame by multiplying errors by even random synaptic weights. This mechanism can transmit teaching signals across multiple layers of neurons and performs as effectively as backpropagation on a variety of tasks. Our results help reopen questions about how the brain could use error signals and dispel long-held assumptions about algorithmic constraints on learning.}
}
@article{ScellierandBengioCoRR-16a,
       author = {Benjamin Scellier and Yoshua Bengio},
        title = {Towards a Biologically Plausible Backprop},
      journal = {CoRR},
       volume = {arXiv:1602.05179v2},
         year = {2016},
     abstract = {Neuroscientists have long criticised deep learning algorithms as incompatible with current knowledge of neurobiology. We explore more biologically plausible versions of deep representation learning, focusing here mostly on unsupervised learning but developing a learning mechanism that could account for supervised, unsupervised and reinforcement learning. The starting point is that the basic learning rule believed to govern synaptic weight updates (Spike-TimingDependent Plasticity) arises out of a simple update rule that makes a lot of sense from a machine learning point of view and can be interpreted as gradient descent on some objective function so long as the neuronal dynamics push firing rates towards better values of the objective function (be it supervised, unsupervised, or rewarddriven). The second main idea is that this corresponds to a form of the variational EM algorithm, i.e., with approximate rather than exact posteriors, implemented by neural dynamics. Another contribution of this paper is that the gradients required for updating the hidden states in the above variational interpretation can be estimated using an approximation that only requires propagating activations forward and backward, with pairs of layers learning to form a denoising auto-encoder. Finally, we extend the theory about the probabilistic interpretation of auto-encoders to justify improved sampling schemes based on the generative interpretation of denoising auto-encoders, and we validate all these ideas on generative learning tasks.},
}
@article{ScellierandBengioCoRR-16b,
       author = {Benjamin Scellier and Yoshua Bengio},
        title = {Equilibrium Propagation: Bridging the Gap Between Energy-Based Models and Backpropagation},
      journal = {CoRR},
       volume = {arXiv:1602.05179v4},
         year = {2016},
     abstract = {We introduce Equilibrium Propagation (e-prop), a learning algorithm for energy-based models. This algorithm involves only one kind of neural computation both for the first phase (when the prediction is made) and the second phase (after the target is revealed) of training. Contrary to backpropagation in feedforward networks, there is no need for special computation in the second phase of our learning algorithm. Equilibrium Propagation combines features of Contrastive Hebbian Learning and Contrastive Divergence while solving the theoretical issues of both algorithms: the algorithm computes the exact gradient of a well defined objective function. Because the objective function is defined in terms of local perturbations, the second phase of e-prop corresponds to only nudging the first-phase fixed point towards a configuration that has lower cost value. In the case of a multi-layer supervised neural network, the output units are slightly nudged towards their target, and the perturbation introduced at the output layer propagates backward in the network. The theory developed in this paper shows that the signal 'back-propagated' during this second phase actually contains information about the error derivatives, which we use to implement a learning rule proved to perform gradient descent with respect to the objective function. Thus, this work makes it more plausible that a mechanism similar to backpropagation could be implemented by brains.}
}
@article{XuetalTNNLS-17,
       author = {David Xu, Andrew Clappison, Cameron Seth, and Jeff Orchard
        title = {Symmetric Predictive Estimator for Biologically Plausible Neural Learning},
      journal = {IEEE Transactions on Neural Networks and Learning Systems},
       volume = {28},
        issue = {10},
         year = {2017},
     abstract = {In a real brain, the act of perception is a bidirectional process, depending on both feedforward sensory pathways and feedback pathways that carry expectations. We are interested in how such a neural network might emerge from a biologically plausible learning rule. Other neural network learning methods either only apply to feedforward networks, or employ assumptions (such as weight copying) that render them unlikely in a real brain. Predictive estimators (PEs) offer a better solution to this bidirectional learning scenario. However, PEs also depend on weight copying. In this paper, we propose the symmetric PE (SPE), an architecture that can learn both feedforward and feedback connection weights individually using only locally available information. We demonstrate that the SPE can learn complicated mappings without the use of weight copying. The SPE networks also show promise in deeper architectures.},
}

¹² Here are my rough notes including selected exchanges with colleagues concerning the different roles of mycroglia in the developing brain. These notes follow up my on my earlier post here. For your convenience, here are a few common technical terms relating to immunology and cell signaling that cover some of the primary tools employed by microglia in their different computational and immunological roles in the human brain: chemokines¹³, cytokines¹⁴, intereukins¹⁵, microglia¹⁶, trophic factors¹⁷.

DM: Thanks for connecting with your colleague Christof on the project. I've listened to a few of his talks over the last few years and find him, and the work of the Allen Institute totally fascinating. I'm traveling today to NYC, and will try to jot down my thoughts on the plane as to how his thinking can be accommodated within the new science regarding microglia. Here are a few thoughts off the top of my head: The questions posed by someone like Christof, interested, as he is, in understanding the biological correlates of consciousness, don't overlap completely with those posed by folks, like, say, David Hubel, Torsten Wiesel, Carla Shatz, and more recently Tobias Bonhoeffer, who are interested in the biological correlates of learning (and memory). I think consciousness is inherently neuronal and digital-computational and depends on functions which occur at the speeds Christof is interested in; the speed of ionotropic signaling between neurons — presynaptic spikes leading to ePSPs and iPSPs as sodium ions and chloride rush down their concentration gradients. Millisecond time scale.

TLD: It's not clear to me that Christof's comments were specific to consciousness, despite his deep interest in the topic. What occurred to me on reading his note had to do with the nature of the computations performed by microglia in service to shaping the interactions between neurons. What do these neural interlocutors need to know about the neurons whose activity they influence? It would appear that the synaptic alterations they produce require seconds or minutes to carry out. Is that because they are simultaneously sampling the activity of tens, hundreds or thousands of dendrites in order to initiate synaptic alterations periodically carried out on the scale of minutes? Do the computations they perform require sampling at millisecond resolution? Setting aside the possibility of some sort of quantum information processing¹⁸, neural refractory periods suggest 1000Hz as an upper bound on how fast human neurons can possibly fire, but 200Hz is a more reasonable upper bound for most human cortical neurons with auditory cortex neurons a possible exception given they appear able to resolve frequencies up to 1000Hz. Still, that's pretty fast.

DM: But learning depends upon functions that occur at slower time scales, beginning with Hebbian changes to synapse efficiency which depend on metabotropic as opposed to ionotropic speeds. Metabolic signaling rather than electrochemical signaling — and I suspect even here, when we are talking about the earliest and most fundamental elements of learning — NMDA receptor/calcium flux mediated changes to synaptic efficiency by AMPA receptor synthesis and insertion — we are already above a timescale of the second. And we haven't even left the confines of the neuron. Structural plasticity, the experience and activity mediated changes in dendritic spine and axonal collateral densities — intuited by Ramòn y Cajal's, and confirmed definitively with post-2000 two photon microscopy of the living, behaving dendritic spine (Trachtenberg et al, 2002, Tobias Bonhoeffer 2001) takes place at time scales greater than seconds, minutes, hours, days. These structural morphological changes to living/behaving neurons were correlated to LTP/LTD and learning before the 2005 re-discovery of microglia (which depended on the same two photon microscopy). I'll point you to the papers that correlate the activity of microglia to these neuronal changes which are the biological substrate of learning. But for now I just want to make the point, that the timescale of the observed structural changes to neurons map well with the timescale of the observed behavior of microglia as then project and retract their arm-like processes to and from the synaptic connections between neurons.

TLD: Does a microglial cell passively sample and analyze extracellular fluid in the vicinity of each spine in its purview for 30 seconds or so and then decide on an intervention modifying all its spines simultaneously? Or possibly the analysis is more interactive as if the glia and neurons are solving a recurrent combinatorial reconstruction problem jointly and initiating reconstruction upon arriving at a solution. From what I understand, the synapses a given microglial cell is responsible for don't overlap with those of other microglial cells. This combinatorial problem could be fully stationary requiring no knowledge of the history of its neuronal / dendritic charges not already encoded in the state of the neurons / dendrites or made apparent during pre-intervention analytical stage. What is the nature of the intervention? Does it entail a specific cocktail of cell-signalling proteins and trophic factors tailored to each dendrite? And why this repeated cycle of projection and retraction as if its target neurons / dendrites are not fixed but rather determined dynamically, possibly sampled randomly to ensure coverage?

I'm particularly interested in the possibility of exploiting the way in which microglia tile the brain, effectively covering all of the synapses. Apparently, it is not simply a cover in the mathematical sense of a set of sets whose union is the whole set, but specifically it is a set of sets that are disjoint and whose union is the whole set. The sets that comprise the microglial tiling are also local in the sense that the elements of the set are bounded in space by the microglial span and the processes of mature microglia in the CNS tend to be compact. In keeping with the terminology introduced in [60], I now think of these sets as domains and imagine they would serve as modular functional domains as defined in [60]. It would not surprise me if these domains overlap in space despite being disjoint, a conjecture I would like to pursue perhaps with your help or the assistance of one of the microglia experts you've introduced me to.

Computational neuroscientists are interested in how learning is implemented in the brain and machine learning researchers have been trying for years to come up with biologically plausible-version of gradient descent. Several of the recent attempts have used strategies involving random weights. One surprisingly simple algorithm for deep learning assigns credit or blame by multiplying error signals by random synaptic weights. The local computations involve adaptation of a sort microglia could be capable of performing. Lillicrap et al [185] show that a network can learn to extract useful information from signals sent through these random feedback connections. In essence, the network learns to learn. For your convenience, I've collected the titles and abstracts for several recent papers describing biologically-plausible back-propagation-like algorithms and made them available here¹⁹.

DM: Neurosteroids also play an important role in regulating neural activity. Like much of what goes on in the field, the scientists studyng the brain's amazing capacity to manufacture its own pregnane steroids aren't talking to the folks who study microglia, and neither group are really talking to the folks who study learning from a computational neuroscience perspective. Neurosteroids could provide the glue, in a way, which brings these often disjointed disciplines together.

Regarding neurosteroids and microglia, the brain can produce neurosteroids (NS) in a very localized manner in order to modulate brain activity. Perhaps at the scale of the individual neuron. And there is some evidence that microglia, which also act at the scale of the individual dendritic spine of individual neurons, are crucial to regulation of the local production of NS. I'm pretty sure that the upregulation of NS is one of the essential tools through which the brain homeostatically keeps microglia from adopting, in error, a macrophage-like, immune reactive phenotype in the face of immune challenges. NS is upregulated by the brain in the face of immune challenge with an anti-inflammatory effect — helping microglia to remain on task interacting with the synaptic landscape. The anxiolytic-driven potentiation of NS biosynthesis is almost certainly how it modulates the behavior of microglia.

Regarding neurosteroids and computational neuroscience: But the brain also clearly upregulates NS so as to modulate the balance of neuronal excitation and inhibition. Allopregnanolone, for instance, is an incredibly potent (nano-molar potency) and efficient positive allosteric modulator of the GABA A receptor. To date, this fact has only been of interest to neuroscientists and biotechs (like SAGE Therapeutics) because they can dose an animal systemically with Allo to achieve profound anti-epileptic, and anxiolytic effect. But the brain itself doesn't make NS for the purpose of depressing excitation globally. In the context of excitation / inhibition, the brain produces NS to modulate specific circuits, potentiating the synaptic weight of specific interneurons relative to the weight of excitatory neurons synapting to the same post synaptic cell. And this should be of interest to the likes of computational neuroscientists like Markus Meister whose work has proven the importance of interneurons to the predictive coding wired into the brain's circuitry genetically through evolution.

It would be really neat to discover that microglia not only deploy their immunological tools to reshape (by Hebbian rules) the synaptic connections of excitatory presynaptic neurons to their post synaptic partner as new experience dictates (the type of learning described by Bonhoeffer), but also that microglia deploy their endocrine-like neurosteroidal tools to modulate the relative importance of the predictive, expectational, information stored in the strength of inhibitory interneuron synapses connecting to the same post synaptic cell.

²⁰ How Google's early failure to sell itself shows why we can't recognize good ideas [SOURCE]

By Lila MacLellan
January 29, 2018

A newly published interview with David Cheriton, the Stanford University computer science professor known for writing Google's second check, offers a reminder about how bad most gatekeepers are at picking out the winners among new ideas. When the Globe and Mail (paywall) asked Cheriton about the original pitch Sergey Brin and Larry Page made to him in 2004, he responded:

Actually, they wanted advice about licensing their search technology to other companies. I'd seen people try that before. The view I take is, you give birth to this baby technology, and you think somebody would like to adopt it because you think it's beautiful. But it's very hard to get anyone else to adopt your baby. I told them, "You have to raise your baby yourself." They came back some months later, and I don't think they said I was right, but they'd decided to start their own company because nobody was interested in their baby.

Stories like these are irresistible. Maybe it's the schadenfreude that comes with imagining Silicon Valley investors missing out on a product that would come to be worth billions. Or it's just that we like to believe that had we been sitting in the room when Page and Brin demonstrated their ranking algorithm, we would have recognized their search technology as genius, history-changing genius.

But studies show that's just wishful thinking. Had you been an executive at, say, Yahoo, one of the companies that was not interested in PageRank, which became Google, you likely would have made the same mistake. Justin Berg, a professor at Stanford University's School of Business, calls it a failure of "creative forecasting," something managers and other decision makers are notorious for. In public lectures on his research, Berg runs through a list of famous rejections: the sitcom Seinfeld, the first Harry Potter book, even the graphical user interface. In literature, there's a multitude of such examples where editors missed a genre-bending or stylistically fresh new voice.

Berg researched how this happens, finding a rich body of literature explaining the natural biases against creativity and innovation that exist in most of us, and are not completely irrational. Novel ideas actually make us "deeply uncomfortable," scholars have found, even when we believe we welcome them. Our inclination to avoid risk means managers are unlikely to endorse an idea that doesn't resemble any that came before it. It's safer not to jeopardize your reputation as someone who can be trusted with the company's time and money.

But Berg wanted to answer a different question about creativity within organizations. He wanted to find out who was good at spotting strong, imaginative ideas, and why, so he began by studying artists and circus managers at Cirque Du Soleil, analyzing their predictions about which new acts introduced by performers would be most appreciated by audiences. Berg found that the artists, and not the managers, had a better sense of when they were watching a future hit. As Adam Grant explained in an Aspens Ideas speech that covered Berg's findings, fellow artists "are much more likely to say, 'What are the reasons that I should consider this idea?' as opposed to, 'Why should I walk away from it?'"

Sadly for managers, it seems there's something about the role itself that can close the mind. This is true even for students asked to think like managers, Berg found in lab studies. But there's also famous proof of this in an anecdote about another early Google investor, Jeff Bezos, founder of Amazon. When he visited Harvard Business School in 1997, he was told by the students in a class called "Managing the Marketspace" that he ought to sell Amazon to Barnes and Noble. In his book, The Everything Store, Brad Stone explains that the Harvard students believed Amazon would be squashed by more established brands when they moved online. They were obviously bright students, but also human. They sided with the known.

On a more optimistic note, however, Berg also found that students playing "managers" got better at a creative forecasting task when they took on the role of of "creator," for even five minutes, by spending some time generating new ideas before listening and judging those of peers. In other words, company executives, editors, producers—all the gatekeepers of the world—ought to make research and innovation part of their jobs. Exactly the way professors like Cheriton do.

²¹ By "reasonably easy" I mean that extending the Zinn hierarchical planner to handle a carefully circumscribed subset of programmer's apprentice-related actions, interventions and misunderstandings suitable for simple demos and testing might take a good programmer a few months effort and a great programmer substantially less. Writing such DMS "stubs" in order to carve out a suitably safe demo space is an art but one that can be relatively easily learned.

²² Using training data and examples collected from Google code repositories would very probably preclude any substantive external demos and complicate publication given the proprietary nature and specialized applications of the Google code base.

²³ Suppose you have a stream of items of large and unknown length that you can only iterate over once. Reservoir sampling is an algorithm that randomly chooses an item from this stream such that each item is equally likely to be selected.

²⁴ Here is a small sample of recent titles and references that straddle disciplinary boundaries to offer new insights: Associative Long Short-Term Memory [53], Neural Turing Machines [105], DRAW: A Recurrent Neural Network For Image Generation [119], Recurrent Models of Visual Attention [206], Using Fast Weights to Attend to the Recent Past [10], Learning Model-based Planning From Scratch [223], The Construction System of the Brain [132], Deconstructing Episodic Memory with Construction [131], Imagination-Augmented Agents for Deep Reinforcement Learning [297], The Future of Memory: Remembering, Imagining, and the Brain [254], Human-level Concept Learning Through Probabilistic Program Induction [171], The Consciousness Prior [25], What is Consciousness, and Could Machines Have It? [73], The Attention Schema Theory: A Mechanistic Account of Subjective Awareness [116]

²⁵ In a memorandum issued in 2000, the NIH Biomedical Information Science and Technology Initiative Consortium agreed on the following definition: Computational Biology: The development and application of data-analytical and theoretical methods, mathematical modeling and computational simulation techniques to the study of biological, behavioral, and social systems.

²⁶ John Rader Platt was an American physicist and biophysicist, a professor at the University of Chicago and an advocate of periodically changing scientific fields to keep fresh, gain new perspectives and facilitate the spread of ideas. His early interests were in the field of molecular biophysics and biophysics, in the 1960s he shifted to philosophy of science, vision and perception, and the study of social trends. He was known for his pioneering work on strong inference in the 1960s and his analysis of social science in the 1970s. He wrote that:

I think there are thousands of scientist who would like to change to less crowded and more interesting fields if they thought the change would not be disapproved and if they could plan how to make a living and how to get research support while making the change. I think this would be a good thing. Mobility spreads the skills in a labor market and mobility would spread the skills in science. Kant, Helmholtz and Pasteur all changed fields. Enrico Fermi once said that a scientist should change fields every ten years; that in the first place his ideas were exhausted by then, and he owed it to the younger men in the field to let them advance; and that in the second place his ideas might be of greater value in bringing a fresh viewpoint to a different field. — John Rader Platt [32]

John W. Gardner, the former cabinet secretary and founder of Common Cause, the nonpartisan public-interest lobby for greater political transparency and accountability, first described a career strategy he referred to as "repotting" as a way to stay engaged and innovative. The idea is that a career reboot not only helps prevent managers from staying in one position too long, being lulled into complacency or leadership fatigue, but that it also pushes leaders to keep learning, to see new challenges with a fresh perspective and ultimately find meaningful work that leaves a lasting legacy:

Most of us have potentialities that have never been developed simply because the circumstances of our lives never called them forth. Exploration of the full range of our own potentialities is not something that we can safely leave to the chances of life. It is something to be pursued systematically, or at least avidly, to the end of our days. We should look forward to an endless and unpredictable dialogue between our potentialities and the claims of life — not only the claims we encounter but the claims we invent. And by the potentialities I mean not just skills, but the full range capacities for sensing, wondering, learning, understanding, loving, and aspiring. — John Gardener [97]

²⁷ By academic career, I am referring to the period from 1986 to 2006 during which I was on the faculty of the Computer Science Department at Brown University. During my first ten years at Brown, I focused exclusively on research and teaching. From 1997 until 2002, I served as the chair of Brown's Computer Science Department and simultaneously as Acting Vice President for Computing and Information Services from 2001 until 2002. I then served as the Deputy Provost of Brown University from 2003 to 2005. Once I became involved in administration, the time I had available for research dwindled and, while I continued working with my graduate students, the intensity with which I pursued new research ideas declined. Peter Norvig suggested I spend my sabbatical at Google in 2006 and subsequently invited me to continue at Google full time to pursue research in computational neuroscience for which I will always be grateful to him. While I work full time as a Research Scientist at Google, I spend approximately 20% of my time teaching and advising students at Stanford thanks to Google's enlightened attitude towards collaboration and university relations.

²⁸ Herbert Simon [261] introduced the idea of bounded rationality as an alternative basis for the sort of mathematical modeling that dominated economics during much of the 20th century. It complements the notion of rationality as optimization, which views decision-making as a fully rational process of finding an optimal choice given the information available. Independently of Simon, procedural limitations have also been pointed out by Bayesian authors most notably I.J. Good, who introduced a distinction between what he calls Type I and Type II rationality. Type I rationality is the ideal of Bayesian decision theory, while Type II rationality takes into account the time and cost of analyzing a problem and thus may depart from the Type I ideal. Type II rationality often justifies a compromise between Bayesian and non-Bayesian methods in statistics [150].

²⁹ Here are some excerpts from Mary Shelley's Frankenstein. Shelley (1797–1851) started writing the story when she was 18. The first edition of the novel was published anonymously in London on January 1, 1818, when she was 20. Her name first appeared on the second edition, published in France in 1823. This is the 200th anniversary of the first edition.

The book has been mentioned in several news stories and appeared in book-club lists focusing on the modern Prometheus theme and helping to fuel concerns about rogue artificial intelligence. These excerpts caught my attention for their nuanced view of a young man caught up with the idea of controlling nature and altering the fate of humanity:

Frankenstein (Mary Shelley) — Loc. 343-54

If, instead of this remark, my father had taken the pains to explain to me that the principles of Agrippa had been entirely exploded and that a modern system of science had been introduced which possessed much greater powers than the ancient, because the powers of the latter were chimerical, while those of the former were real and practical, under such circumstances I should certainly have thrown Agrippa aside and have contented my imagination, warmed as it was, by returning with greater ardour to my former studies. It is even possible that the train of my ideas would never have received the fatal impulse that led to my ruin. But the cursory glance my father had taken of my volume by no means assured me that he was acquainted with its contents, and I continued to read with the greatest avidity.

When I returned home my first care was to procure the whole works of this author, and afterwards of Paracelsus and Albertus Magnus. I read and studied the wild fancies of these writers with delight; they appeared to me treasures known to few besides myself. I have described myself as always having been imbued with a fervent longing to penetrate the secrets of nature. In spite of the intense labour and wonderful discoveries of modern philosophers, I always came from my studies discontented and unsatisfied. Sir Isaac Newton is said to have avowed that he felt like a child picking up shells beside the great and unexplored ocean of truth. Those of his successors in each branch of natural philosophy with whom I was acquainted appeared even to my boy's apprehensions as tyros engaged in the same pursuit.

Frankenstein (Mary Shelley) — Loc. 355-63

The most learned philosopher knew little more. He had partially unveiled the face of Nature, but her immortal lineaments were still a wonder and a mystery. He might dissect, anatomize, and give names; but, not to speak of a final cause, causes in their secondary and tertiary grades were utterly unknown to him. I had gazed upon the fortifications and impediments that seemed to keep human beings from entering the citadel of nature, and rashly and ignorantly I had repined. But here were books, and here were men who had penetrated deeper and knew more. I took their word for all that they averred, and I became their disciple.

It may appear strange that such should arise in the eighteenth century; but while I followed the routine of education in the schools of Geneva, I was, to a great degree, self-taught with regard to my favourite studies. My father was not scientific, and I was left to struggle with a child's blindness, added to a student's thirst for knowledge. Under the guidance of my new preceptors I entered with the greatest diligence into the search of the philosopher's stone and the elixir of life; but the latter soon obtained my undivided attention.

Frankenstein (Mary Shelley) — Loc. 366-68

And thus for a time I was occupied by exploded systems, mingling, like an unadept, a thousand contradictory theories and floundering desperately in a very slough of multifarious knowledge, guided by an ardent imagination and childish reasoning, till an accident again changed the current of my ideas.

Frankenstein (Mary Shelley) — Loc. 438-42

I replied in the affirmative. "Every minute," continued M. Krempe with warmth, "every instant that you have wasted on those books is utterly and entirely lost. You have burdened your memory with exploded systems and useless names. Good God! In what desert land have you lived, where no one was kind enough to inform you that these fancies which you have so greedily imbibed are a thousand years old and as musty as they are ancient? I little expected, in this enlightened and scientific age, to find a disciple of Albertus Magnus and Paracelsus. My dear sir, you must begin your studies entirely anew."

Frankenstein (Mary Shelley) — Loc. 464-69

"The ancient teachers of this science," said he, "promised impossibilities and performed nothing. The modern masters promise very little; they know that metals cannot be transmuted and that the elixir of life is a chimera but these philosophers, whose hands seem only made to dabble in dirt, and their eyes to pore over the microscope or crucible, have indeed performed miracles. They penetrate into the recesses of nature and show how she works in her hiding-places. They ascend into the heavens; they have discovered how the blood circulates, and the nature of the air we breathe. They have acquired new and almost unlimited powers; they can command the thunders of heaven, mimic the earthquake, and even mock the invisible world with its own shadows."

Frankenstein (Mary Shelley) — Loc. 546-52

I prepared myself for a multitude of reverses; my operations might be incessantly baffled, and at last my work be imperfect, yet when I considered the improvement which every day takes place in science and mechanics, I was encouraged to hope my present attempts would at least lay the foundations of future success. Nor could I consider the magnitude and complexity of my plan as any argument of its impracticability. It was with these feelings that I began the creation of a human being. As the minuteness of the parts formed a great hindrance to my speed, I resolved, contrary to my first intention, to make the being of a gigantic stature, that is to say, about eight feet in height, and proportionably large. After having formed this determination and having spent some months in successfully collecting and arranging my materials, I began.

Frankenstein (Mary Shelley) — Loc. 569-73

It was a most beautiful season; never did the fields bestow a more plentiful harvest or the vines yield a more luxuriant vintage, but my eyes were insensible to the charms of nature. And the same feelings which made me neglect the scenes around me caused me also to forget those friends who were so many miles absent, and whom I had not seen for so long a time. I knew my silence disquieted them, and I well remembered the words of my father: "I know that while you are pleased with yourself you will think of us with affection, and we shall hear regularly from you. You must pardon me if I regard any interruption in your correspondence as a proof that your other duties are equally neglected."

Frankenstein (Mary Shelley) — Loc. 578-83

A human being in perfection ought always to preserve a calm and peaceful mind and never to allow passion or a transitory desire to disturb his tranquillity. I do not think that the pursuit of knowledge is an exception to this rule. If the study to which you apply yourself has a tendency to weaken your affections and to destroy your taste for those simple pleasures in which no alloy can possibly mix, then that study is certainly unlawful, that is to say, not befitting the human mind. If this rule were always observed; if no man allowed any pursuit whatsoever to interfere with the tranquillity of his domestic affections, Greece had not been enslaved, Caesar would have spared his country, America would have been discovered more gradually, and the empires of Mexico and Peru had not been destroyed.

Frankenstein (Mary Shelley) — Loc. 781-85

A selfish pursuit had cramped and narrowed me, until your gentleness and affection warmed and opened my senses; I became the same happy creature who, a few years ago, loved and beloved by all, had no sorrow or care. When happy, inanimate nature had the power of bestowing on me the most delightful sensations. A serene sky and verdant fields filled me with ecstasy. The present season was indeed divine; the flowers of spring bloomed in the hedges, while those of summer were already in bud. I was undisturbed by thoughts which during the preceding year had pressed upon me, notwithstanding my endeavours to throw them off, with an invincible burden.

³⁰ For the purpose of thinking about episodic memory, the prefrontal cortex is primarily associated with executive function and attention and memory which we cover elsewhere in the research notes. SOURCE

³¹ The temporal lobe (Brodmann area 22) is involved in processing sensory input into derived meanings for the appropriate retention of visual memory, language comprehension, and emotion association. It adjoins Wernnicke's area associated with language comprehension. SOURCE

³² Wernicke's area is classically located in the posterior section of the superior temporal gyrus (STG) in the (most commonly) left cerebral hemisphere. This area encircles the auditory cortex on the lateral sulcus (the part of the brain where the temporal lobe and parietal lobe meet). SOURCE

³³ The retrosplenial cortex (Brodmann area 29 and 30) is located close to visual areas and also to the hippocampal spatial/memory system suggest it may have a role in mediating between perceptual and memory functions. SOURCE

³⁴ The occipital lobe is one of the four major lobes of the cerebral cortex in the brain of mammals. The occipital lobe is the visual processing center of the mammalian brain containing most of the anatomical region of the visual cortex SOURCE

³⁵ The cuneus (Brodmann area 17) receives visual information from the same-sided superior quandrantic retina (corresponding to contralateral inferior visual field). It is most known for its involvement in basic visual processing. The mid-level visual processing that occurs in the extrastriate projection fields of the cuneus are modulated by extraretinal effects, like attention, working memory, and reward expectation. SOURCE

³⁶ The precuneus is located on the inside between the two cerebral hemispheres in the rear region between the somatosensory cortex and forward of the cuneus (which contains the visual cortex). The mental imagery concerning the self has been located in the forward part of the precuneus with posterior areas being involved with episodic memory. Another area has been linked to visuospatial imagery. SOURCE

³⁷ Here is an extended quotation from Conway [48] that, while it doesn't reveal any additional anatomical detail, does provide some useful framing insights from the perspective of cognitive neuroscience:

A distinction might be drawn between a conceptual based frontal-anterior temporal memory system and a posterior temporal-occipital EM system. In the human brain a unique way to access and manipulate EMs is through the conceptual knowledge system. It is this relation between conceptual knowledge and EM that supports two major and fairly amazing abilities: the ability to imagine how the past might have been otherwise and the ability to imagine the future. The ability to imagine how a specific experience might have unfolded differently with different outcomes conveys a huge survival advantage. Different courses of actions can be played out in memory with no physical consequences.

The ability to manipulate memories in this way provides the basis of planning and the projection of alternate forms of the past into the future. It is a powerful way of anticipating possible outcomes. Indeed, it provides a meaning by which long-term goals can be generated and planned for. It is then especially interesting that recent neuroimaging studies comparing brain activations during the recall of autobiographical memories with brain activations during the generation of false but plausible memories has found extensive commonalities between the brain areas active in both.

These commonalities are so marked that it might be more appropriate to refer to a remembering-imaging system rather than simply a "memory system." The remembering-imaging system is a cognitive system in which the various components that are usually constructed into an autobiographical memory (autobiographical knowledge and episodic memory) can be constructed into other mental representations of imaginary scenarios. For example, specific EMs might be constructed into novel conceptual context where "dinner with X" becomes "dinner with Y." or perhaps generic knowledge of locations, actors, and actions are configured into potential EMs of unexperienced events. Recent findings indicate then that EMs and autobiographical knowledge can provide the materials, the content, for the construction of imagined tasks and futures.

³⁸ Here is a short list of books and papers on theory of mind and episodic memory. The Handbook of Episodic Memory is a wonderful reference covering much of what is known about episodic memory. Unlike what is known about how the brain represents physical objects in our environment — the shape, color and motion of visual stimuli are represented in an orderly fashion in the visual association areas, episodic memory, is a great deal more complicated, emerges gradually and develops over many years integrating information from a great many sources. The selected journal articles are but a small sample of what I've read or skimmed, even if I read and understood all of the articles I've skimmed and memorized all 600 plus pages of the handbook, it would be years of work to develop a working model of episodic memory that even approximates human memory.

@book{DereetalHANDBOOK-of-EPISODIC-MEMORY-2008,
       author = {Dere, E. and Easton, A. and Nadel, L. and Huston, J.P.},
        title = {Handbook of Episodic Memory},
    publisher = {Elsevier Science},
       series = {Handbook of Behavioral Neuroscience},
         year = {2008},
}
@article{SaxeetalARP-04,
       author = {R. Saxe and S. Carey and N. Kanwisher},
        title = {Understanding Other Minds: Linking Developmental Psychology and Functional Neuroimaging},
      journal = {Annual Review of Psychology},
       volume = {55},
       number = {1},
         year = {2004},
        pages = {87-124},
     abstract = {Evidence from developmental psychology suggests that understanding other minds constitutes a special domain of cognition with at least two components: an early-developing system for reasoning about goals, perceptions, and emotions, and a later-developing system for representing the contents of beliefs. Neuroimaging reinforces and elaborates upon this view by providing evidence that (a) domain-specific brain regions exist for representing belief contents, (b) these regions are apparently distinct from other regions engaged in reasoning about goals and actions (suggesting that the two developmental stages reflect the emergence of two distinct systems, rather than the elaboration of a single system), and (c) these regions are distinct from brain regions engaged in inhibitory control and in syntactic processing. The clear neural distinction between these processes is evidence that belief attribution is not dependent on either inhibitory control or syntax, but is subserved by a specialized neural system for theory of mind. }
}
@article{DuboisandAdolphsPNAS-16,
       author = {Dubois, Julien and Adolphs, Ralph},
        title = {How the brain represents other minds},
      journal = {Proceedings of the National Academy of Sciences},
       volume = {113},
       number = {1},
         year = {2016},
        pages = {19-21},
     abstract = {How does the brain represent the world? Sensory neuroscience has given us a detailed window into how the brain represents physical objects in our environment: For instance, the shape, color, and direction of motion of visual stimuli are represented in an orderly fashion in higher order visual cortices. But we also represent social objects: other people and their thoughts, beliefs, and feelings. How is that kind of knowledge represented in the brain? In an ambitious new study in PNAS, Tamir et al. used functional MRI (fMRI) to argue that our brains represent other minds along three broad dimensions: social impact, rationality, and valence.}
}
@article{MahyetalDCN-14,
        title = {How and where: Theory-of-mind in the brain},
       author = {Caitlin E.V. Mahy and Louis J. Moses and Jennifer H. Pfeifer},
      journal = {Developmental Cognitive Neuroscience},
       volume = {9},
         year = {2014},
        pages = {68-81},
     abstract = {Theory of mind (ToM) is a core topic in both social neuroscience and developmental psychology, yet theory and data from each field have only minimally constrained thinking in the other. The two fields might be fruitfully integrated, however, if social neuroscientists sought evidence directly relevant to current accounts of ToM development: modularity, simulation, executive, and theory theory accounts. Here we extend the distinct predictions made by each theory to the neural level, describe neuroimaging evidence that in principle would be relevant to testing each account, and discuss such evidence where it exists. We propose that it would be mutually beneficial for both fields if ToM neuroimaging studies focused more on integrating developmental accounts of ToM acquisition with neuroimaging approaches, and suggest ways this might be achieved.}
}
@article{FerreretalFiN-09,
       author = {Ferrer, Emilio and O'Hare, Elizabeth D. and Bunge, Silvia A.},
        title = {Fluid Reasoning and the Developing Brain},
      journal = {Frontiers in Neuroscience},
    publisher = {Frontiers Research Foundation},
       volume = {3},
        issue = {1},
         year = {2009},
        pages = {46-51},
     abstract = {Fluid reasoning is the cornerstone of human cognition, both during development and in adulthood. Despite this, the neural mechanisms underlying the development of fluid reasoning are largely unknown. In this review, we provide an overview of this important cognitive ability, the method of measurement, its changes over the childhood and adolescence of an individual, and its underlying neurobiological underpinnings. We review important findings from psychometric, cognitive, and neuroscientific literatures, and outline important future directions for this interdisciplinary research.},
}
@article{DumontheilDCN-14,
        title = {Development of abstract thinking during childhood and adolescence: The role of rostrolateral prefrontal cortex},
       author = {Iroise Dumontheil},
      journal = {Developmental Cognitive Neuroscience},
       volume = {10},
         year = {2014},
        pages = {57-76},
     abstract = {Rostral prefrontal cortex (RPFC) has increased in size and changed in terms of its cellular organisation during primate evolution. In parallel emerged the ability to detach oneself from the immediate environment to process abstract thoughts and solve problems and to understand other individuals’ thoughts and intentions. Rostrolateral prefrontal cortex (RLPFC) is thought to play an important role in supporting the integration of abstract, often self-generated, thoughts. Thoughts can be temporally abstract and relate to long term goals, or past or future events, or relationally abstract and focus on the relationships between representations rather than simple stimulus features. Behavioural studies have provided evidence of a prolonged development of the cognitive functions associated with RLPFC, in particular logical and relational reasoning, but also episodic memory retrieval and prospective memory. Functional and structural neuroimaging studies provide further support for a prolonged development of RLPFC during adolescence, with some evidence of increased specialisation of RLPFC activation for relational integration and aspects of episodic memory retrieval. Topics for future research will be discussed, such as the role of medial RPFC in processing abstract thoughts in the social domain, the possibility of training abstract thinking in the domain of reasoning, and links to education.}
}

³⁹ Preliminary sample of papers relating to episodic memory and related applications including question answering:

@article{AnonymousICLR-18a,
        title = {Integrating Episodic Memory into a Reinforcement Learning Agent Using Reservoir Sampling},
       author = {Anonymous},
      journal = {International Conference on Learning Representations},
         year = {2018},
          url = {https://openreview.net/forum?id=ByJDAIe0b},
     abstract = {Episodic memory is a psychology term which refers to the ability to recall specific events from the past. We suggest one advantage of this particular type of memory is the ability to easily assign credit to a specific state when remembered information is found to be useful. Inspired by this idea, and the increasing popularity of external memory mechanisms to handle long-term dependencies in deep learning systems, we propose a novel algorithm which uses a reservoir sampling procedure to maintain an external memory consisting of a fixed number of past states. The algorithm allows a deep reinforcement learning agent to learn online to preferentially remember those states which are found to be useful to recall later on. Critically this method allows for efficient online computation of gradient estimates with respect to the write process of the external memory. Thus unlike most prior mechanisms for external memory it is feasible to use in an online reinforcement learning setting.}
}
@article{AnonymousICLR-18b,
        title = {Memory Architectures in Recurrent Neural Network Language Models},
       author = {Anonymous},
      journal = {International Conference on Learning Representations},
         year = {2018},
          url = {https://openreview.net/forum?id=SkFqf0lAZ},
     abstract = {We compare and analyze sequential, random access, and stack memory architectures for recurrent neural network language models. Our experiments on the Penn Treebank and Wikitext-2 datasets show that stack-based memory architectures consistently achieve the best performance in terms of held out perplexity. We also propose a generalization to existing continuous stack models (Joulin & Mikolov,2015; Grefenstette et al., 2015)  to allow a variable number of pop operations more naturally that further improves performance. We further evaluate these language models in terms of their ability to capture non-local syntactic dependencies on a subject-verb agreement dataset  (Linzen et al., 2016) and establish new state of the art results using memory augmented language models. Our results demonstrate the value of stack-structured memory for explaining the distribution of words in natural language, in line with linguistic theories claiming a context-free backbone for natural language.}
}
@misc{DynamicMemoryNetworks-16,
        title = {Dynamic Memory Networks},
       author = {Anonymous},
         year = {2016},
 howpublished = {{LDGN} Notes and Thoughts on Machine Learning and Artificial Intelligence},
          url = {https://ldgn.wordpress.com/2016/01/04/attention-and-memory-in-deep-learning/},
     abstract = {Dynamic Memory Networks (DMN) are a general framework for question answering over inputs. Conceptually, a difference is made between inputs and questions. The DMN takes a sequence of inputs and a question and then employs an iterative attention process to compute the answer. The sequence of inputs can be seen as the history, which complements the general world knowledge (see semantic memory module). The DNM framework consists of five components: (i) input module: processes raw input and maps it to a useful representation, (ii) semantic memory module: stores general knowledge about concepts and facts. It can be instantiated by word embeddings or knowledge bases, (iii) question module: maps a question into a representation, (iv) episodic memory module: an iterative component that in each iteration focuses on different parts of the input, updates its internal state and finally outputs an answer representation (vi) answer module: generates the answer to return.}
}
@article{KumaretalCoRR-16,
       author = {Ankit Kumar and Ozan Irsoy and Jonathan Su and James Bradbury and Robert English and Brian Pierce and Peter Ondruska and Ishaan Gulrajani and Richard Socher},
        title = {Ask Me Anything: Dynamic Memory Networks for Natural Language Processing},
      journal = {CoRR},
       volume = {arXiv:1506.07285},
         year = {2015},
     abstract = {Most tasks in natural language processing can be cast into question answering (QA) problems over language input. We introduce the dynamic memory network (DMN), a neural network architecture which processes input sequences and questions, forms episodic memories, and generates relevant answers. Questions trigger an iterative attention process which allows the model to condition its attention on the inputs and the result of previous iterations. These results are then reasoned over in a hierarchical recurrent sequence model to generate answers. The DMN can be trained end-to-end and obtains state-of-the-art results on several types of tasks and datasets: question answering (Facebook's bAbI dataset), text classification for sentiment analysis (Stanford Sentiment Treebank) and sequence modeling for part-of-speech tagging (WSJ-PTB). The training for these different tasks relies exclusively on trained word vector representations and input-question-answer triplets.}
}
@inproceedings{ChangandTanIJCAI-17,
       author = {Poo-Hee Chang, Ah-Hwee Tan},
        title = {Encoding and Recall of Spatio-Temporal Episodic Memory in Real Time},
    booktitle = {Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence, {IJCAI-17}},
        pages = {1490--1496},
         year = {2017},
     abstract = {Episodic memory enables a cognitive system to improve its performance by reflecting upon past events. In this paper, we propose a computational model called STEM for encoding and recall of episodic events together with the associated con- textual information in real time. Based on a class of self-organizing neural networks, STEM is de- signed to learn memory chunks or cognitive nodes, each encoding a set of co-occurring multi-modal activity patterns across multiple pattern channels. We present algorithms for recall of events based on partial and inexact input patterns. Our empirical results based on a public domain data set show that STEM displays a high level of efficiency and ro- bustness in encoding and retrieval with both partial and noisy search cues when compared with a state- of-the-art associative memory model.},
}
@article{TaniJCS-98,
       author = {Tani, Jun},
        title = {An interpretation of the 'self' from the dynamical systems perspective: {A} constructivist approach},
    booktitle = {Journal of Consciousness Studies},
       volume = {5},
         year = {1998},
        pages = {516-542},
     abstract = {This study attempts to describe the notion of the 'self' using dynamical systems language based on the results of our robot learning experiments. A neural network model consisting of multiple modules is proposed, in which the interactive dynamics between the bottom-up perception and the top-down prediction are investigated. Our experiments with a real mobile robot showed that the incremental learning of the robot switches spontaneously between steady and unsteady phases. In the steady phase, the top-down prediction for the bottom-up perception works well when coherence is achieved between the internal and the environmental dynamics. In the unsteady phase, conflicts arise between the bottom-up perception and the top-down prediction; the coherence is lost, and a chaotic attractor is observed in the internal neural dynamics. By investigating possible analogies between this result and the phenomenological literature on the 'self', we draw the conclusions that (1) the structure of the 'self' corresponds to the 'open dynamic structure' which is characterized by co-existence of stability in terms of goal-directedness and instability caused by embodiment; (2) the open dynamic structure causes the system's spontaneous transition to the unsteady phase where the 'self' becomes aware.}
}
@article{PritzeletalCoRR-17,
       author = {Alexander Pritzel and Benigno Uria and Sriram Srinivasan and Adri{\`{a}} Puigdom{\`{e}}nech Badia and Oriol Vinyals and Demis Hassabis and Daan Wierstra and Charles Blundell},
        title = {Neural Episodic Control},
      journal = {CoRR},
       volume = {arXiv:1703.01988},
         year = {2017},
     abstract = {Deep reinforcement learning methods attain super-human performance in a wide range of environments. Such methods are grossly inefficient, often taking orders of magnitudes more data than humans to achieve reasonable performance. We propose Neural Episodic Control: a deep reinforcement learning agent that is able to rapidly assimilate new experiences and act upon them. Our agent uses a semi-tabular representation of the value function: a buffer of past experience containing slowly changing state representations and rapidly updated estimates of the value function. We show across a wide range of environments that our agent learns significantly faster than other state-of-the-art, general purpose deep reinforcement learning agents.}
}
@book{GallagherMODELS-OF-THE-SELF-13,
        title = {Models of the Self},
       author = {Gallagher, Shaun},
         year = {2013},
    publisher = {Imprint Academic}
}
@book{OKeefeandNadelHIPPOCAMPUS-78,
        title = {The hippocampus as a cognitive map},
       author = {O'Keefe, John and Nadel, Lynn},
         year = {1978},
    publisher = {Clarendon Press}
}

⁴⁰ Here is a sample of papers relating to attention and theory-of-mind reasoning focusing on the separate research agendas of Josh Tenenbaum at MIT and Michael Graziano at Princeton along with their many students and colleagues:

@article{GershmanetalPLoS-16,
       author = {Sam Gershman and Tobias Gerstenberg and Chris Baker and Fiery Cushman},
      journal = {PLoS ONE},
        title = {Plans, habits, and theory of mind},
         year = {2016},
     abstract = {Human success and even survival depends on our ability to predict what others will do by guessing what they are thinking. If I accelerate, will he yield? If I propose, will she accept? If I confess, will they forgive? Psychologists call this capacity "theory of mind." According to current theories, we solve this problem by assuming that others are rational actors. That is, we assume that others design and execute efficient plans to achieve their goals, given their knowledge. But if this view is correct, then our theory of mind is startlingly incomplete. Human action is not always a product of rational planning, and we would be mistaken to always interpret others' behaviors as such. A wealth of evidence indicates that we often act habitually—a form of behavioral control that depends not on rational planning, but rather on a history of reinforcement. We aim to test whether the human theory of mind includes a theory of habitual action and to assess when and how it is deployed. In a series of studies, we show that human theory of mind is sensitive to factors influencing the balance between habitual and planned behavior.},
}
@article{LakeetalCoRR-16,
       author = {Brenden M. Lake and Tomer D. Ullman and Joshua B. Tenenbaum and Samuel J. Gershman},
        title = {Building Machines That Learn and Think Like People},
      journal = {CoRR},
       volume = {arXiv:1604.00289},
         year = {2016},
     abstract = {Recent progress in artificial intelligence (AI) has renewed interest in building systems that learn and think like people. Many advances have come from using deep neural networks trained end-to-end in tasks such as object recognition, video games, and board games, achieving performance that equals or even beats humans in some respects. Despite their biological inspiration and performance achievements, these systems differ from human intelligence in crucial ways. We review progress in cognitive science suggesting that truly human-like learning and thinking machines will have to reach beyond current engineering trends in both what they learn, and how they learn it. Specifically, we argue that these machines should (a) build causal models of the world that support explanation and understanding, rather than merely solving pattern recognition problems; (b) ground learning in intuitive theories of physics and psychology, to support and enrich the knowledge that is learned; and (c) harness compositionality and learning-to-learn to rapidly acquire and generalize knowledge to new tasks and situations. We suggest concrete challenges and promising routes towards these goals that can combine the strengths of recent neural network advances with more structured cognitive models.}
}
@article{StuhlmuellerandGoodmanCS-13,
       author = {Stuhlm\"{u}ller, A. and Goodman, N. D.},
      journal = {Journal of Cognitive Systems Research},
        title = {Reasoning about Reasoning by Nested Conditioning: Modeling Theory of Mind with Probabilistic Programs},
       volume = {28},
         year = {2014},
        pages = {80-99},
     abstract = {A wide range of human reasoning patterns can be explained as conditioning in probabilistic models; however, conditioning has traditionally been viewed as an operation applied to such models, not represented in such models. We describe how probabilistic programs can explicitly represent conditioning as part of a model. This enables us to describe reasoning about others’ reasoning using nested conditioning. Much of human reasoning is about the beliefs, desires, and intentions of other people; we use probabilistic programs to formalize these inferences in a way that captures the flexibility and inherent uncertainty of reasoning about other agents. We express examples from game theory, artificial intelligence, and linguistics as recursive probabilistic programs and illustrate how this representation language makes it easy to explore new directions in each of these fields. We discuss the algorithmic challenges posed by these kinds of models and describe how Dynamic Programming techniques can help address these challenges.}
}
@inbook{Goodmanetal2015,
       author = {Goodman, Noah D. and Tenenbaum, Joshua B. and Gerstenberg, T.},
    booktitle = {The Conceptual Mind: New Directions in the Study of Concepts},
       editor = {Morgolis and Lawrence},
    publisher = {MIT Press},
        title = {Concepts in a probabilistic language of thought},
         year = {2015}
}
@inproceedings{ErikvandenBoogaardetal2017,
       author = {Erik van den Boogaard and Jan Treur and Maxim Turpijn},
        title = {A Neurologically Inspired Network Model for Graziano's Attention Schema Theory for Consciousness},
    booktitle = {Natural and Artificial Computation for Biomedicine and Neuroscience - International Work-Conference on the Interplay Between Natural and Artificial Computation},
        pages = {10-21},
         year = {2017},
     abstract = {This paper describes a network-oriented model based on the neuroscientist Graziano’s Attention Schema Theory for consciousness. This theory describes an attention schema as an internal model of the attention process supporting the control of attention, similar to how our mind uses a body schema as an internal model of the body to control its movements. The Attention Schema Theory comes with a number of testable predictions. After designing a neurologically inspired temporal- causal network model for the Attention schema Theory, a few simula- tions were conducted to verify some of these predictions. One prediction is that a noticeable attention control deficit occurs when using attention without awareness. Another is that a noticeable attention control deficit occurs when using only bottom-up influence (from the sensory repre- sentations) without any top-down influence (for example, from goal or control states). The presented model is illustrated by a scenario where a hunter imagines (using internal simulation) a prey which he wants to attend to and catch, but shortly after he or she imagines a predator which he then wants to attend to and avoid. The outcomes of the simulations support the predictions that were made.}
}
@article{GrazianoFiRAI-17,
       author = {Graziano, Michael S. A.},
        title = {The Attention Schema Theory: A Foundation for Engineering Artificial Consciousness},
      journal = {Frontiers in Robotics and AI},
       volume = {4},
        pages = {60},
         year = {2017},
     abstract = {The purpose of the attention schema theory is to explain how an information-processing device, the brain, arrives at the claim that it possesses a non-physical, subjective awareness, and assigns a high degree of certainty to that extraordinary claim. The theory does not address how the brain might actually possess a non-physical essence. It is not a theory that deals in the non-physical. It is about the computations that cause a machine to make a claim and to assign a high degree of certainty to the claim. The theory is offered as a possible starting point for building artificial consciousness. Given current technology, it should be possible to build a machine that contains a rich internal model of what consciousness is, attributes that property of consciousness to itself and to the people it interacts with, and uses that attribution to make predictions about human behavior. Such a machine would “believe” it is conscious and act like it is conscious, in the same sense that the human machine believes and acts.}
}
@article{IgelstrometalENEURO-16,
       author = {Igelstr\"{o}m, Kajsa M. and Webb, Taylor W. and Kelly, Yin T. and Graziano, Michael S. A.},
        title = {Topographical Organization of Attentional, Social, and Memory Processes in the Human Temporoparietal Cortex},
      journal = {eNeuro},
    publisher = {Society for Neuroscience},
       volume = {3},
       number = {2},
         year = {2016},
     abstract = {The temporoparietal junction (TPJ) is activated in association with a large range of functions, including social cognition, episodic memory retrieval, and attentional reorienting. An ongoing debate is whether the TPJ performs an overarching, domain-general computation, or whether functions reside in domain-specific subdivisions. We scanned subjects with fMRI during five tasks known to activate the TPJ, probing social, attentional, and memory functions, and used data-driven parcellation (independent component analysis) to isolate task-related functional processes in the bilateral TPJ. We found that one dorsal component in the right TPJ, which was connected with the frontoparietal control network, was activated in all of the tasks. Other TPJ subregions were specific for attentional reorienting, oddball target detection, or social attribution of belief. The TPJ components that participated in attentional reorienting and oddball target detection appeared spatially separated, but both were connected with the ventral attention network. The TPJ component that participated in the theory-of-mind task was part of the default-mode network. Further, we found that the BOLD response in the domain-general dorsal component had a longer latency than responses in the domain-specific components, suggesting an involvement in distinct, perhaps postperceptual, computations. These findings suggest that the TPJ performs both domain-general and domain-specific computations that reside within spatially distinct functional components.},
}
@article{GrazianoandWebb-15,
       author = {Graziano, Michael S. A. and Webb, Taylor W.},
        title = {The attention schema theory: a mechanistic account of subjective awareness},
      journal = {Frontiers in Psychology},
       volume = {6},
        pages = {500},
         year = {2015},
     abstract = {We recently proposed the attention schema theory, a novel way to explain the brain basis of subjective awareness in a mechanistic and scientifically testable manner. The theory begins with attention, the process by which signals compete for the brain’s limited computing resources. This internal signal competition is partly under a bottom-up influence and partly under top-down control. We propose that the top-down control of attention is improved when the brain has access to a simplified model of attention itself. The brain therefore constructs a schematic model of the process of attention, the ‘attention schema’, in much the same way that it constructs a schematic model of the body, the ‘body schema’. The content of this internal model leads a brain to conclude that it has a subjective experience. One advantage of this theory is that it explains how awareness and attention can sometimes become dissociated; the brain’s internal models are never perfect, and sometimes a model becomes dissociated from the object being modeled. A second advantage of this theory is that it explains how we can be aware of both internal and external events. The brain can apply attention to many types of information including external sensory information and internal information about emotions and cognitive states. If awareness is a model of attention, then this model should pertain to the same domains of information to which attention pertains. A third advantage of this theory is that it provides testable predictions. If awareness is the internal model of attention, used to help control attention, then without awareness, attention should still be possible but should suffer deficits in control. In this article, we review the existing literature on the relationship between attention and awareness, and suggest that at least some of the predictions of the theory are borne out by the evidence.}
}

⁴¹ The abstract syntax tree of the example in the text is a simplified version of the AST for a WHILE loop:

⁴² Titles and abstracts of papers on prefrontal structures relevant to attention and executive control, including recent research on whether or to what extent working memory is situated in the frontal cortex [172]:

@article{BalletalFiN-11,
       author = {Ball, Gareth and Stokes, Paul R. and Rhodes, Rebecca A. and Bose, Subrata K. and Rezek, Iead and Wink, Alle-Meije and Lord, Louis-David and Mehta, Mitul A. and Grasby, Paul M. and Turkheimer, Federico E.},
        title = {Executive Functions and Prefrontal Cortex: A Matter of Persistence?},
      journal = {Frontiers in Systems Neuroscience},
         year = {2011},
    publisher = {Frontiers Research Foundation},
       volume = {5},
        pages = {3},
     abstract = {Executive function is thought to originates from the dynamics of frontal cortical networks. We examined the dynamic properties of the blood oxygen level dependent time-series measured with functional MRI (fMRI) within the prefrontal cortex (PFC) to test the hypothesis that temporally persistent neural activity underlies performance in three tasks of executive function. A numerical estimate of signal persistence, the Hurst exponent, postulated to represent the coherent firing of cortical networks, was determined and correlated with task performance. Increasing persistence in the lateral PFC was shown to correlate with improved performance during an n-back task. Conversely, we observed a correlation between persistence and increasing commission error --- indicating a failure to inhibit a prepotent response --- during a Go/No-Go task. We propose that persistence within the PFC reflects dynamic network formation and these findings underline the importance of frequency analysis of fMRI time-series in the study of executive functions.},
}
@article{BalstersetalNEUROIMAGE-09,
     title = {Evolution of the cerebellar cortex: {T}he selective expansion of prefrontal-projecting cerebellar lobules},
    author = {Balsters, J. H. and Cussans, E. and Diedrichsen, J. and Phillips, K. A. and Preuss, T. M. and Rilling, J. K. and Ramnani, N.},
   journal = {NeuroImage},
     issue = 3,
     pages = {2045-2052},
    volume = 49,
      year = 2010,
  abstract = {It has been suggested that interconnected brain areas evolve in tandem because evolutionary pressures act on complete functional systems rather than on individual brain areas. The cerebellar cortex has reciprocal connections with both the prefrontal cortex and motor cortex, forming independent loops with each. Specifically, in capuchin monkeys cerebellar cortical lobules Crus I and Crus II connect with prefrontal cortex  whereas the primary motor cortex connects with cerebellar lobules V, VI, VIIb, and VIIIa. Comparisons of extant primate species suggest that the prefrontal cortex has expanded more than cortical motor areas in human evolution. Given the enlargement of the prefrontal cortex relative to motor cortex in humans, our hypothesis would predict corresponding volumetric increases in the parts of the cerebellum connected to the prefrontal cortex, relative to cerebellar lobules connected to the motor cortex. We tested the hypothesis by comparing the volumes of cerebellar lobules in structural MRI scans in capuchins, chimpanzees and humans. The fractions of cerebellar volume occupied by Crus I and Crus II were significantly larger in humans compared to chimpanzees and capuchins. Our results therefore support the hypothesis that in the cortico-cerebellar system, functionally related structures evolve in concert with each other. The evolutionary expansion of these prefrontal-projecting cerebellar territories might contribute to the evolution of the higher cognitive functions of humans.},
}
@article{BaraketalPIN-13,
       author = {Omri Barak and David Sussillo and Ranulfo Romo and Misha Tsodyks and L.F. Abbott},
        title = {From fixed points to chaos: Three models of delayed discrimination},
      journal = {Progress in Neurobiology},
       volume = {103},
        pages = {214-222},
         year = {2013},
         note = {Conversion of Sensory Signals into Perceptions, Memories and Decisions},
      comment = {This manuscript argues persuasively that mixed selectivity, a signature of high-dimensional neural representations, is a fundamental component of the computational power of prefrontal cortex.},
     abstract = {Working memory is a crucial component of most cognitive tasks. Its neuronal mechanisms are still unclear despite intensive experimental and theoretical explorations. Most theoretical models of working memory assume both time-invariant neural representations and precise connectivity schemes based on the tuning properties of network neurons. A different, more recent class of models assumes randomly connected neurons that have no tuning to any particular task, and bases task performance purely on adjustment of network readout. Intermediate between these schemes are networks that start out random but are trained by a learning scheme. Experimental studies of a delayed vibrotactile discrimination task indicate that some of the neurons in prefrontal cortex are persistently tuned to the frequency of a remembered stimulus, but the majority exhibit more complex relationships to the stimulus that vary considerably across time. We compare three models, ranging from a highly organized line attractor model to a randomly connected network with chaotic activity, with data recorded during this task. The random network does a surprisingly good job of both performing the task and matching certain aspects of the data. The intermediate model, in which an initially random network is partially trained to perform the working memory task by tuning its recurrent and readout connections, provides a better description, although none of the models matches all features of the data. Our results suggest that prefrontal networks may begin in a random state relative to the task and initially rely on modified readout for task performance. With further training, however, more tuned neurons with less time-varying responses should emerge as the networks become more structured.}
}
@article{CarlenSCIENCE-17,
       author = {Carl\`{e}n, Marie},
        title = {What constitutes the prefrontal cortex?},
      journal = {Science},
    publisher = {American Association for the Advancement of Science},
       volume = {358},
       number = {6362},
         year = {2017},
        pages = {478-482},
     abstract = {During evolution, the prefrontal region grew in size relative to the rest of the cortex. It reached its largest extent in the human brain, where it constitutes 30\% of the total cortical area. This growth was accompanied by phylogenetic differentiation of the cortical areas. It has been argued that the human brain holds prefrontal regions that are both qualitatively and functionally unique. Present-day neuroscientists studying the prefrontal cortex increasingly use mice. An important goal is to reveal how the prefrontal cortex enables complex behavior. However, the prefrontal cortex still lacks a conclusive definition. The structure and function of this brain area across species remain unresolved. This state of affairs is often overlooked, warranting renewed focus on what the prefrontal cortex is and does.},
}
@article{ChaudhurietalNEURON-15,
       author = {Rishidev Chaudhuri and Kenneth Knoblauch and Marie-Alice Gariel and Henry Kennedy and Xiao-Jing Wang},
        title = {A Large-Scale Circuit Mechanism for Hierarchical Dynamical Processing in the Primate Cortex},
      journal = {Neuron},
       volume = {88},
       number = {2},
         year = {2015},
        pages = {419-431},
     abstract = {Summary We developed a large-scale dynamical model of the macaque neocortex, which is based on recently acquired directed- and weighted-connectivity data from tract-tracing experiments, and which incorporates heterogeneity across areas. A hierarchy of timescales naturally emerges from this system: sensory areas show brief, transient responses to input (appropriate for sensory processing), whereas association areas integrate inputs over time and exhibit persistent activity (suitable for decision-making and working memory). The model displays multiple temporal hierarchies, as evidenced by contrasting responses to visual versus somatosensory stimulation. Moreover, slower prefrontal and temporal areas have a disproportionate impact on global brain dynamics. These findings establish a circuit mechanism for "temporal receptive windows" that are progressively enlarged along the cortical hierarchy, suggest an extension of time integration in decision making from local to large circuits, and should prompt a re-evaluation of the analysis of functional connectivity (measured by fMRI or electroencephalography/magnetoencephalography) by taking into account inter-areal heterogeneity.}
}
@article{DehaeneChangeuxNEURON-11,
       author = {Dehaene, Stanislas and Changeux, Jean-Pierre},
        title = {Experimental and Theoretical Approaches to Conscious Processing},
      journal = {Neuron},
         year = {2017},
    publisher = {Elsevier},
       volume = {70},
        issue = {2},
        pages = {200-227},
     abstract = {Recent experimental studies and theoretical models have begun to address the challenge of establishing a causal link between subjective conscious experience and measurable neuronal activity. The present review focuses on the well-delimited issue of how an external or internal piece of information goes beyond nonconscious processing and gains access to conscious processing, a transition characterized by the existence of a reportable subjective experience. Converging neuroimaging and neurophysiological data, acquired during minimal experimental contrasts between conscious and nonconscious processing, point to objective neural measures of conscious access: late amplification of relevant sensory activity, long-distance cortico-cortical synchronization at beta and gamma frequencies, and ?ignition? of a large-scale prefronto-parietal network. We compare these findings to current theoretical models of conscious processing, including the Global Neuronal Workspace (GNW) model according to which conscious access occurs when incoming information is made globally available to multiple brain systems through a network of neurons with long-range axons densely distributed in prefrontal, parieto-temporal, and cingulate cortices. The clinical implications of these results for general anesthesia, coma, vegetative state, and schizophrenia are discussed.},
}
@article{JohnsonetalNATURE-16,
       author = {Johnson, Carolyn M. and Peckler, Hannah and Tai, Lung-Hao and Wilbrecht, Linda},
        title = {Rule learning enhances structural plasticity of long-range axons in frontal cortex},
      journal = {Nature Communications},
    publisher = {Nature Publishing Group},
       volume = {7},
         year = {2016},
     abstract = {Rules encompass cue-action-outcome associations used to guide decisions and strategies in a specific context. Subregions of the frontal cortex including the orbitofrontal cortex (OFC) and dorsomedial prefrontal cortex (dmPFC) are implicated in rule learning, although changes in structural connectivity underlying rule learning are poorly understood. We imaged OFC axonal projections to dmPFC during training in a multiple choice foraging task and used a reinforcement learning model to quantify explore-exploit strategy use and prediction error magnitude. Here we show that rule training, but not experience of reward alone, enhances OFC bouton plasticity. Baseline bouton density and gains during training correlate with rule exploitation, while bouton loss correlates with exploration and scales with the magnitude of experienced prediction errors. We conclude that rule learning sculpts frontal cortex interconnectivity and adjusts a thermostat for the explore-exploit balance.},
}
@article{KoechlinandJubaultNEURON-06,
     title = {Broca's Area and the Hierarchical Organization of Human Behavior},
    author = {Koechlin, Etienne and Jubault, Thomas},
   journal = {Neuron},
    volume = 50,
     issue = 6,
      year = 2006,
     pages = {963-974},
  abstract = {The prefrontal cortex subserves executive control, i.e., the organization of action or thought in relation to internal goals. This brain region hosts a system of executive processes extending from premotor to the most anterior prefrontal regions that governs the temporal organization of behavior. Little is known, however, about the prefrontal executive system involved in the hierarchical organization of behavior. Here, we show using magnetic resonance imaging in humans that the posterior portion of the prefrontal cortex, including Broca's area and its homolog in the right hemisphere, contains a system of executive processes that control start and end states and the nesting of functional segments that combine in hierarchically organized action plans. Our results indicate that Broca's area and its right homolog process hierarchically structured behaviors regardless of their temporal organization, suggesting a fundamental segregation between prefrontal executive systems involved in the hierarchical and temporal organization of goal-directed behaviors.},
}
@article{KoechlinetalPNAS-00,
      journal = {Proceedings of the National Academy of Sciences},
       author = {Koechlin, Etienne and Corrado, Gregory and Pietrini, Pietro and Grafman, Jordan},
        title = {Dissociating the role of the medial and lateral anterior prefrontal cortex in human planning},
       volume = 97,
        issue = 13,
         year = 2000,
        pages = {7651-7656},
     abstract = {The anterior prefrontal cortex is known to subserve higher cognitive functions such as task management and planning. Less is known, however, about the functional specialization of this cortical region in humans. Using functional MRI, we report a double dissociation: the medial anterior prefrontal cortex, in association with the ventral striatum, was engaged preferentially when subjects executed tasks in sequences that were expected, whereas the polar prefrontal cortex, in association with the dorsolateral striatum, was  involved preferentially when subjects performed tasks in sequences that  were contingent on unpredictable events. These results parallel the  functional segregation previously described between the medial and  lateral premotor cortex underlying planned and contingent motor control  and extend this division to the anterior prefrontal cortex, when task  management and planning are required. Thus, our findings support the  assumption that common frontal organizational principles underlie motor  and higher executive functions in humans.},
}
@article{KoechlinetalSCIENCE-03,
       author = {Etienne Koechlin and Chryst\`{e}le Ody and Fr\'{e}d\'{e}rique Kouneiher},
        title = {The architecture of cognitive control in the human prefrontal cortex},
      journal = {Science},
       volume = 302,
         year = 2003,
        pages = {1181-1185},
     abstract = {The prefrontal cortex (PFC) subserves cognitive control: the ability to coordinate thoughts or actions in relation with internal goals. Its functional architecture, however, remains poorly understood. Using brain imaging in humans, we showed that the lateral PFC is organized as a cascade of executive processes from premotor to anterior PFC regions that control behavior according to stimuli, the present perceptual context, and the temporal episode in which stimuli occur, respectively. The results support an unified modular model of cognitive control that describes the overall functional organization of the human lateral PFC and has basic methodological and theoretical implications.},
}
@article{LaraandWallisFiSN-15,
       author = {Lara, Antonio H. and Wallis, Jonathan D.},
        title = {The Role of Prefrontal Cortex in Working Memory: A Mini Review},
      journal = {Frontiers in System Neuroscience},
    publisher = {Frontiers Media S.A.},
       volume = {9},
         year = {2015},
        pages = {173},
     abstract = {A prominent account of prefrontal cortex (PFC) function is that single neurons within the PFC maintain representations of task-relevant stimuli in working memory. Evidence for this view comes from studies in which subjects hold a stimulus across a delay lasting up to several seconds. Persistent elevated activity in the PFC has been observed in animal models as well as in humans performing these tasks. This persistent activity has been interpreted as evidence for the encoding of the stimulus itself in working memory. However, recent findings have posed a challenge to this notion. A number of recent studies have examined neural data from the PFC and posterior sensory areas, both at the single neuron level in primates, and at a larger scale in humans, and have failed to find encoding of stimulus information in the PFC during tasks with a substantial working memory component. Strong stimulus related information, however, was seen in posterior sensory areas. These results suggest that delay period activity in the PFC might be better understood not as a signature of memory storage per se, but as a top down signal that influences posterior sensory areas where the actual working memory representations are maintained.},
}
@article{OReillySCIENCE-06,
     title = {Biologically Based Computational Models of High-Level Cognition},
    author = {O'Reilly, Randall C.},
   journal = {Science},
    volume = 314,
     issue = 5796,
      year = 2006,
     pages = {91-94},
  abstract = {Computer models based on the detailed biology of the brain can help us understand the myriad complexities of human cognition and intelligence. Here, we review models of the higher level aspects of human intelligence, which depend critically on the prefrontal cortex and associated subcortical areas. The picture emerging from a convergence of detailed mechanistic models and more abstract functional models represents a synthesis between analog and digital forms of computation. Specifically, the need for robust active maintenance and rapid updating of information in the prefrontal cortex appears to be satisfied by bistable activation states and dynamic gating mechanisms. These mechanisms are fundamental to digital computers and may be critical for the distinctive aspects of human intelligence.},
}
@article{RottschyetalNEUROIMAGE-12,
       author = {Rottschy, C. and Langner, R. and Dogan, I. and Reetz, K. and Laird, A. R. and Schulz, J. B. and Fox, P. T. and Eickhoff, S. B.},
        title = {Modelling neural correlates of working memory: A coordinate-based meta-analysis},
      journal = {Neuroimage},
         year = {2012},
       volume = {60},
        issue = {1},
        pages = {830-846},
     abstract = {Working memory subsumes the capability to memorize, retrieve and utilize information for a limited period of time which is essential to many human behaviours. Moreover, impairments of working memory functions may be found in nearly all neurological and psychiatric diseases. To examine what brain regions are commonly and differently active during various working memory tasks, we performed a coordinate-based meta-analysis over 189 fMRI experiments on healthy subjects. The main effect yielded a widespread bilateral fronto-parietal network. Further meta-analyses revealed that several regions were sensitive to specific task components, e.g. Broca's region was selectively active during verbal tasks or ventral and dorsal premotor cortex were preferentially involved in memory for object identity and location, respectively. Moreover, the lateral prefrontal cortex showed a division in a rostral and a caudal part based on differential involvement in task-set and load effects. Nevertheless, a consistent but more restricted core network emerged from conjunctions across analyses of specific task designs and contrasts. This core network appears to comprise the quintessence of regions, which are necessary during working memory tasks. It may be argued that the core regions form a distributed executive network with potentially generalized functions for focusing on competing representations in the brain. The present study demonstrates that meta-analyses are a powerful tool to integrate the data of functional imaging studies on a (broader) psychological construct, probing the consistency across various paradigms as well as the differential effects of different experimental implementations.},
}
@incollection{TefferandSemendeferiPBR-12,
       author = {Kate Teffer and Katerina Semendeferi},
        title = {Chapter 9 - Human prefrontal cortex: Evolution, development and pathology},
       editor = {Michel A. Hofman and Dean Falk},
       series = {Progress in Brain Research},
    publisher = {Elsevier},
       volume = {195},
        pages = {191-218},
         year = {2012},
    booktitle = {Evolution of the Primate Brain},
     abstract = {The prefrontal cortex is critical to many cognitive abilities that are considered particularly human, and forms a large part of a neural system crucial for normal socio-emotional and executive functioning in humans and other primates. In this chapter, we survey the literature regarding prefrontal development and pathology in humans as well as comparative studies of the region in humans and closely related primate species. The prefrontal cortex matures later in development than more caudal regions, and some of its neuronal subpopulations exhibit more complex dendritic arborizations. Comparative work suggests that the human prefrontal cortex differs from that of closely related primate species less in relative size than it does in organization. Specific reorganizational events in neural circuitry may have taken place either as a consequence of adjusting to increases in size or as adaptive responses to specific selection pressures. Living in complex environments has been recognized as a considerable factor in the evolution of primate cognition. Normal frontal lobe development and function are also compromised in several neurological and psychiatric disorders. A phylogenetically recent reorganization of frontal cortical circuitry may have been critical to the emergence of human-specific executive and social-emotional functions, and developmental pathology in these same systems underlies many psychiatric and neurological disorders, including autism and schizophrenia.}
}
@article{WatsonetalFiSN-14,
       author = {Watson, Thomas C. and Becker, Nadine and Apps, Richard and Jones, Matthew W.},
        title = {Back to front: cerebellar connections and interactions with the prefrontal cortex},
      journal = {Frontiers in Systems Neuroscience},
         year = {2014},
    publisher = {Frontiers Media S.A.},
       volume = {8},
        pages = {4},
     abstract = {Although recent neuroanatomical evidence has demonstrated closed-loop connectivity between prefrontal cortex and the cerebellum, the physiology of cerebello-cerebral circuits and the extent to which cerebellar output modulates neuronal activity in neocortex during behavior remain relatively unexplored. We show that electrical stimulation of the contralateral cerebellar fastigial nucleus (FN) in awake, behaving rats evokes distinct local field potential (LFP) responses (onset latency {\textasciitilde}13 ms) in the prelimbic (PrL) subdivision of the medial prefrontal cortex. Trains of FN stimulation evoke heterogeneous patterns of response in putative pyramidal cells in frontal and prefrontal regions in both urethane-anesthetized and awake, behaving rats. However, the majority of cells showed decreased firing rates during stimulation and subsequent rebound increases; more than 90\% of cells showed significant changes in response. Simultaneous recording of on-going LFP activity from FN and PrL while rats were at rest or actively exploring an open field arena revealed significant network coherence restricted to the theta frequency range (5-10 Hz). Granger causality analysis indicated that this coherence was significantly directed from cerebellum to PrL during active locomotion. Our results demonstrate the presence of a cerebello-prefrontal pathway in rat and reveal behaviorally dependent coordinated network activity between the two structures, which could facilitate transfer of sensorimotor information into ongoing neocortical processing during goal directed behaviors.},
}
@article{ZhaoetalNATURE-16,
       author = {Zhou, Xin and Zhu, Dantong and Qi, Xue-Lian and Li, Sihai and King, Samson G. and Salinas, Emilio and Stanford, Terrence R. and Constantinidis, Christos},
        title = {Neural correlates of working memory development in adolescent primates},
      journal = {Nature Communications},
    publisher = {Nature Publishing Group},
         year = {2016},
       volume = {9},
        issue = {7},
        pages = {13423},
     abstract = {Working memory ability matures after puberty, in parallel with structural changes in the prefrontal cortex, but little is known about how changes in prefrontal neuronal activity mediate this cognitive improvement in primates. To address this issue, we compare behavioural performance and neurophysiological activity in monkeys as they transitioned from puberty into adulthood. Here we report that monkeys perform working memory tasks reliably during puberty and show modest improvement in adulthood. The adult prefrontal cortex is characterized by increased activity during the delay period of the task but no change in the representation of stimuli. Activity evoked by distracting stimuli also decreases in the adult prefrontal cortex. The increase in delay period activity relative to the baseline activity of prefrontal neurons is the best correlate of maturation and is not merely a consequence of improved performance. Our results reveal neural correlates of the working memory improvement typical of primate adolescence.},
}

⁴³ Here are two excerpts from Dehaene [68] relating to attention, working memory and executive control. Note that working memory and conscious perception were thought to share similar brain mechanisms. Trübutschek et al [283] examine recent reports of non-conscious working memory that challenge this view. They combine visual masking with magnetoencephalography to demonstrate the reality of nonconscious working memory and dissect its neural mechanisms:

The component of the mind that psychologists call working memory is one of the dominant functions of the dorsolateral prefrontal cortex and the areas that it connects with, thus making these areas strong candidates for the depositories of our conscious knowledge. These regions pop up in brain imaging experiments whenever we briefly hold on to a piece of information: a phone number, a color, or the shape of a flashed picture. Prefrontal neurons implement an active memory: long after the picture is gone, they continue to fire throughout the short-term memory task—sometimes as long as dozens of seconds later. And when the prefrontal cortex is impaired or distracted, this memory is lost—it falls into unconscious oblivion.

Together with the physicists Mariano Sigman and Ariel Zylberberg, I have begun to explore the computational properties that such a device would possess. It closely resembles what computer scientists call a "production system," a type of program introduced in the 1960s to implement artificial intelligence tasks. A production system comprises a database, also called "working memory," and a vast array of if-then production rules (e.g., if there is an A in working memory, then change it to the sequence BC). At each step, the system examines whether a rule matches the current state of its working memory. If multiple rules match, then they compete under the aegis of a stochastic prioritizing system. Finally, the winning rule ignites and is allowed to change the contents of working memory before the entire process resumes. Thus this sequence of steps amounts to serial cycles of unconscious competition, conscious ignition, and broadcasting.

⁴⁴ Here is Stanislas Dehaene writing about the relationship between consciousness and maintaining persistent thoughts:

There may be a very good reason why our consciousness condenses sensory messages into a synthetic code, devoid of gaps and ambiguities: such a code is compact enough to be carried forward in time, entering what we usually call “working memory.” Working memory and consciousness seem to be tightly related. One may even argue, with Daniel Dennett, that a main role of consciousness is to create lasting thoughts. Once a piece of information is conscious, it stays fresh in our mind for as long as we care to attend to it and remember it. The conscious brief must be kept stable enough to inform our decisions, even if they take a few minutes to form. This extended duration, thickening of the present moment, is characteristic of our conscious thoughts. — Stanislas Dehaene [68]

⁴⁵ Bibliography references for the four tutorials on neural networks used as supplements in CS379C including URL and date of the cached PDF on the course website:

@misc{KarparthyCONVOLUTIONAL-NEURAL-NETWORKS-16,
        title = {Convolutional Neural Networks for Visual Recognition},
       author = {Andrej Karparthy},
 howpublished = {http://cs231n.github.io/convolutional-networks/},
         year = {2016},
}	 
@Misc{KarparthyUNREASONABLY-EFFECTIVE-RNN-15,
        title = {The Unreasonable Effectivenss of Recursive Neutal Networks},
       author = {Andrej Karparthy},
 howpublished = {http://karpathy.github.io/2015/05/21/rnn-effectiveness/},
         year = {2015},
}	 
@article{OlahandCarterATTENTION-TUTORIAL-16,
       author = {Olah, Chris and Carter, Shan},
        title = {Attention and Augmented Recurrent Neural Networks},
      journal = {Distill},
          url = {http://distill.pub/2016/augmented-rnns},
         year = {2016},
}
@misc{OlahLSTM-NEURAL-NETWORK-TUTORIAL-15,
        title = {Understanding LSTM Networks},
       author = {Christopher Olah},
 howpublished = {http://colah.github.io/posts/2015-08-Understanding-LSTMs},
         year = {2015},
} 

⁴⁶ In Greek mythology, Pan is the god of the mountain wilds and wooded glens, the champion of rustic music and the frequent consort of the woodland nymphs. Pan is also the god of theatrical criticism and impromptus. In addition to the usual duties of godhood in Greek mythology, Pan has a reputation of being slyly mischievous and sexually promiscuous, characteristics we will not attempt to recreate in our digital apprentice. After Pan our next choice is Panache, a word of French origin that carries the connotation of flamboyant manner, confidence and reckless courage, and might serve as an acronym for the unwieldy full name [P]rogrammer's [A]pprentice [N]eural-network [A]rchitecture [H]ierarchical [E]xecutive [C]ontroller, allowing only one word out of place to reflect Pan's mischievous wit.

⁴⁷ Recent work and seminal papers relating to the notion of dynamic links — also called fast weights — to attend to the recent past [289, 290]:

@incollection{vonderMalsburgPNN-94,
        title = {The Correlation Theory of Brain Function},
       author = {Christoph von der Malsburg},
    booktitle = {Models of Neural Networks: Physics of Neural Networks},
       editor = {Domany, E. and van Hemmen, J.L. and Schulten, K.},
    publisher = {Springer},
         year = {1994},
     abstract = {A summary of brain theory is given so far as it is contained within the framework of Localization Theory. Diffculties of this "conventional theory" are traced back to a specific deficiency: there is no way to express relations between active cells (as for instance their representing parts of the same object). A new theory is proposed to cure this deficiency. It introduces a new kind of dynamical control, termed synaptic modulation, according to which synapses switch between a conducting and a non- conducting state. The dynamics of this variable is controlled on a fast time scale by correlations in the temporal fine structure of cellular signals. Furthermore, conventional synaptic plasticity is replaced by a refined version. Synaptic modulation and plasticity form the basis for short-term and long-term memory, respectively. Signal correlations, shaped by the variable network, express structure and relationships within objects. In particular, the figure-ground problem may be solved in this way. Synaptic modulation introduces flexibility into cerebral networks which is necessary to solve the invariance problem. Since momentarily useless connections are deactivated, interference between different memory traces can be reduced, and memory capacity increased, in comparison with conventional associative memory.}
}
@techreport{vonderMalsburgMPI-81,
        title = {The Correlation Theory of Brain Function},
       author = {Christoph von der Malsburg},
        issue = {{Internal Report 81-2}},
  institution = {Max Planck Institute for Biophysical Chemistry},
         year = 1981
}
@incollection{HintonandPlautCSS-87,
       author = {Hinton, G. E. and Plaut, D. C.},
        title = {Using fast weights to deblur old},
    booktitle = {Proceedings of the 9th Annual Conference of the Cognitive Science Society},
    publisher = {Lawrence Erlbaum Associates},
         year = {1987},
        pages = {177-186},
     abstract = {Connectionist models usually have a single weight on each connection. Some interesting new properties emerge if each connection has two weights: A slowly changing, plastic weight which stores long-term knowledge and a fast-changing, elastic weight which stores temporary knowledge and spontaneously decays towards zero. If a network learns a set of associations and then these associations are "blurred" by subsequent learning, all the original associations can be "deblurred" by rehearsing on just a few of them. The rehearsal allows the fast weights to take on values that temporarily cancel out the changes in the slow weights caused by the subsequent learning.}
}
@article{BaetalCoRR-16,
       author = {Jimmy Ba and Geoffrey Hinton and Volodymyr Mnih and Joel Z. Leibo and Catalin Ionescu},
        title = {Using Fast Weights to Attend to the Recent Past},
      journal = {CoRR},
       volume = {arXiv:1610.06258},
         year = {2016},
     abstract = {Until recently, research on artificial neural networks was largely restricted to systems with only two types of variable: Neural activities that represent the current or recent input and weights that learn to capture regularities among inputs, outputs and payoffs. There is no good reason for this restriction. Synapses have dynamics at many different time-scales and this suggests that artificial neural networks might benefit from variables that change slower than activities but much faster than the standard weights. These "fast weights" can be used to store temporary memories of the recent past and they provide a neurally plausible way of implementing the type of attention to the past that has recently proved very helpful in sequence-to-sequence models. By using fast weights we can avoid the need to store copies of neural activity patterns.}
}
@article{AnonymousICLR-17b,
        title = {Gated Fast Weights for Associative Retrieval},
       author = {Anonymous},
      journal = {International Conference on Learning Representations},
         year = {2017},
     abstract = {We improve previous end-to-end differentiable neural networks (NNs) with fast weight memories. A gate mechanism updates fast weights at every time step of a sequence through two separate outer-product-based matrices generated by slow parts of the net. The system is trained on a complex sequence to sequence variation of the Associative Retrieval Problem with roughly 70 times more temporal memory (i.e. time-varying variables) than similar-sized standard recurrent NNs (RNNs). In terms of accuracy and number of parameters, our architecture outperforms a variety of RNNs, including Long Short-Term Memory, Hypernetworks, and related fast weight architectures.}
}
@article{ZhuandvonderMalsburgNEUROCOMPUTING-02,
       author = {Junmei Zhu and Christoph von der Malsburg},
        title = {Synapto–synaptic interactions speed up dynamic link matching},
      journal = {Neurocomputing},
       volume = {44-46},
         year = {2002},
        pages = {721-728},
     abstract = {An extended dynamic link matching (DLM) model is presented for the fast creation of mappings invariant to position, scale, deformation and orientation, a process underlying various visual tasks that are fast and effortless for humans but are computationally very difficult. Structured interactions between synapses are represented by control units. Each control unit stands for a group of synapses that are consistent with each other in terms of relative position, scale and orientation. Specifically, it is shown in simulations how these high-order interactions between synapses lead to very fast DLM.}
}
@article{HuangetalCoRR-17,
       author = {{Huang}, Q. and {Smolensky}, P. and {He}, X. and {Deng}, L. and {Wu}, D.},
        title = {{Tensor Product Generation Networks}},
      journal = {CoRR},
         year = 2017,
     abstract = {We present a new tensor product generation network (TPGN) that generates natural language descriptions for images. The model has a novel architecture that instantiates a general framework for encoding and processing symbolic structure through neural network computation. This framework is built on Tensor Product Representations (TPRs). We evaluated the proposed TPGN on the MS COCO image captioning task. The exeprimental results show that the TPGN outperforms the LSTM based state-of-the-art baseline with a significant margin. Further, we show that our caption generation model can be interpreted as generating sequences of grammatical categories and retrieving words by their categories from a plan encoded as a distributed representation.}
}
@book{SmolenskyandLegendre2011,
        title = {The Harmonic Mind: From Neural Computation to Optimality-theoretic Grammar. Linguistic and philosophical implications},
       author = {Smolensky, P. and Legendre, G.},
       number = {{II}},
         year = {2011},
    publisher = {MIT Press}
}

⁴⁹ K. L. Parker, Y. C. Kim, R. M. Kelley, A. J. Nessler, K-H Chen, V. A. Muller-Ewald, N. C. Andreasen, N. S. Narayanan. Delta-frequency stimulation of cerebellar projections can compensate for schizophrenia-related medial frontal dysfunction. Molecular Psychiatry, (2017).

The latest findings by Parker and colleagues provide fresh insights into how the cerebellum influences neural networks in the frontal lobes and the role of the cerebellum in cognitive processing. This study also suggests that delta-frequency cerebellar stimulation might help improve cognitive problems in human patients with schizophrenia.

Mark J. Wagner, Tony Hyun Kim, Joan Savall, Mark J. Schnitzer, Liqun Luo. Cerebellar granule cells encode the expectation of reward. Nature (2017)

Along this line, last week, neuroscientists at Stanford University serendipitously discovered (for the first time) that granule cells in the cerebellum encode and predict rewards. The Stanford study, "Cerebellar Granule Cells Encode the Expectation of Reward," was published March 20 online ahead of print in the journal Nature.

Manish Saggar, Eve-Marie Quintin, Nicholas T. Bott, Eliza Kienitz, Yin-hsuan Chien, Daniel W-C. Hong, Ning Liu, Adam Royalty, Grace Hawthorne, Allan L. Reiss. Changes in Brain Activation Associated with Spontaneous Improvization and Figural Creativity After Design-Thinking-Based Training: A Longitudinal fMRI Study. Cerebral Cortex (2016).

Another team of researchers at Stanford University, led by Manish Saggar, reported that optimizing cerebral-cerebellar connectivity increases creative capacity. Saggar's team found that suppressing the executive-control functions of the cerebrum — while encouraging the cerebellum to become the "controller" — increased spontaneous creative capacity.

⁴⁸ Excerpts and page links from Wikipedia on brain structures relevant to attention and executive control. They are provided here as a convenient reference for quick review. Follow the supplied links for the full story and remember that this is Wikipedia and not the latest — 5th edition (October 26, 2012) — edition of Kandel, Schwartz, Jessell, Siegelbaum and Hudspeth [156]:

1. Tectum — In adult humans, it only consists of the inferior and the superior colliculi.

   1. The superior colliculus is involved in preliminary visual processing and control of eye movements. In non-mammalian vertebrates it serves as the main visual area of the brain, functionally analogous to the visual areas of the cerebral cortex in mammals.

   2. The inferior colliculus is involved in auditory processing. It receives input from various brain stem nuclei and projects to the medial geniculate nucleus of the thalamus, which relays auditory information to the primary auditory cortex.

   Both colliculi also have descending projections to the paramedian pontine reticular formation and spinal cord, and thus can be involved in responses to stimuli faster than cortical processing would allow. Collectively the colliculi are referred to as the corpora quadrigemina.

2. Thalamus — The thalamus is the large mass of gray matter in the dorsal part of the diencephalon of the brain with several functions such as relaying of sensory signals, including motor signals, to the cerebral cortex, and the regulation of consciousness, sleep, and alertness.

   1. The thalamus has multiple functions, generally believed to act as a relay station, or hub, relaying information between different subcortical areas and the cerebral cortex. In particular, every sensory system (with the exception of the olfactory system) includes a thalamic nucleus that receives sensory signals and sends them to the associated primary cortical area.

   2. For the visual system, for example, inputs from the retina are sent to the lateral geniculate nucleus of the thalamus, which in turn projects to the visual cortex in the occipital lobe. The thalamus is believed to both process sensory information as well as relay it—each of the primary sensory relay areas receives strong feedback connections from the cerebral cortex.

3. Basal Ganglia — The basal ganglia (or basal nuclei) is a group of subcortical nuclei, of varied origin, in the brains of vertebrates including humans, which are situated at the base of the forebrain. Basal ganglia are strongly interconnected with the cerebral cortex, thalamus, and brainstem, as well as several other brain areas. The basal ganglia are associated with a variety of functions including: control of voluntary motor movements, procedural learning, routine behaviors or "habits" such as teeth grinding, eye movements, cognition, and emotion.

   1. Popular theories implicate the basal ganglia primarily in action selection — in helping to decide which of several possible behaviors to execute at any given time. In more specific terms, the basal ganglia's primary function is likely to control and regulate activities of the motor and premotor cortical areas so that voluntary movements can be performed smoothly.

   2. Experimental studies show that the basal ganglia exert an inhibitory influence on a number of motor systems, and that a release of this inhibition permits a motor system to become active. The "behavior switching" that takes place within the basal ganglia is influenced by signals from many parts of the brain, including the prefrontal cortex, which plays a key role in executive functions⁴⁹.

4. Prefrontal Cortex — This brain region has been implicated in planning complex cognitive behavior, personality expression, decision making, and moderating social behavior. The basic activity of this brain region is considered to be orchestration of thoughts and actions in accordance with internal goals.

   The most typical psychological term for functions carried out by the prefrontal cortex area is executive function. Executive function relates to abilities to differentiate among conflicting thoughts, determine good and bad, better and best, same and different, future consequences of current activities, working toward a defined goal, prediction of outcomes, expectation based on actions, and social "control" (the ability to suppress urges that, if not suppressed, could lead to socially unacceptable outcomes).

⁵⁰ The concept of working memory is defined in terms of how it facilitates problem solving. Its relationship to the concepts of short- and long-term memory are described in [49] and summarized in the paper's abstract below. Cowan [49] emphasizes the connection between working memory and attention. O'Reilly [220] emphasizes the active maintenance and rapid updating of information required for planning and decision making in terms of different proposed gating mechanisms, and uses related mechanisms to explain dynamic variable binding. Eliasmith [86] and Stocco et al [271] provide additional detail concerning the role of the basal ganglia.

The exact location of working memory in the brain is still a matter of some controversy due, in no small part, to fuzziness about what constitutes working memory and how we might go about localizing it. Rottschy et al [249] performed a coordinate-based meta-analysis over 189 fMRI experiments involving various tasks relying on the exercise of working memory tasks. The main effects across all of these experiments involving working memory reveal consistent bilateral activity of fronto-parietal networks clearly visible in the fMRI data accompanying the Rottschy et al [249] paper and can explored using the 3-D graphics tools provided here.

@article{CowanPBR-08,
       author = {Cowan, Nelson},
        title = {What are the differences between long-term, short-term and working memory?},
      journal = {Progress in Brain Research},
         year = {2008},
       volume = {169},
        pages = {323-338},
     abstract = {In the recent literature there has been considerable confusion about the three types of memory: long-term, short-term, and working memory. This chapter strives to reduce that confusion and makes up-to-date assessments of these types of memory. Long- and short-term memory could differ in two fundamental ways, with only short-term memory demonstrating (1) temporal decay and (2) chunk capacity limits. Both properties of short-term memory are still controversial but the current literature is rather encouraging regarding the existence of both decay and capacity limits. Working memory has been conceived and defined in three different, slightly discrepant ways: as short-term memory applied to cognitive tasks, as a multi-component system that holds and manipulates information in short-term memory, and as the use of attention to manage short-term memory. Regardless of the definition, there are some measures of memory in the short term that seem routine and do not correlate well with cognitive aptitudes and other measures (those usually identified with the term 'working memory') that seem more attention demanding and do correlate well with these aptitudes. The evidence is evaluated and placed within a theoretical framework depicted in Figure 1.},
}
@book{Eliasmith2013,
        title = {How to Build a Brain: A Neural Architecture for Biological Cognition},
       author = {Eliasmith, Chris},
       series = {Oxford Series on Cognitive Modeling},
         year = {2013},
    publisher = {Oxford University Press {USA}},
     abstract = {One goal of researchers in neuroscience, psychology, and artificial intelligence is to build theoretical models that are able to explain the flexibility and adaptiveness of biological systems. How to build a brain provides a detailed guided exploration of a new cognitive architecture that takes biological detail seriously, while addressing cognitive phenomena. The Semantic Pointer Architecture (SPA) introduced in this book provides a set of tools for constructing a wide range of biologically constrained perceptual, cognitive, and motor models.}
}
@article{OReillySCIENCE-06,
     title = {Biologically Based Computational Models of High-Level Cognition},
    author = {O'Reilly, Randall C.},
   journal = {Science},
    volume = 314,
     issue = 5796,
      year = 2006,
     pages = {91-94},
  abstract = {Computer models based on the detailed biology of the brain can help us understand the myriad complexities of human cognition and intelligence. Here, we review models of the higher level aspects of human intelligence, which depend critically on the prefrontal cortex and associated subcortical areas. The picture emerging from a convergence of detailed mechanistic models and more abstract functional models represents a synthesis between analog and digital forms of computation. Specifically, the need for robust active maintenance and rapid updating of information in the prefrontal cortex appears to be satisfied by bistable activation states and dynamic gating mechanisms. These mechanisms are fundamental to digital computers and may be critical for the distinctive aspects of human intelligence.},
}
@article{RottschyetalNEUROIMAGE-12,
       author = {Rottschy, C. and Langner, R. and Dogan, I. and Reetz, K. and Laird, A. R. and Schulz, J. B. and Fox, P. T. and Eickhoff, S. B.},
        title = {Modelling neural correlates of working memory: A coordinate-based meta-analysis},
      journal = {Neuroimage},
         year = {2012},
       volume = {60},
        issue = {1},
        pages = {830-846},
     abstract = {Working memory subsumes the capability to memorize, retrieve and utilize information for a limited period of time which is essential to many human behaviours. Moreover, impairments of working memory functions may be found in nearly all neurological and psychiatric diseases. To examine what brain regions are commonly and differently active during various working memory tasks, we performed a coordinate-based meta-analysis over 189 fMRI experiments on healthy subjects. The main effect yielded a widespread bilateral fronto-parietal network. Further meta-analyses revealed that several regions were sensitive to specific task components, e.g. Broca's region was selectively active during verbal tasks or ventral and dorsal premotor cortex were preferentially involved in memory for object identity and location, respectively. Moreover, the lateral prefrontal cortex showed a division in a rostral and a caudal part based on differential involvement in task-set and load effects. Nevertheless, a consistent but more restricted core network emerged from conjunctions across analyses of specific task designs and contrasts. This core network appears to comprise the quintessence of regions, which are necessary during working memory tasks. It may be argued that the core regions form a distributed executive network with potentially generalized functions for focusing on competing representations in the brain. The present study demonstrates that meta-analyses are a powerful tool to integrate the data of functional imaging studies on a (broader) psychological construct, probing the consistency across various paradigms as well as the differential effects of different experimental implementations.},
}
@article{StoccoetalPR-10,
       author = {Stocco, Andrea and Lebiere, Christian and Anderson, John R.},
        title = {Conditional Routing of Information to the Cortex: A Model of the Basal Ganglia's Role in Cognitive Coordination},
      journal = {Psychology Review},
         year = {2010},
       volume = {117},
        issue = {2},
        pages = {541-574},
     abstract = {The basal ganglia play a central role in cognition and are involved in such general functions as action selection and reinforcement learning. Here, we present a model exploring the hypothesis that the basal ganglia implement a conditional information-routing system. The system directs the transmission of cortical signals between pairs of regions by manipulating separately the selection of sources and destinations of information transfers. We suggest that such a mechanism provides an account for several cognitive functions of the basal ganglia. The model also incorporates a possible mechanism by which subsequent transfers of information control the release of dopamine. This signal is used to produce novel stimulus-response associations by internalizing transferred cortical representations in the striatum. We discuss how the model is related to production systems and cognitive architectures. A series of simulations is presented to illustrate how the model can perform simple stimulus-response tasks, develop automatic behaviors, and provide an account of impairments in Parkinson's and Huntington's diseases.},
}
@article{TrubutschekBIORXIV-16,
       author = {Tr\"{u}butschek, Darinka and Marti, S\'{e}bastien and Ojeda, Andr\'{e}s and King, Jean-R\'{e}mi and Mi, Yuanyuan and Tsodyks, Misha and Dehaene, Stanislas},
        title = {A theory of working memory without consciousness or sustained activity},
      journal = {Biorxiv},
         year = {2016},
     abstract = {Working memory and conscious perception are thought to share similar brain mechanisms, yet recent reports of non-conscious working memory challenge this view. Combining visual masking with magnetoencephalography, we demonstrate the reality of nonconscious working memory and dissect its neural mechanisms. In a spatial delayed-response task, participants reported the location of a subjectively unseen target above chance-level after a long delay. Conscious perception and conscious working memory were characterized by similar signatures: a sustained desynchronization in the alpha/beta band over frontal cortex, and a decodable representation of target location in posterior sensors. During non-conscious working memory, such activity vanished. Our findings contradict models that identify working memory with sustained neural firing, but are compatible with recent proposals of ‘activity-silent’ working memory. We present a theoretical framework and simulations showing how slowly decaying synaptic changes allow cell assemblies to go dormant during the delay, yet be retrieved above chance-level after several seconds.}
}
@article{ZhaoetalNATURE-16,
       author = {Zhou, Xin and Zhu, Dantong and Qi, Xue-Lian and Li, Sihai and King, Samson G. and Salinas, Emilio and Stanford, Terrence R. and Constantinidis, Christos},
        title = {Neural correlates of working memory development in adolescent primates},
      journal = {Nature Communications},
    publisher = {Nature Publishing Group},
         year = {2016},
       volume = {9},
        issue = {7},
        pages = {13423},
     abstract = {Working memory ability matures after puberty, in parallel with structural changes in the prefrontal cortex, but little is known about how changes in prefrontal neuronal activity mediate this cognitive improvement in primates. To address this issue, we compare behavioural performance and neurophysiological activity in monkeys as they transitioned from puberty into adulthood. Here we report that monkeys perform working memory tasks reliably during puberty and show modest improvement in adulthood. The adult prefrontal cortex is characterized by increased activity during the delay period of the task but no change in the representation of stimuli. Activity evoked by distracting stimuli also decreases in the adult prefrontal cortex. The increase in delay period activity relative to the baseline activity of prefrontal neurons is the best correlate of maturation and is not merely a consequence of improved performance. Our results reveal neural correlates of the working memory improvement typical of primate adolescence.},
}

⁵¹ Two relative recent papers on simulating and reasoning about physical systems with an emphasis of continuous models that can be easily visualized:

% @article{BattagliaetalCoRR-16,
       author = {Peter W. Battaglia and Razvan Pascanu and Matthew Lai and Danilo Jimenez Rezende and Koray Kavukcuoglu},
        title = {Interaction Networks for Learning about Objects, Relations and Physics},
      journal = {CoRR},
       volume = {arXiv:1612.00222},
         year = {2016},
     abstract = {Reasoning about objects, relations, and physics is central to human intelligence, and a key goal of artificial intelligence. Here we introduce the interaction network, a model which can reason about how objects in complex systems interact, supporting dynamical predictions, as well as inferences about the abstract properties of the system. Our model takes graphs as input, performs object- and relation-centric reasoning in a way that is analogous to a simulation, and is implemented using deep neural networks. We evaluate its ability to reason about several challenging physical domains: n-body problems, rigid-body collision, and non-rigid dynamics. Our results show it can be trained to accurately simulate the physical trajectories of dozens of objects over thousands of time steps, estimate abstract quantities such as energy, and generalize automatically to systems with different numbers and configurations of objects and relations. Our interaction network implementation is the first general-purpose, learnable physics engine, and a powerful general framework for reasoning about object and relations in a wide variety of complex real-world domains.}
}
@article{ChangetalCoRR-16,
       author = {Michael B. Chang and Tomer Ullman and Antonio Torralba and Joshua B. Tenenbaum},
        title = {A Compositional Object-Based Approach to Learning Physical Dynamics},
      journal = {CoRR},
       volume = {arXiv:1612.00341},
         year = {2016},
     abstract = {We present the Neural Physics Engine (NPE), a framework for learning simulators of intuitive physics that naturally generalize across variable object count and different scene configurations. We propose a factorization of a physical scene into composable object-based representations and a neural network architecture whose compositional structure factorizes object dynamics into pairwise interactions. Like a symbolic physics engine, the NPE is endowed with generic notions of objects and their interactions; realized as a neural network, it can be trained via stochastic gradient descent to adapt to specific object properties and dynamics of different worlds. We evaluate the efficacy of our approach on simple rigid body dynamics in two-dimensional worlds. By comparing to less structured architectures, we show that the NPE's compositional representation of the structure in physical interactions improves its ability to predict movement, generalize across variable object count and different scene configurations, and infer latent properties of objects such as mass.},
}

⁵² Maseo Ito's 2012 monograph summarizing 50 years of work on the cerebellum and related cortical structures:

@book{Ito2012,
        title = {The Cerebellum: Brain for an Implicit Self},
       author = {Ito, Masao},
    publisher = {FT Press},
         year = {2012},
}

⁵³ Review articles relating to the control of non-motor activities that involve closed loop connections between the cerebral corted and in particular prefrontal region and the cerebellar cortex via the basal ganglia, dentate gyrus and thalamus:

@article{ItoNRN-08,
       author = {Ito, Masao},
        title = {Control of mental activities by internal models in the cerebellum},
      journal = {Nature Reviews Neuroscience},
    publisher = {Nature Publishing Group},
       volume = 9,
        issue = 4,
         year = 2008,
        pages = {304-313},
     abstract = {The intricate neuronal circuitry of the cerebellum is thought to encode internal models that reproduce the dynamic properties of body parts. These models are essential for controlling the movement of these body parts: they allow the brain to precisely control the movement without the need for sensory feedback. It is thought that the cerebellum might also encode internal models that reproduce the essential properties of mental representations in the cerebral cortex. This hypothesis suggests a possible mechanism by which intuition and implicit thought might function and explains some of the symptoms that are exhibited by psychiatric patients. This article examines the conceptual bases and experimental evidence for this hypothesis.},
}
@article{StricketalARN-09,
       author = {Peter L. Strick and Richard P. Dum and Julie A. Fiez},
        title = {Cerebellum and Nonmotor Function},
      journal = {Annual Review of Neuroscience},
       volume = {32},
       number = {1},
        pages = {413-434},
         year = {2009},
     abstract = {Does the cerebellum influence nonmotor behavior? Recent anatomical studies demonstrate that the output of the cerebellum targets multiple nonmotor areas in the prefrontal and posterior parietal cortex, as well as the cortical motor areas. The projections to different cortical areas originate from distinct output channels within the cerebellar nuclei. The cerebral cortical area that is the main target of each output channel is a major source of input to the channel. Thus, a closed-loop circuit represents the major architectural unit of cerebro-cerebellar interactions. The outputs of these loops provide the cerebellum with the anatomical substrate to influence the control of movement and cognition. Neuroimaging and neuropsychological data supply compelling support for this view. The range of tasks associated with cerebellar activation is remarkable and includes tasks designed to assess attention, executive control, language, working memory, learning, pain, emotion, and addiction. These data, along with the revelations about cerebro-cerebellar circuitry, provide a new framework for exploring the contribution of the cerebellum to diverse aspects of behavior. }
}
@article{WatsonetalFiSN-14,
       author = {Watson, Thomas C. and Becker, Nadine and Apps, Richard and Jones, Matthew W.},
        title = {Back to front: cerebellar connections and interactions with the prefrontal cortex},
      journal = {Frontiers in Systems Neuroscience},
         year = {2014},
    publisher = {Frontiers Media S.A.},
       volume = {8},
        pages = {4},
     abstract = {Although recent neuroanatomical evidence has demonstrated closed-loop connectivity between prefrontal cortex and the cerebellum, the physiology of cerebello-cerebral circuits and the extent to which cerebellar output modulates neuronal activity in neocortex during behavior remain relatively unexplored. We show that electrical stimulation of the contralateral cerebellar fastigial nucleus (FN) in awake, behaving rats evokes distinct local field potential (LFP) responses (onset latency {\textasciitilde}13 ms) in the prelimbic (PrL) subdivision of the medial prefrontal cortex. Trains of FN stimulation evoke heterogeneous patterns of response in putative pyramidal cells in frontal and prefrontal regions in both urethane-anesthetized and awake, behaving rats. However, the majority of cells showed decreased firing rates during stimulation and subsequent rebound increases; more than 90\% of cells showed significant changes in response. Simultaneous recording of on-going LFP activity from FN and PrL while rats were at rest or actively exploring an open field arena revealed significant network coherence restricted to the theta frequency range (5-10 Hz). Granger causality analysis indicated that this coherence was significantly directed from cerebellum to PrL during active locomotion. Our results demonstrate the presence of a cerebello-prefrontal pathway in rat and reveal behaviorally dependent coordinated network activity between the two structures, which could facilitate transfer of sensorimotor information into ongoing neocortical processing during goal directed behaviors.},
}

⁵⁴ Original papers by David Marr and James Albus on theories of learning in the cerebellar cortex plus early related work by Eccles, Itō, and Szentágothai and a more recent overview by Raymond et al in Science:

@article{AlbusMB-71,
       author = {Albus, James S.},
        title = {A Theory of Cerebellar Functions},
      journal = {Mathematical Biology},
       volume = 10,
         year = 1971,
        pages = {25-61},
}
@book{Ecclesetal1967,
        title = {The Cerebellum as a Neuronal Machine},
       author = {Eccles, J.C. and It\={o}, M. and Szent\'{a}gothai, J.},
         year = {1967},
    publisher = {Springer-Verlag}
}
@article{ItoNRN-08,
       author = {Ito, Masao},
        title = {Control of mental activities by internal models in the cerebellum},
      journal = {Nature Reviews Neuroscience},
    publisher = {Nature Publishing Group},
       volume = 9,
        issue = 4,
         year = 2008,
        pages = {304-313},
     abstract = {The intricate neuronal circuitry of the cerebellum is thought to encode internal models that reproduce the dynamic properties of body parts. These models are essential for controlling the movement of these body parts: they allow the brain to precisely control the movement without the need for sensory feedback. It is thought that the cerebellum might also encode internal models that reproduce the essential properties of mental representations in the cerebral cortex. This hypothesis suggests a possible mechanism by which intuition and implicit thought might function and explains some of the symptoms that are exhibited by psychiatric patients. This article examines the conceptual bases and experimental evidence for this hypothesis.},
}
@article{MarrJoP-69,
       author = {Marr, David},
        title = {A Theory of Cerebellar Cortex},
      journal = {Journal of Physiology},
       volume = 202,
         year = 1969,
        pages = {437-470},
}  
@article{RaymondetalSCIENCE-96,
       author = {Jennifer L. Raymond and Stephen G. Lisberger and Michael D. Mauk},
        title = {The Cerebellum: A Neuronal Learning Machine?},
      journal = {Science},
       volume = 722,
        issue = 6158,
         year = 1996,
     abstract = {The comparison of two seemingly quite different behaviors yields a surprisingly consistent picture of the role of the cerebellum in motorlearning. Behavioral and physiological data about classical conditioning of the eye lid response and motor learning in the vestibulo-ocular reflex suggest that (i) plasticity is distributed between the cerebellar cortex and the deep cerebellar nuclei; (i) the cerebellar cortex plays a special role in learning the timing of movement; and (i) the cerebellar cortex guides learning in the deep nuclei, which may for movement coordination allow learning to be transferred from the cortex to the deep nuclei. Because many of the similarities in the data from the two systems typify general features of cerebellar organization, the cerebellar mechanisms of learning in these two systems may represent principles that apply to many motor systems.}
}

⁵⁵ For your convenience, here is the citation including the abstract of the Reed and de Freitas paper that serves as one of the primary inspirations for analyzing execution traces. You might also find convenient these links relating to technologies for analyzing and generating execution traces: software tracing, Unix Kernel function tracers, control flow graph and more conventional data-flow analysis.

@article{ReedandDeFreitasCoRR-15,
       author = {Scott E. Reed and Nando de Freitas},
        title = {Neural Programmer-Interpreters},
      journal = {CoRR},
       volume = {arXiv:1511.06279},
         year = {2015},
     abstract = {We propose the neural programmer-interpreter (NPI): a recurrent and compositional neural network that learns to represent and execute programs. NPI has three learnable components: a task-agnostic recurrent core, a persistent key-value program memory, and domain-specific encoders that enable a single NPI to operate in multiple perceptually diverse environments with distinct affordances. By learning to compose lower-level programs to express higher-level programs, NPI reduces sample complexity and increases generalization ability compared to sequence-to-sequence LSTMs. The program memory allows efficient learning of additional tasks by building on existing programs. NPI can also harness the environment (e.g. a scratch pad with read-write pointers) to cache intermediate results of computation, lessening the long-term memory burden on recurrent hidden units. In this work we train the NPI with fully-supervised execution traces; each program has example sequences of calls to the immediate subprograms conditioned on the input. Rather than training on a huge number of relatively weak labels, NPI learns from a small number of rich examples. We demonstrate the capability of our model to learn several types of compositional programs: addition, sorting, and canonicalizing 3D models. Furthermore, a single NPI learns to execute these programs and all 21 associated subprograms.}
}

⁵⁶ Here is a relatively simple program written in Emacs Lisp that normalizes BibTeX entries for readability and compatibility with the different Emacs modes that I routinely use in my Emacs environment for coding, writing papers and maintaining the repository that I use for my research notes:

;; bibtex-fill-entry = formats fill according to following parameters which are required by the code below  

(defcustom bibtex-align-at-equal-sign t
  "If non-nil, align fields at equal sign instead of field text. If non-nil, the
   column for the equal sign is the value of `bibtex-text-indentation', minus 2."
  :group 'bibtex
  :type 'boolean)

(defcustom bibtex-comma-after-last-field t
  "If non-nil, a comma is put at end of last field in the entry template."
  :group 'bibtex
  :type 'boolean)

;; Offsets are based on longest field name, in this case "organization":

(defcustom bibtex-field-indentation 1
  "Starting column for the name part in BibTeX fields."
  :group 'bibtex
  :type 'integer)

(defcustom bibtex-text-indentation
  (+ bibtex-field-indentation
     (length "organization = "))
  "Starting column for the text part in BibTeX fields."
  :group 'bibtex
  :type 'integer)

(defvar bibtex-field-alignment-adjustments
  '([ " author       = " "       author = " ]
    [ " abstract     = " "     abstract = " ]
    [ " address      = " "      address = " ]
    [ " booktitle    = " "    booktitle = " ]
    [ " comment      = " "      comment = " ]
    [ " crossref     = " "     crossref = " ]
    [ " editor       = " "       editor = " ]
    [ " institution  = " "  institution = " ]
    [ " issue        = " "        issue = " ]
    [ " journal      = " "      journal = " ]
    [ " location     = " "     location = " ]    
    [ " number       = " "       number = " ]
    [ " organization = " " organization = " ]    
    [ " pages        = " "        pages = " ]
    [ " pdf          = " "          pdf = " ]
    [ " publisher    = " "    publisher = " ]
    [ " series       = " "       series = " ]
    [ " title        = " "        title = " ]
    [ " school       = " "       school = " ]
    [ " url          = " "          url = " ]
    [ " volume       = " "       volume = " ]
    [ " year         = " "         year = " ]))

;; /Applications/Emacs.app/Contents/Resources/lisp/textmodes/bibtex.el

(defun bibtex-format-fields-region ()
  (interactive)
  (bibtex-fill-entry)
  (let ((start (bibtex-beginning-of-entry)) (end (bibtex-end-of-entry)))
    (untabify start end)
    (replace-pairs-region start end bibtex-field-alignment-adjustments)))

(defun replace-pairs-region (p1 p2 pairs)
  (let ((mapping-point-seed 1000) index (mapping-points '()))
    (setq index 0)
    (while (< index (length pairs))
      (setq mapping-points (cons (number-to-string (+ mapping-point-seed index)) mapping-points ))
      (setq index (1+ index)))
    (save-excursion
      (save-restriction
        (narrow-to-region p1 p2)
        ;; Replace each find string by corresponding item in mapping-points:
        (setq index 0)
        (while (< index (length pairs))
          (goto-char (point-min))
          (while (search-forward (elt (elt pairs index) 0) nil t)
            (replace-match (elt mapping-points index) t t) )
          (setq index (1+ index)) )
        ;; Replace each mapping-point by the corresponding replacement string:
        (setq index 0)
        (while (< index (length pairs))
          (goto-char (point-min))
          (while (search-forward (elt mapping-points index) nil t)
            (replace-match (elt (elt pairs index) 1) t t) )
          (setq index (1+ index)))
        ;; Replace all double quotes used as delimiters with curly brackets:
        (goto-char (point-min))
        (while (search-forward-regexp "= \"\\([^\"]+\\)\"" nil t) (replace-match "= {\\1}" t nil)) ))))

⁵⁷ A list of BibTeX items relating to automated or assistive programming research. Gulwani et al [122] on program synthesis — foundations and trends — is added to the mix as it provides some useful background on dinfferent methods of both historical and current interest:

@inproceedings{DongandLapataACL-16,
       author = {Li Dong and Mirella Lapata},
        title = {Language to Logical Form with Neural Attention},
    booktitle = {Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics},
        pages = {33-43},
         year = {2016},
    publisher = {Association for Computational Linguistics},
      address = {Stroudsburg, PA, USA},
     abstract = {Semantic parsing aims at mapping natural language to machine interpretable meaning representations. Traditional approaches rely on high-quality lexicons, manually-built templates, and linguistic features which are either domain- or representation-specific. In this paper we present a general method based on an attention-enhanced encoder-decoder model. We encode input utterances into vector representations, and generate their logical forms by conditioning the output sequences or trees on the encoding vectors. Experimental results on four datasets show that our approach performs competitively without using hand-engineered features and is easy to adapt across domains and meaning representations.}
}
@article{LuongetalEMNLP-15,
       author = {Minh{-}Thang Luong and Hieu Pham and Christopher D. Manning},
        title = {Effective Approaches to Attention-based Neural Machine Translation},
    booktitle = {Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing},
        pages = {1412-1421},
         year = {2015},
     abstract = {An attentional mechanism has lately been used to improve neural machine translation (NMT) by selectively focusing on parts of the source sentence during translation. However, there has been little work exploring useful architectures for attention-based NMT. This paper examines two simple and effective classes of attentional mechanism: a global approach which always attends to all source words and a local one that only looks at a subset of source words at a time. We demonstrate the effectiveness of both approaches over the WMT translation tasks between English and German in both directions. With local attention, we achieve a significant gain of 5.0 BLEU points over non-attentional systems which already incorporate known techniques such as dropout. Our ensemble model using different attention architectures has established a new state-of-the-art result in the WMT'15 English to German translation task with 25.9 BLEU points, an improvement of 1.0 BLEU points over the existing best system backed by NMT and an n-gram reranker.}
}
@article{GulwanietalFaTiPL-17,
       author = {Sumit Gulwani and Oleksandr Polozov and Rishabh Singh},
        title = {Program Synthesis},
      journal = {Foundations and Trends in Programming Languages},
       volume = {4},
       number = {1-2},
         year = {2017},
        pages = {1-119},
     abstract = {Program synthesis is the task of automatically finding a program in the underlying programming language that satisfies the user intent expressed in the form of some specification. Since the inception of AI in the 1950s, this problem has been considered the holy grail of Computer Science. Despite inherent challenges in the problem such as ambiguity of user intent and a typically enormous search space of programs, the field of program synthesis has developed many different techniques that enable program synthesis in different real-life application domains. It is now used successfully in software engineering, biological discovery, computeraided education, end-user programming, and data cleaning. In the last decade, several applications of synthesis in the field of programming by examples have been deployed in mass-market industrial products. This survey is a general overview of the state-of-the-art approaches to program synthesis, its applications, and subfields. We discuss the general principles common to all modern synthesis approaches such as syntactic bias, oracle-guided inductive `xbibsearch, and optimization techniques. We then present a literature review covering the four most common state-of-the-art techniques in program synthesis: enumerative search, constraint solving, stochastic search, and deduction-based programming by examples. We conclude with a brief list of future horizons for the field.}
}
@article{GauntetalCoRR-16,
       author = {Alexander L. Gaunt and Marc Brockschmidt and Rishabh Singh and Nate Kushman and Pushmeet Kohli and Jonathan Taylor and Daniel Tarlow},
        title = {{TerpreT}: {A} Probabilistic Programming Language for Program Induction},
      journal = {CoRR},
       volume = {arXiv:1612.00817},
         year = {2016},
     abstract = {We study machine learning formulations of inductive program synthesis; that is, given input-output examples, synthesize source code that maps inputs to corresponding outputs. Our key contribution is TerpreT, a domain-specific language for expressing program synthesis problems. A TerpreT model is composed of a specification of a program representation and an interpreter that describes how programs map inputs to outputs. The inference task is to observe a set of input-output examples and infer the underlying program. From a TerpreT model we automatically perform inference using four different back-ends: gradient descent (thus each TerpreT model can be seen as defining a differentiable interpreter), linear program (LP) relaxations for graphical models, discrete satisfiability solving, and the Sketch program synthesis system. TerpreT has two main benefits. First, it enables rapid exploration of a range of domains, program representations, and interpreter models. Second, it separates the model specification from the inference algorithm, allowing proper comparisons between different approaches to inference. We illustrate the value of TerpreT by developing several interpreter models and performing an extensive empirical comparison between alternative inference algorithms on a variety of program models. To our knowledge, this is the first work to compare gradient-based search over program space to traditional search-based alternatives. Our key empirical finding is that constraint solvers dominate the gradient descent and LP-based formulations.}
}
@article{AnonymousICLR-17,
        title = {Neural Program Search: Solving Programming Tasks from Description and Examples},
       author = {Anonymous},
      journal = {International Conference on Learning Representations},
         year = {2017},
          url = {https://openreview.net/forum?id=B1KJJf-R-},
     abstract = {We present Neural Program Search, an algorithm to generate programs from natural language description and a small number of input / output examples. The algorithm combines methods from Deep Learning and Program Synthesis fields by designing rich domain-specific language (DSL) and defining efficient search algorithm guided by a Seq2Tree model on it. To evaluate the quality of the approach we also present a semi-synthetic dataset of descriptions with test examples and corresponding programs. We show that our algorithm significantly outperforms sequence-to-sequence model with attention baseline.}
}
@article{Alvarez-MelisandJaakkolaCoRR-17,
       author = {David Alvarez{-}Melis and Tommi S. Jaakkola},
        title = {A causal framework for explaining the predictions of black-box sequence-to-sequence models},
      journal = {CoRR},
       volume = {arXiv:1707.01943},
         year = {2017},
     abstract = {We interpret the predictions of any black-box structured input-structured output model around a specific input-output pair. Our method returns an "explanation" consisting of groups of input-output tokens that are causally related. These dependencies are inferred by querying the black-box model with perturbed inputs, generating a graph over tokens from the responses, and solving a partitioning problem to select the most relevant components. We focus the general approach on sequence-to-sequence problems, adopting a variational autoencoder to yield meaningful input perturbations. We test our method across several NLP sequence generation tasks.}
}
@techreport{LinetalUW-CSE-TR-17,
       author = {Xi Victoria Lin and Chenglong Wang and Deric Pang and Kevin Vu and Luke Zettlemoyer and Michael D. Ernst},
        title = {Program synthesis from natural language using recurrent neural networks},
  institution = {University of Washington Department of Computer Science and Engineering},
       number = {UW-CSE-17-03-01},
      address = {Seattle, WA, USA},
         year = {2017},
     abstract = {Oftentimes, a programmer may have difficulty implementing a desired operation. Even when the programmer can describe her goal in English, it can be difficult to translate into code. Existing resources, such as question-and-answer websites, tabulate specific operations that someone has wanted to perform in the past, but they are not effective in generalizing to new tasks, to compound tasks that require combining previous questions, or sometimes even to variations of listed tasks. Our goal is to make programming easier and more productive by letting programmers use their own words and concepts to express the intended operation, rather than forcing them to accommodate the machine by memorizing its grammar. We have built a system that lets a programmer describe a desired operation in natural language, then automatically translates it to a programming language for review and approval by the programmer. Our system, Tellina, does the translation using recurrent neural networks (RNNs), a state-of-the-art natural language processing technique that we augmented with slot (argument) filling and other enhancements. We evaluated Tellina in the context of shell scripting. We trained Tellina’s RNNs on textual descriptions of file system operations and bash one-liners, scraped from the web. Although recovering completely correct commands is challenging, Tellina achieves top-3 accuracy of 80\% for producing the correct command structure. In a controlled study, programmers who had access to Tellina outperformed those who did not, even when Tellina’s predictions were not completely correct, to a statistically significant degree.}
}

⁵⁸ In humans, the cerebellum plays an important role in motor control, and it may also be involved in some cognitive functions such as attention and language as well as in regulating fear and pleasure responses, but its movement-related functions are the most solidly established. The human cerebellum does not initiate movement, but contributes to coordination, precision, and accurate timing: it receives input from sensory systems of the spinal cord and from other parts of the brain, and integrates these inputs to fine-tune motor activity. (SOURCE)

⁵⁹ As this article on Botton rhetorically mentions, "Where else can one find a video on the philosophy of Heidegger with over 350,000 views?" and goes on to suggest that, along with a diverse collection of frequently viewed self help material, they have a dedicated and growing audience as evidenced by the adoption and use of their smartphone applications.

⁶⁰ Here are few examples from The School of Life YouTube Channel that combine clever animated cartoons and Alain de Botton's narrative to encourage emotional intelligence. They are the sort of thing I would have liked to have on hand as a faculty advisor to help distraught graduate students, often living apart from friends and family for the first time in their lives, feeling out of place in a strange culture, under a great deal of academic pressure and questioning their choices. While I don't imagine you being in particular need of such advice, I thought you might enjoy the clever drawings and offered wisdom nonetheless:

• How Emotionally Healthy Are You?    Emotional health is defined by four markers: our degree of self-love, of openness, of communication and of trust.

• Why You Shouldn't Trust Your Feelings    It can be very hard to detect just how much our judgement is constantly affected by our feelings. We should — at points — take care to be very skeptical of our first impulses.

• How to Process Your Emotions    In order to be calm and at ease with ourselves, we need regular periods where we do something rather strange-sounding: process our emotions. Here is a guide to this essential psychological move.

• The Dangers of Thinking Too Much; And Thinking Too Little    There are dangers associated both with thinking too much — and thinking too little. The trick is to use our minds to access our most sincere, authentic and original thoughts.

• Is It Better to Be Polite or Frank?    We live in an age that thinks highly of frankness and directness. But there are — nevertheless — a few reasons why politeness remains a hugely important quality.

• How to Remain Calm With People    Remaining calm around people who annoy us is one of the great life skills. It's also a teachable and learnable skill.

• How to Deal With A Crisis of Meaning    Many of us are regularly thrown off course by what we might term 'crisis of meaning'; periods when what we are up to seems not to connect up with anything purposeful or properly dignified. It's at such moments that we need to lean on a bigger theory of what meaning is, where it comes from, and how our lives relate to it.

⁶¹ Here are few more examples from The School of Life YouTube Channel covering broader social and political issues. The YouTube channel has plenty fodder for its more intellectually inclined subscribers, e.g., the History of Ideas category including Capitalism, Manners, Rituals and Religion, and the Political Theory category covering Adam Smith, Karl Marx, John Locke, John Maynard Keynes and host of other historically important thinkers. The variety reflects Alain de Botton's eclectic taste, but has expanded over the years to cover a broader set of social issues leveraging different voices and media to reach a wider, perhaps more modern audience:

• On the History of Consumerism    It’s only very recently in history that we’ve been able to buy more than the bare necessities. Can the history of consumption guide us to a wiser future?

• How to Find Fulfilling Work     The key to finding fulfilling work is to think a lot, analyse one's fears, understand the market, reflect on capitalism.

• How to Find Meaningful Work     Contrary to some expectations, it isn’t only money we want from work. We also need our work to feel ‘meaningful’. But what exactly is meaningful work, and where can we find more of it?

⁶² Here is an assortment of papers that use semantic-embedding techniques to encode computer programs for a variety of applications:

@inproceedings{ChistyakovetalICLR-17,
       author = {Alexander Chistyakov and Ekaterina Lobacheva and Arseny Kuznetsov and Alexey Romanenko},
        title = {Semantic embeddings for program behaviour patterns},
    booktitle = {ICLR Workshop},
         year = {2017},
     abstract = {In this paper, we propose a new feature extraction technique for program execution logs. First, we automatically extract complex patterns from a program's behaviour graph. Then, we embed these patterns into a continuous space by training an autoencoder. We evaluate the proposed features on a real-world malicious software detection task. We also find that the embedding space captures interpretable structures in the space of pattern parts.}
}
@article{WangetalCoRR-17,
       author = {Ke Wang and Rishabh Singh and Zhendong Su},
        title = {Dynamic Neural Program Embedding for Program Repair},
      journal = {CoRR},
       volume = {arXiv:1711.07163},
         year = {2017},
     abstract = {Neural program embeddings have shown much promise recently for a variety of program analysis tasks, including program synthesis, program repair, fault localization, etc. However, most existing program embeddings are based on syntactic features of programs, such as raw token sequences or abstract syntax trees. Unlike images and text, a program has an unambiguous semantic meaning that can be difficult to capture by only considering its syntax (i.e. syntactically similar pro- grams can exhibit vastly different run-time behavior), which makes syntax-based program embeddings fundamentally limited. This paper proposes a novel semantic program embedding that is learned from program execution traces. Our key insight is that program states expressed as sequential tuples of live variable values not only captures program semantics more precisely, but also offer a more natural fit for Recurrent Neural Networks to model. We evaluate different syntactic and semantic program embeddings on predicting the types of errors that students make in their submissions to an introductory programming class and two exercises on the CodeHunt education platform. Evaluation results show that our new semantic program embedding significantly outperforms the syntactic program embeddings based on token sequences and abstract syntax trees. In addition, we augment a search-based program repair system with the predictions obtained from our se- mantic embedding, and show that search efficiency is also significantly improved.}
}
@article{XuetalCoRR-17,
       author = {Xiaojun Xu and Chang Liu and Qian Feng and Heng Yin and Le Song and Dawn Song},
        title = {Neural Network-based Graph Embedding for Cross-Platform Binary Code Similarity Detection},
      journal = {CoRR},
       volume = {arXiv:1708.06525},
         year = {2017},
     abstract = {The problem of cross-platform binary code similarity detection aims at detecting whether two binary functions coming from different platforms are similar or not. It has many security applications, including plagiarism detection, malware detection, vulnerability search, etc. Existing approaches rely on approximate graph matching algorithms, which are inevitably slow and sometimes inaccurate, and hard to adapt to a new task. To address these issues, in this work, we propose a novel neural network-based approach to compute the embedding, i.e., a numeric vector, based on the control flow graph of each binary function, then the similarity detection can be done efficiently by measuring the distance between the embeddings for two functions. We implement a prototype called Gemini. Our extensive evaluation shows that Gemini outperforms the state-of-the-art approaches by large margins with respect to similarity detection accuracy. Further, Gemini can speed up prior art's embedding generation time by 3 to 4 orders of magnitude and reduce the required training time from more than 1 week down to 30 minutes to 10 hours. Our real world case studies demonstrate that Gemini can identify significantly more vulnerable firmware images than the state-of-the-art, i.e., Genius. Our research showcases a successful application of deep learning on computer security problems.}
}
@article{PiechetalCoRR-15,
       author = {Chris Piech and Jonathan Huang and Andy Nguyen and Mike Phulsuksombati and Mehran Sahami and Leonidas J. Guibas},
        title = {Learning Program Embeddings to Propagate Feedback on Student Code},
      journal = {CoRR},
       volume = {arXiv:1505.05969},
         year = {2015},
     abstract = {Providing feedback, both assessing final work and giving hints to stuck students, is difficult for open-ended assignments in massive online classes which can range from thousands to millions of students. We introduce a neural network method to encode programs as a linear mapping from an embedded precondition space to an embedded postcondition space and propose an algorithm for feedback at scale using these linear maps as features. We apply our algorithm to assessments from the Code.org Hour of Code and Stanford University's CS1 course, where we propagate human comments on student assignments to orders of magnitude more submissions.}
}
@inproceedings{BalogetalICLR-17,
       author = {Matej Balog and Alexander L. Gaunt and Marc Brockschmidt and Sebastian Nowozin and Daniel Tarlow},
        title = {DeepCoder: Learning to Write Programs},
    booktitle = {International Conference on Learning Representations},
         year = 2017,
     abstract = {We develop a first line of attack for solving programming competition-style problems from input-output examples using deep learning. The approach is to train a neural network to predict properties of the program that generated the outputs from the inputs. We use the neural network's predictions to augment search techniques from the programming languages community, including enumerative search and an SMT-based solver. Empirically, we show that our approach leads to an order of magnitude speedup over the strong non-augmented baselines and a Recurrent Neural Network approach, and that we are able to solve problems of difficulty comparable to the simplest problems on programming competition websites.}
}
@article{AllamanisetalCoRR-17,
       author = {Miltiadis Allamanis and Marc Brockschmidt and Mahmoud Khademi},
        title = {Learning to Represent Programs with Graphs},
      journal = {CoRR},
       volume = {arXiv:1711.00740},
         year = {2017},
     abstract = {Learning tasks on source code (i.e., formal languages) have been considered recently, but most work has tried to transfer natural language methods and does not capitalize on the unique opportunities offered by code's known syntax. For example, long-range dependencies induced by using the same variable or function in distant locations are often not considered. We propose to use graphs to represent both the syntactic and semantic structure of code and use graph-based deep learning methods to learn to reason over program structures.\\
                 In this work, we present how to construct graphs from source code and how to scale Gated Graph Neural Networks training to such large graphs. We evaluate our method on two tasks: VarNaming, in which a network attempts to predict the name of a variable given its usage, and VarMisuse, in which the network learns to reason about selecting the correct variable that should be used at a given program location. Our comparison to methods that use less structured program representations shows the advantages of modeling known structure, and suggests that our models learn to infer meaningful names and to solve the VarMisuse task in many cases. Additionally, our testing showed that VarMisuse identifies a number of bugs in mature open-source projects.}
}

⁶³ Here's the BibTeX entry for a recent survey paper Gulwanietal [122] on current trends in program synthesis that Rishabh Singh co-authored:

@article{GulwanietalFaTiPL-17,
       author = {Sumit Gulwani and Oleksandr Polozov and Rishabh Singh},
        title = {Program Synthesis},
      journal = {Foundations and Trends in Programming Languages},
       volume = {4},
       number = {1-2},
         year = {2017},
        pages = {1-119},
     abstract = {Program synthesis is the task of automatically finding a program in the underlying programming language that satisfies the user intent expressed in the form of some specification. Since the inception of AI in the 1950s, this problem has been considered the holy grail of Computer Science. Despite inherent challenges in the problem such as ambiguity of user intent and a typically enormous search space of programs, the field of program synthesis has developed many different techniques that enable program synthesis in different real-life application domains. It is now used successfully in software engineering, biological discovery, computer-aided education, end-user programming, and data cleaning. In the last decade, several applications of synthesis in the field of programming by examples have been deployed in mass-market industrial products. This survey is a general overview of the state-of-the-art approaches to program synthesis, its applications, and subfields. We discuss the general principles common to all modern synthesis approaches such as syntactic bias, oracle-guided inductive search, and optimization techniques. We then present a literature review covering the four most common state-of-the-art techniques in program synthesis: enumerative search, constraint solving, stochastic search, and deduction-based programming by examples. We conclude with a brief list of future horizons for the field.}
}  

⁶⁴ Collection of papers relating to the role of the cerebellum and basal ganglia in language, planning and simulating complex activity:

@book{Lieberman02,
     author = {Philip Lieberman},
      title = {Human language and our reptilian brain: The subcortical bases for speech, syntax and thought},
  publisher = {Harvard University Press},
    address = {Cambridge, MA},
       year = 2002,
}
@article{LiebermanAJPT-02,
     author = {Philip Lieberman},
      title = {On the nature and evolution of the neural bases of human language},
    Journal = {American Journal of Physical Anthropology},
      month = {December},
     volume = 119,
      issue = {S35},
      pages = {36-62},
       year = 2002,
   abstract = {The traditional theory equating the brain bases of language with Broca's and Wernicke's neocortical areas is wrong. Neural circuits linking activity in anatomically segregated populations of neurons in subcortical structures and the neocortex throughout the human brain regulate complex behaviors such as walking, talking, and comprehending the meaning of sentences. When we hear or read a word, neural structures involved in the perception or real-world associations of the word are activated as well as posterior cortical regions adjacent to Wernicke's area. Many areas of the neocortex and subcortical structures support the cortical-striatal-cortical circuits that confer complex syntactic ability, speech production, and a large vocabulary. However, many of these structures also form part of the neural circuits regulating other aspects of behavior. For example, the basal ganglia, which regulate motor control, are also crucial elements in the circuits that confer human linguistic ability and abstract reasoning. The cerebellum, traditionally associated with motor control, is active in motor learning. The basal ganglia are also key elements in reward-based learning. Data from studies of Broca's aphasia, Parkinson's disease, hypoxia, focal brain damage, and a genetically transmitted brain anomaly (the putative ``language gene,'' family KE), and from comparative studies of the brains and behavior of other species, demonstrate that the basal ganglia sequence the discrete elements that constitute a complete motor act, syntactic process, or thought process. Imaging studies of intact human subjects and electrophysiologic and tracer studies of the brains and behavior of other species confirm these findings. As Dobzansky put it, ``Nothing in biology makes sense except in the light of evolution'' (cited in Mayr, 1982). That applies with as much force to the human brain and the neural bases of language as it does to the human foot or jaw. The converse follows: the mark of evolution on the brains of human beings and other species provides insight into the evolution of the brain bases of human language. The neural substrate that regulated motor control in the common ancestor of apes and humans most likely was modified to enhance cognitive and linguistic ability. Speech communication played a central role in this process. However, the process that ultimately resulted in the human brain may have started when our earliest hominid ancestors began to walk.},
}
@article{MerkerBBS-07,
       author = {Merker, Bjorn},
        title = {Consciousness without a cerebral cortex: A challenge for neuroscience and medicine},
      journal = {Behavioral and Brain Sciences},
    publisher = {Cambridge University Press},
       volume = {30},
       number = {1},
         year = {2007},
        pages = {63–81},
     abstract = {A broad range of evidence regarding the functional organization of the vertebrate brain - spanning from comparative neurology to experimental psychology and neurophysiology to clinical data - is reviewed for its bearing on conceptions of the neural organization of consciousness. A novel principle relating target selection, action selection, and motivation to one another, as a means to optimize integration for action in real time, is introduced. With its help, the principal macrosystems of the vertebrate brain can be seen to form a centralized functional design in which an upper brain stem system organized for conscious function performs a penultimate step in action control. This upper brain stem system retained a key role throughout the evolutionary process by which an expanding forebrain - culminating in the cerebral cortex of mammals - came to serve as a medium for the elaboration of conscious contents. This highly conserved upper brainstem system, which extends from the roof of the midbrain to the basal diencephalon, integrates the massively parallel and distributed information capacity of the cerebral hemispheres into the limited-capacity, sequential mode of operation required for coherent behavior. It maintains special connective relations with cortical territories implicated in attentional and conscious functions, but is not rendered nonfunctional in the absence of cortical input. This helps explain the purposive, goal-directed behavior exhibited by mammals after experimental decortication, as well as the evidence that children born without a cortex are conscious. Taken together these circumstances suggest that brainstem mechanisms are integral to the constitution of the conscious state, and that an adequate account of neural mechanisms of conscious function cannot be confined to the thalamocortical complex alone.}
}
@article{NottebohmPLoS-05,
       author = {Nottebohm, Fernando},
      journal = {{PLOS} Biology},
    publisher = {Public Library of Science},
        title = {The Neural Basis of Birdsong},
         year = {2005},
       volume = {3},
     abstract = {There is a tradition in biology of using specific animal models to study generalizable basic properties of a system. For example, the giant axon of squid was used for the pioneering work on nerve transmission; the fruit fly (Drosophila) has played a key role in researchers discovering the role of homeobox genes in embryogenesis; the sea slug (Aplysia) is used to study the molecular biology of learning; and the round worm (Caenorhabditis elegans) is used to study programmed cell death. Basic insights gained from these four systems apply widely to other multicellular animals. Here, I will review basic discoveries made by studying birdsong that have helped answer more general questions in vertebrate neuroscience.},
}
@misc{PattaniBRAINLESS-16,
        title = {Preliminary Research -- Living Without Brain Structures},
       author = {Aneri Pattani},
         year = {2016},
 howpublished = {StoryLab Posting},
          url = {http://www.northeastern.edu/storylab/sp16/2016/01/27/preliminary-research-living-without-brain-structures/},
     abstract = {How much of our brain do we actually need? A number of stories have appeared in the news in recent months about people with chunks of their brains missing or damaged. These cases tell a story about the mind that goes deeper than their initial shock factor. It isn't just that we don't understand how the brain works, but that we may be thinking about it in the entirely wrong way.}
}
@article{Stephenson-JonesetalCURRENT-BIOLOGY-11,
       author = {Stephenson-Jones, M. and Samuelsson, E. and Ericsson, J. and Robertson, B. and Grillner, S.},
        title = {Evolutionary conservation of the basal ganglia as a common vertebrate mechanism for action selection},
      journal = {Current Biology},
         year = {2011},
       volume = {21},
       number = {13},
        pages = {1081-1091},
     abstract = {Although the basal ganglia are thought to play a key role in action selection in mammals, it is unknown whether this mammalian circuitry is present in lower vertebrates as a conserved selection mechanism. We aim here, using lamprey, to elucidate the basal ganglia circuitry in the phylogenetically oldest group of vertebrates (cyclostomes) and determine how this selection architecture evolved to accommodate the increased behavioral repertoires of advanced vertebrates.\\ We show, using immunohistochemistry, tract tracing, and whole-cell recordings, that all parts of the mammalian basal ganglia (striatum, globus pallidus interna [GPi] and externa [GPe], and subthalamic nucleus [STN]) are present in the lamprey forebrain. In addition, the circuit features, molecular markers, and physiological activity patterns are conserved. Thus, GABAergic striatal neurons expressing substance P project directly to the pallidal output layer, whereas enkephalin-expressing striatal neurons project indirectly via nuclei homologous to the GPe and STN. Moreover, pallidal output neurons tonically inhibit tectum, mesencephalic, and diencephalic motor regions.\\ These results show that the detailed basal ganglia circuitry is present in the phylogenetically oldest vertebrates and has been conserved, most likely as a mechanism for action selection used by all vertebrates, for over 560 million years. Our data also suggest that the mammalian basal ganglia evolved through a process of exaptation, where the ancestral core unit has been co-opted for multiple functions, allowing them to process cognitive, emotional, and motor information in parallel and control a broader range of behaviors.}
} 

⁶⁵ The field of automated planning in AI in the 80's was often criticized for focusing on toy problems such as the blocks world. The criticism was leveled by disciplines like operations research and applied control theory that spanned a wide range of real-world problems including inventory management, vehicle routing, chemical process control and robotic assembly, whereas the blocks world was so general it could be used to solve any — AI Complete and NP-hard — symbolic problem.

⁶⁶ Prior work listed on this slide deck presented by Dan Abolafia joint work with Quoc Le and Mohammad Norouzi. Dan introduced this project as part of the code synthesis moonshot, stating that "[w]e want to build an ML system that can learn to write code".

@article{BalogetalCoRR-16,
       author = {Matej Balog and Alexander L. Gaunt and Marc Brockschmidt and Sebastian Nowozin and Daniel Tarlow},
        title = {{DeepCoder}: Learning to Write Programs},
      journal = {CoRR},
       volume = {arXiv:1611.01989},
         year = {2016},
     abstract = {We develop a first line of attack for solving programming competition-style problems from input-output examples using deep learning. The approach is to train a neural network to predict properties of the program that generated the outputs from the inputs. We use the neural network's predictions to augment search techniques from the programming languages community, including enumerative search and an SMT-based solver. Empirically, we show that our approach leads to an order of magnitude speedup over the strong non-augmented baselines and a Recurrent Neural Network approach, and that we are able to solve problems of difficulty comparable to the simplest problems on programming competition websites.}
}
@article{DevlinetalCoRR-17,
       author = {Jacob Devlin and Jonathan Uesato and Surya Bhupatiraju and Rishabh Singh and Abdel{-}rahman Mohamed and Pushmeet Kohli},
        title = {{RobustFill}: Neural Program Learning under Noisy {I/O}},
      journal = {CoRR},
       volume = {arXiv:1703.07469},
         year = {2017},
     abstract = {The problem of automatically generating a computer program from some specification has been studied since the early days of AI. Recently, two competing approaches for automatic program learning have received significant attention: (1) neural program synthesis, where a neural network is conditioned on input/output (I/O) examples and learns to generate a program, and (2) neural program induction, where a neural network generates new outputs directly using a latent program representation. Here, for the first time, we directly compare both approaches on a large-scale, real-world learning task. We additionally contrast to rule-based program synthesis, which uses hand-crafted semantics to guide the program generation. Our neural models use a modified attention RNN to allow encoding of variable-sized sets of I/O pairs. Our best synthesis model achieves 92\% accuracy on a real-world test set, compared to the 34\% accuracy of the previous best neural synthesis approach. The synthesis model also outperforms a comparable induction model on this task, but we more importantly demonstrate that the strength of each approach is highly dependent on the evaluation metric and end-user application. Finally, we show that we can train our neural models to remain very robust to the type of noise expected in real-world data (e.g., typos), while a highly-engineered rule-based system fails entirely.}
}
@article{LiangetalCoRR-16,
       author = {Chen Liang and Jonathan Berant and Quoc Le and Kenneth D. Forbus and Ni Lao},
        title = {Neural Symbolic Machines: Learning Semantic Parsers on Freebase with Weak Supervision},
      journal = {CoRR},
       volume = {arXiv:1611.00020},
         year = {2016},
     abstract = {Harnessing the statistical power of neural networks to perform language understanding and symbolic reasoning is difficult, when it requires executing efficient discrete operations against a large knowledge-base. In this work, we introduce a Neural Symbolic Machine (NSM), which contains (a) a neural "programmer”, i.e., a sequence-to-sequence model that maps language utterances to programs and utilizes a key-variable memory to handle compositionality (b) a symbolic "computer”, i.e., a Lisp interpreter that performs program execution, and helps find good programs by pruning the search space. We apply REINFORCE to directly optimize the task reward of this structured prediction problem. To train with weak supervision and improve the stability of REINFORCE we augment it with an iterative maximum-likelihood training process. NSM outperforms the state-of-the-art on the WEBQUESTIONSSP dataset when trained from question-answer pairs only, without requiring any feature engineering or domain-specific knowledge.}
}
@article{NeelakantanetalCoRR-15,
       author = {Arvind Neelakantan and Quoc V. Le and Ilya Sutskever},
        title = {Neural Programmer: Inducing Latent Programs with Gradient Descent},
      journal = {CoRR},
       volume = {arXiv:1511.04834},
         year = {2015},
     abstract = {Deep neural networks have achieved impressive supervised classification performance in many tasks including image recognition, speech recognition, and sequence to sequence learning. However, this success has not been translated to applications like question answering that may involve complex arithmetic and logic reasoning. A major limitation of these models is in their inability to learn even simple arithmetic and logic operations. For example, it has been shown that neural networks fail to learn to add two binary numbers reliably. In this work, we propose Neural Programmer, an end-to-end differentiable neural network augmented with a small set of basic arithmetic and logic operations. Neural Programmer can call these augmented operations over several steps, thereby inducing compositional programs that are more complex than the built-in operations. The model learns from a weak supervision signal which is the result of execution of the correct program, hence it does not require expensive annotation of the correct program itself. The decisions of what operations to call, and what data segments to apply to are inferred by Neural Programmer. Such decisions, during training, are done in a differentiable fashion so that the entire network can be trained jointly by gradient descent. We find that training the model is difficult, but it can be greatly improved by adding random noise to the gradient. On a fairly complex synthetic table-comprehension dataset, traditional recurrent networks and attentional models perform poorly while Neural Programmer typically obtains nearly perfect accuracy.}
}
@article{ReedandDeFreitasCoRR-15,
       author = {Scott E. Reed and Nando de Freitas},
        title = {Neural Programmer-Interpreters},
      journal = {CoRR},
       volume = {arXiv:1511.06279},
         year = {2015},
     abstract = {We propose the neural programmer-interpreter (NPI): a recurrent and compositional neural network that learns to represent and execute programs. NPI has three learnable components: a task-agnostic recurrent core, a persistent key-value program memory, and domain-specific encoders that enable a single NPI to operate in multiple perceptually diverse environments with distinct affordances. By learning to compose lower-level programs to express higher-level programs, NPI reduces sample complexity and increases generalization ability compared to sequence-tosequence LSTMs. The program memory allows efficient learning of additional tasks by building on existing programs. NPI can also harness the environment (e.g. a scratch pad with read-write pointers) to cache intermediate results of computation, lessening the long-term memory burden on recurrent hidden units. In this work we train the NPI with fully-supervised execution traces; each program has example sequences of calls to the immediate subprograms conditioned on the input. Rather than training on a huge number of relatively weak labels, NPI learns from a small number of rich examples. We demonstrate the capability of our model to learn several types of compositional programs: addition, sorting, and canonicalizing 3D models. Furthermore, a single NPI learns to execute these programs and all 21 associated subprograms.}
}  
@inproceedings{GuadarramaetalICRAS-13,
       author = {Sergio Guadarrama and Lorenzo Riano and Dave Golland and Daniel Gouhring and Yangqing Jia and Dan Klein and Pieter Abbeel and Trevor Darrell},
        title = {Grounding spatial relations for human-robot interaction},
    booktitle = {2013 {IEEE/RSJ} International Conference on Intelligent Robots and Systems},
         year = 2013,
        pages = {1640-1647},
     abstract = {We propose a system for human-robot interaction that learns both models for spatial prepositions and for object recognition. Our system grounds the meaning of an input sentence in terms of visual percepts coming from the robot’s sensors in order to send an appropriate command to the PR2 or respond to spatial queries. To perform this grounding, the system recognizes the objects in the scene, determines which spatial relations hold between those objects, and semantically parses the input sentence. The proposed system uses the visual and spatial information in conjunction with the semantic parse to interpret statements that refer to objects (nouns), their spatial relationships (prepositions), and to execute commands (actions). The semantic parse is inherently compositional, allowing the robot to understand complex commands that refer to multiple objects and relations such as: "Move the cup close to the robot to the area in front of the plate and behind the tea box". Our system correctly parses 94\% of the 210 online test sentences, correctly interprets 91\% of the correctly parsed sentences, and correctly executes 89\% of the correctly interpreted sentences.}
}  
@inproceedings{TamaretalNIPS-16,
       author = {Aviv Tamar and Yi Wu and Garrett Thomas and Sergey Levine and and Pieter Abbeel},
        title = {Value Iteration Networks},
    booktitle = {Proceedings of the 30th International Conference on Neural Information Processing Systems},
    publisher = {Curran Associates Inc.},
         year = {2016},
     location = {Barcelona, Spain},
     abstract = {We introduce the value iteration network (VIN): a fully differentiable neural network with a ‘planning module’ embedded within. VINs can learn to plan, and are suitable for predicting outcomes that involve planning-based reasoning, such as policies for reinforcement learning. Key to our approach is a novel differentiable approximation of the value-iteration algorithm, which can be represented as a convolutional neural network, and trained end-to-end using standard backpropagation. We evaluate VIN based policies on discrete and continuous path-planning domains, and on a natural-language based search task. We show that by learning an explicit planning computation, VIN policies generalize better to new, unseen domains.},
}

⁶⁷ Recent papers from DeepMind on model-based planning that focus on how an agent might construct / imagine an action occuring in a given situation to produce a given outcome:

@article{PascanuCoRR-17,
       author = {Razvan Pascanu and Yujia Li and Oriol Vinyals and Nicolas Heess and Lars Buesing and S{\'{e}}bastien Racani{\`{e}}re and David P. Reichert and Theophane Weber and Daan Wierstra and Peter Battaglia},
        title = {Learning model-based planning from scratch},
      journal = {CoRR},
       volume = {arXiv:1707.06170},
         year = {2017},
     abstract = {Conventional wisdom holds that model-based planning is a powerful approach to sequential decision-making. It is often very challenging in practice, however, because while a model can be used to evaluate a plan, it does not prescribe how to construct a plan. Here we introduce the "Imagination-based Planner", the first model-based, sequential decision-making agent that can learn to construct, evaluate, and execute plans. Before any action, it can perform a variable number of imagination steps, which involve proposing an imagined action and evaluating it with its model-based imagination. All imagined actions and outcomes are aggregated, iteratively, into a "plan context" which conditions future real and imagined actions. The agent can even decide how to imagine: testing out alternative imagined actions, chaining sequences of actions together, or building a more complex "imagination tree" by navigating flexibly among the previously imagined states using a learned policy. And our agent can learn to plan economically, jointly optimizing for external rewards and computational costs associated with using its imagination. We show that our architecture can learn to solve a challenging continuous control problem, and also learn elaborate planning strategies in a discrete maze-solving task. Our work opens a new direction toward learning the components of a model-based planning system and how to use them.}
}  
@article{WeberetalCoRR-17,
       author = {Theophane Weber and  S{\'{e}}bastien Racani{\`{e}}re and David P. Reichert and Lars Buesing and Arthur Guez and Danilo Jimenez Rezende and Adri{\`{a}} Puigdom{\`{e}}nech Badia and Oriol Vinyals and Nicolas Heess and Yujia Li and Razvan Pascanu and Peter Battaglia and David Silver and Daan Wierstra},
        title = {Imagination-Augmented Agents for Deep Reinforcement Learning},
      journal = {CoRR},
       volume = {arXiv:/1707.06203},
         year = {2017},
     abstract = {We introduce Imagination-Augmented Agents (I2As), a novel architecture for deep reinforcement learning combining model-free and model-based aspects. In contrast to most existing model-based reinforcement learning and planning methods, which prescribe how a model should be used to arrive at a policy, I2As learn to interpret predictions from a learned environment model to construct implicit plans in arbitrary ways, by using the predictions as additional context in deep policy networks. I2As show improved data efficiency, performance, and robustness to model misspecification compared to several baselines.}
}
@article{SilveretalCoRR-17,
       author = {David Silver and Hado van Hasselt and Matteo Hessel and Tom Schaul and Arthur Guez and Tim Harley and Gabriel Dulac{-}Arnold and David P. Reichert and Neil C. Rabinowitz and Andr{\'{e}} Barreto and Thomas Degris},
        title = {The Predictron: End-To-End Learning and Planning},
      journal = {CoRR},
       volume = {arXiv:1612.08810},
      comment = {ICML 2017 preprint},
         year = {2017},
     abstract = {One of the key challenges of artificial intelligence is to learn models that are effective in the context of planning. In this document we introduce the predictron architecture. The predictron consists of a fully abstract model, represented by a Markov reward process, that can be rolled forward multiple "imagined" planning steps. Each forward pass of the predictron accumulates internal rewards and values over multiple planning depths. The predictron is trained end-to-end so as to make these accumulated values accurately approximate the true value function. We applied the predictron to procedurally generated random mazes and a simulator for the game of pool. The predictron yielded significantly more accurate predictions than conventional deep neural network architectures.}
}

⁶⁸ Collection of papers on attention, imagination and executive control from various researchers in cognitive neuroscience:

@article{HassabisetalNEURON-17,
       author = {Demis Hassabis and Dharshan Kumaran and Christopher Summerfield and Matthew Botvinick},
        title = {Neuroscience-Inspired Artificial Intelligence},
      journal = {Neuron},
       volume = {95},
       number = {2},
        pages = {245-258},
         year = {2017},
     abstract = {The fields of neuroscience and artificial intelligence (AI) have a long and intertwined history. In more recent times, however, communication and collaboration between the two fields has become less commonplace. In this article, we argue that better understanding biological brains could play a vital role in building intelligent machines. We survey historical interactions between the AI and neuroscience fields and emphasize current advances in AI that have been inspired by the study of neural computation in humans and other animals. We conclude by highlighting shared themes that may be key for advancing future research in both fields.},
}
@article{HassabisandMaguireTiCS-07,
        title = {Deconstructing episodic memory with construction},
       author = {Hassabis, Demis and Maguire, Eleanor A.},
      journal = {Trends in Cognitive Science},
    publisher = {Elsevier Current Trends},
       volume = 11,
        issue = 7
         year = 2007,
        pages = {299-306},
     abstract = {It has recently been observed that the brain network supporting recall of episodic memories shares much in common with other cognitive functions such as episodic future thinking, navigation and theory of mind. It has been speculated that ‘self-projection’ is the key common process. However, in this Opinion article, we note that other functions (e.g. imagining fictitious experiences) not explicitly connected to either the self or a subjective sense of time, activate a similar brain network. Hence, we argue that the process of ‘scene construction’ is better able to account for the commonalities in the brain areas engaged by an extended range of disparate functions. In light of this, we re-evaluate our understanding of episodic memory, the processes underpinning it and other related cognitive functions.} 
}
@article{HassabisandMaguirePTRS_B-09,
        title = {The construction system of the brain},
       author = {Hassabis, Demis and Maguire, Eleanor A.},
      journal = {Philosophical Transactions of the Royal Society London B Biological Science},
    publisher = {The Royal Society},
       volume = 364,
        issue = 1521,
         year = 2009,
        pages = {1263-1271},
     abstract = {The ability to construct a hypothetical situation in one's imagination prior to it actually occurring may afford greater accuracy in predicting its eventual outcome. The recollection of past experiences is also considered to be a reconstructive process with memories recreated from their component parts. Construction, therefore, plays a critical role in allowing us to plan for the future and remember the past. Conceptually, construction can be broken down into a number of constituent processes although little is known about their neural correlates. Moreover, it has been suggested that some of these processes may be shared by a number of other cognitive functions including spatial navigation and imagination. Recently, novel paradigms have been developed that allow for the isolation and characterization of these underlying processes and their associated neuroanatomy. Here, we selectively review this fast-growing literature and consider some implications for remembering the past and predicting the future.},
}
@article{PritzeletalCoRR-17,
       author = {Alexander Pritzel and Benigno Uria and  Sriram Srinivasan and Adri\`{a} Puigdom\`{e}nech Badia and Oriol Vinyals and Demis Hassabis and Daan Wierstra and Charles Blundell},
        title = {Neural Episodic Control},
      journal = {CoRR},
       volume = {arXiv:1703.01988},
         year = {2017},
     abstract = {Deep reinforcement learning methods attain super-human performance in a wide range of environments. Such methods are grossly inefficient, often taking orders of magnitudes more data than humans to achieve reasonable performance. We propose Neural Episodic Control: a deep reinforcement learning agent that is able to rapidly assimilate new experiences and act upon them. Our agent uses a semi-tabular representation of the value function: a buffer of past experience containing slowly changing state representations and rapidly updated estimates of the value function. We show across a wide range of environments that our agent learns significantly faster than other state-of-the-art, general purpose deep reinforcement learning agents.}
}
@article{OReillyNC-96,
       author = {O'Reilly, Randall C.},
        title = {Biologically Plausible Error-driven Learning Using Local Activation Differences: The Generalized Recirculation Algorithm},
      journal = {Neural Computation},
    publisher = {MIT Press},
      address = {Cambridge, MA, USA},
       volume = {8},
       number = {5},
         year = {1996},
        pages = {895-938},
     abstract = {The error backpropagation learning algorithm (BP) is generally considered biologically implausible because it does not use locally available, activation-based variables. A version of BP that can be computed locally using bidirectional activation recirculation (Hinton and McClelland 1988) instead of backpropagated error derivatives is more biologically plausible. This paper presents a generalized version of the recirculation algorithm (GeneRec), which overcomes several limitations of the earlier algorithm by using a generic recurrent network with sigmoidal units that can learn arbitrary input/output mappings. However, the contrastive Hebbian learning algorithm (CHL, also known as DBM or mean field learning) also uses local variables to perform error-driven learning in a sigmoidal recurrent network. CHL was derived in a stochastic framework (the Boltzmann machine), but has been extended to the deterministic case in various ways, all of which rely on problematic approximations and assumptions, leading some to conclude that it is fundamentally flawed. This paper shows that CHL can be derived instead from within the BP framework via the GeneRec algorithm. CHL is a symmetry-preserving version of GeneRec that uses a simple approximation to the midpoint or second-order accurate Runge-Kutta method of numerical integration, which explains the generally faster learning speed of CHL compared to BI. Thus, all known fully general error-driven learning algorithms that use local activation-based variables in deterministic networks can be considered variations of the GeneRec algorithm (and indirectly, of the backpropagation algorithm). GeneRec therefore provides a promising framework for thinking about how the brain might perform error-driven learning. To further this goal, an explicit biological mechanism is proposed that would be capable of implementing GeneRec-style learning. This mechanism is consistent with available evidence regarding synaptic modification in neurons in the neocortex and hippocampus, and makes further predictions.}
}
@article{OReillySCIENCE-06,
     title = {Biologically Based Computational Models of High-Level Cognition},
    author = {O'Reilly, Randall C.},
   journal = {Science},
    volume = 314,
     issue = 5796,
      year = 2006,
     pages = {91-94},
  abstract = {Computer models based on the detailed biology of the brain can help us understand the myriad complexities of human cognition and intelligence. Here, we review models of the higher level aspects of human intelligence, which depend critically on the prefrontal cortex and associated subcortical areas. The picture emerging from a convergence of detailed mechanistic models and more abstract functional models represents a synthesis between analog and digital forms of computation. Specifically, the need for robust active maintenance and rapid updating of information in the prefrontal cortex appears to be satisfied by bistable activation states and dynamic gating mechanisms. These mechanisms are fundamental to digital computers and may be critical for the distinctive aspects of human intelligence.},
}
@article{OReillyTiCS-98,
       author = {O'Reilly, R. C.},
        title = {{{S}ix principles for biologically based computational models of cortical cognition}},
      journal = {Trends Cognitive Science)},
         year = {1998},
       volume = {2},
       number = {11},
        pages = {455-462},
     abstract = {This review describes and motivates six principles for computational cognitive neuroscience models: biological realism, distributed representations, inhibitory competition, bidirectional activation propagation, error-driven task learning, and Hebbian model learning. Although these principles are supported by a number of cognitive, computational and biological motivations, the prototypical neural-network model (a feedforward back-propagation network) incorporates only two of them, and no widely used model incorporates all of them. It is argued here that these principles should be integrated into a coherent overall framework, and some potential synergies and conflicts in doing so are discussed.}
}
@article{OReillyandFrankNC-06,
       author = {O'Reilly, Randall C. and Frank, Michael J.},
        title = {Making Working Memory Work: A Computational Model of Learning in the Prefrontal Cortex and Basal Ganglia},
      journal = {Neural Computation},
       volume = 18,
        issue = 2,
         year = 2006,
        pages = {283-328},
          url = {http://psych.colorado.edu/~oreilly/papers/OReillyFrank06_pbwm.pdf},
    publisher = {MIT Press},
      address = {Cambridge, MA, USA},
     abstract = {The prefrontal cortex has long been thought to subserve both working memory (the holding of information online for processing) and executive functions (deciding how to manipulate working memory and perform processing). Although many computational models of working memory have been developed, the mechanistic basis of executive function remains elusive, often amounting to a homunculus. This article presents an attempt to deconstruct this homunculus through powerful learning mechanisms that allow a computational model of the prefrontal cortex to control both itself and other brain areas in a strategic, task-appropriate manner. These learning mechanisms are based on subcortical structures in the midbrain, basal ganglia, and amygdala, which together form an actor-critic architecture. The critic system learns which prefrontal representations are task relevant and trains the actor, which in turn provides a dynamic gating mechanism for controlling working memory updating. Computationally, the learning mechanism is designed to simultaneously solve the temporal and structural credit assignment problems. The model's performance compares favorably with standard backpropagation-based temporal learning mechanisms on the challenging 1-2-AX working memory task and other benchmark working memory tasks.}
}
@book{OReillyandMunakata00,
     title = {Computational Explorations in Cognitive Neuroscience: Understanding the Mind by Simulating the Brain},
    author = {Randall O'Reilly and Yuko Munakata},
 publisher = {MIT Press},
   address = {Cambridge, Massachusetts},
      year = 2000,
}
@article{OReillyetalCON-10,
     title = {Computational models of cognitive control},
    author = {Randall C. O'Reilly and Seth A. Herd and Wolfgang M. Pauli},
   journal = {Current Opinion in Neurobiology},
    volume = 20,
     issue = 2,
      year = 2010,
     pages = {257-261},
}
@article{RosenfeldetalCoRR-17,
        title = {Priming Neural Networks},
       author = {Amir Rosenfeld, Mahdi Biparva, John K. Tsotsos},
      journal = {CoRR},
       volume = {arXiv:1711.05918},
         year = 2017,
     abstract = {Visual priming is known to affect the human visual system to allow detection of scene elements, even those that may have been near unnoticeable before, such as the presence of camouflaged animals. This process has been shown to be an effect of top-down signaling in the visual system triggered by the said cue. In this paper, we propose a mechanism to mimic the process of priming in the context of object detection and segmentation. We view priming as having a modulatory, cue dependent effect on layers of features within a network. Our results show how such a process can be complementary to, and at times more effective than simple post-processing applied to the output of the network, notably so in cases where the object is hard to detect such as in severe noise. Moreover, we find the effects of priming are sometimes stronger when early visual layers are affected. Overall, our experiments confirm that top-down signals can go a long way in improving object detection and segmentation.}
}

⁶⁹ In a recent discussion, Christof mentioned that AIBS scientists regularly work with both mouse and human samples of neural tissue and that, apart from their size, the cortical tissue from mice and humans are virtually identical to trained anatomists and histologists. This makes him wonder whether mice would develop, say, consciousness or more sophisticated signaling if they had a cortex as large as ours. As far as we know, there is nothing special about spindle / von Economo neurons or the elusive mirror neurons, except that the — the former — are larger and are noteworthy for their size and location. Simple neural network architectures including convolutional networks work efficiently to learn structure, induce sparse feature libraries, construct concept hierarchies, and can even learn value functions to support reinforcement learning and attentional mechanisms.

⁷⁰ This is another example of progress primarily due to improvements in computing hardware. The basic idea gets invented a dozen times and never gains a purchase until Moore’s law catches up enabling lots of researchers to essentially “play” in the conceptual space that opens up due to the confluence of a good idea and adequate computing resources. It's not just the length of time that an engineer has to sit on his or her hands waiting while some mysterious process plays out over an indeterminate interval of time and, at the end, provides an inscrutable and generally unsatisfying answer, it's the anxiety, frustration and helplessness that this induces in the engineer. When you can launch millions of experiments and get millions of answers in the same or a shorter amount of time, everything changes. An anomaly in the result of your only experiment is a problem, while an anomaly in one result out of millions is an outlier or a bug in your experimental design.

⁷¹ Here is a sample of recent review papers relating to the idea of dual-loop models that integrate sensorimotor maps with semantic processing:

@incollection{WeilleretalNL-16,
       author = {Cornelius Weiller and Tobias Bormann and Dorothee Kuemmerer and Mariachristina Musso and Michel Rijntjes},
        title = {The Dual Loop Model in Language},
       editor = {Hickok, Gregory  and Small, Steven L.},
    booktitle = {Neurobiology of Language},
    publisher = {Academic Press},
      address = {San Diego},
         year = {2016},
        pages = {325-337},
     abstract = {A dual loop model for the processing of language in the brain was proposed early in history and has also formed the basis for many neuropsychological models. These models incorporate a (direct) route for sensorimotor mapping and a (indirect) route for "semantic" processing. Dual loop models also emerged in the field of visual processing, motor control, or spatial attention. Thus, a general dual loop system may provide the framework for the interpretation of cognition in human and primate brains independent of the modality and species. Modern imaging techniques like diffusion tensor imaging (DTI)-based fiber tracking identified long human association tracts for ventral and dorsal pathways. The extreme capsule (EmC) and uncinate fascicle (UF) are part of the ventral system, and the superior longitudinal fasciculi (SLF I, II, III) and the arcuate fasciculus (AF) are all dorsal pathways. This chapter reviews language processing in the brain in the context of a domain general dual loop system.}
}
@incollection{RizzolattiandRozziNL-16,
       author = {Giacomo Rizzolatti and Stefano Rozzi},
        title = {Motor Cortex and Mirror System in Monkeys and Humans},
       editor = {Hickok, Gregory  and Small, Steven L.},
    booktitle = {Neurobiology of Language},
    publisher = {Academic Press},
      address = {San Diego},
         year = {2016},
        pages = {59-72},
     abstract = {The view of the functions of the motor system has radically changed in past years. It is now clear that the motor system not only is a "producer" of movements but also is involved in cognitive functions. In this chapter we review the anatomical and functional organization of the cortical motor system in monkeys and humans with particular emphasis on those areas that are endowed of the mirror mechanism and involved in communication. We discuss the possible evolutionary link between basic unintentional communication and voluntary communication. We conclude examining the hypothesis that purports that motor activation is involved in phonemes and understanding of action words.}
}
@article{HamzeietalCC-16,
       author = {Hamzei, Farsin and Vry, Magnus-Sebastian and Saur, Dorothee and Glauche, Volkmar and Hoeren, Markus and Mader, Irina and Weiller, Cornelius and Rijntjes, Michel},
        title = {The Dual-Loop Model and the Human Mirror Neuron System: an Exploratory Combined fMRI and DTI Study of the Inferior Frontal Gyrus},
      journal = {Cerebral Cortex},
       volume = {26},
       number = {5},
        pages = {2215-2224},
         year = {2016},
     abstract = {The inferior frontal gyrus (IFG) is active during both goal-directed action and while observing the same motor act, leading to the idea that also the meaning of a motor act (action understanding) is represented in this "mirror neuron system" (MNS). However, in the dual-loop model, based on dorsal and ventral visual streams, the MNS is thought to be a function of the dorsal steam, projecting to pars opercularis (BA44) of IFG, while recent studies suggest that conceptual meaning and semantic analysis are a function of ventral connections, projecting mainly to pars triangularis (BA45) of IFG. To resolve this discrepancy, we investigated action observation (AO) and imitation (IMI) using fMRI in a large group of subjects. A grasping task (GR) assessed the contribution from movement without AO. We analyzed connections of the MNS-related areas within IFG with postrolandic areas with the use of activation-based DTI. We found that action observation with imitation are mainly a function of the dorsal stream centered on dorsal part of BA44, but also involve BA45, which is dorsally and ventrally connected to the same postrolandic regions. The current finding suggests that BA45 is the crucial part where the MNS and the dual-loop system interact.}
}
@article{RijntjesetalFiEN-12,
       author = {Rijntjes, Michel and Weiller, Cornelius and Bormann, Tobias and Musso, Mariachristina},
        title = {The dual loop model: its relation to language and other modalities},
      journal = {Frontiers in Evolutionary Neuroscience},
       volume = {4},
        pages = {9},
         year = {2012},
     abstract = {The current neurobiological consensus of a general dual-loop system scaffolding human and primate brains gives evidence that the dorsal and ventral connections subserve similar functions, independent of the modality and species. However, most current commentators agree that, although bees dance and chimpanzees grunt, these systems of communication differ qualitatively from human language. So why is language in humans unique? We discuss anatomical differences between humans and other animals, the meaning of lesion studies in patients, the role of inner speech, and compare functional imaging studies in language with other modalities in respect to the dual loop model. These aspects might be helpful for understanding what kind of biological system the language faculty is, and how it relates to other systems in our own species and others.}
}
%

⁷² Examples of qualia include the perceived sensation of pain of a headache, the taste of wine, as well as the redness of an evening sky. As qualitative characters of sensation, qualia stand in contrast to "propositional attitudes", where the focus is on beliefs about experience rather than what is it directly like to be experiencing. (SOURCE)

⁷³ Natural language programming (NLP) is an ontology-assisted way of programming in terms of natural language sentences, e.g. English. A structured document with Content, sections and subsections for explanations of sentences forms a NLP document, which is actually a computer program. Natural languages and natural language user interfaces include Inform7, a natural programming language for making interactive fiction, Ring, a general purpose language, Shakespeare, an esoteric natural programming language in the style of the plays of William Shakespeare, and Wolfram Alpha, a computational knowledge engine, using natural language input. The smallest unit of statement in NLP is a sentence. Each sentence is stated in terms of concepts from the underlying ontology, attributes in that ontology and named objects in capital letters. In an NLP text every sentence unambiguously compiles into a procedure call in the underlying high-level programming language such as MATLAB, Octave, SciLab, Python, etc. (SOURCE)

⁷⁴ Here are the BibTeX references and abstracts for several papers on using natural language to write programs:

@book{Veres2008,
        title = {{Natural Language Programming of Agents and Robotic Devices: Publishing for agents and humans in sEnglish}},
       author = {S\'{a}ndor M. Veres},
    publisher = {SysBrain Ltd},
      address = {London},
         year = {2008},
     abstract = {This paper is an application of ontology theory, conceptual graphs and programming language theories to develop the theoretical foundations of natural language programming (NLP) that has in recent years been used to produce natural language documents for intelligent agents and human readers. The analysis given reveals three benefits of NLP. First, it is “conceptualized” programming that enables developers to write less bug prone programs due to clarity of code presentation and enforced structuring of data. Secondly, NLP can aid programming of the all important abstractions for robots: event, action and world model abstractions can be created by sentences. Thirdly, NLP can be used to publish natural language documents by researchers, i.e. English language documents on control theory and procedures, on the Internet or in printed documents. This theoretical paper also defines a large class of intelligent agents that can read such documents. This enables human users and agents to have shared understanding of how application systems work.}
}
@inproceedings{LeietalACL-13,
       author = {Tao Lei, Fan Long, Regina Barzilay and Martin Rinard},
        title = {From Natural Language Specifications to Program Input Parsers},
    booktitle = {The 51st Annual Meeting of the Association for Computational Linguistics},
     location = {Sofia, Bulgaria},
         year = {2013},
     abstract = {We present a method for automatically generating input parsers from English specifications of input file formats. We use a Bayesian generative model to capture relevant natural language phenomena and translate the English specification into a specification tree, which is then translated into a C++ input parser. We model the problem as a joint dependency parsing and semantic role labeling task. Our method is based on two sources of information: (1) the correlation between the text and the specification tree and (2) noisy supervision as determined by the success of the generated C++ parser in reading input examples. Our results show that our approach achieves 80.0\% F-Score accuracy compared to an F-Score of 66.7\% produced by a state-of-the-art semantic parser on a dataset of input format speci- fications from the ACM International Collegiate Programming Contest (which were written in English for humans with no intention of providing support for automated processing).}
}
@article{LocascioetalCoRR-16,
       author = {Nicholas Locascio and Karthik Narasimhan and Eduardo DeLeon and Nate Kushman and Regina Barzilay},
        title = {Neural Generation of Regular Expressions from Natural Language with Minimal Domain Knowledge},
      journal = {CoRR},
       volume = {arxiv:1608.03000},
         year = {2016},
     abstract = {This paper explores the task of translating natural language queries into regular expressions which embody their meaning. In contrast to prior work, the proposed neural model does not utilize domain-specific crafting, learning to translate directly from a parallel corpus. To fully explore the potential of neural models, we propose a methodology for collecting a large corpus of regular expression, natural language pairs. Our resulting model achieves a performance gain of 19.6\% over previous state-of-the-art models.}
}
@inproceedings{KushmanandBarzilayACL-13,
       author = {Kushman, Nate and Barzilay, Regina},
        title = {Using Semantic Unification to Generate Regular Expressions from Natural Language},
    booktitle = {Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies},
         year = {2013},
    publisher = {Association for Computational Linguistics},
        pages = {826-836},
     location = {Atlanta, Georgia},
     abstract = {We consider the problem of translating natural language text queries into regular expressions which represent their meaning. The mismatch in the level of abstraction between the natural language representation and the regular expression representation make this a novel and challenging problem. However, a given regular expression can be written in many semantically equivalent forms, and we exploit this flexibility to facilitate translation by finding a form which more directly corresponds to the natural language. We evaluate our technique on a set of natural language queries and their associated regular expressions which we gathered from Amazon Mechanical Turk. Our model substantially outperforms a stateof-the-art semantic parsing baseline, yielding a 29\% absolute improvement in accuracy}
}
@inproceedings{ErnstSNAPL-17,
       author = {Michael D. Ernst},
        title = {Natural language is a programming language: Applying natural language processing to software development},
    booktitle = {{SNAPL 2017: the 2nd Summit oN Advances in Programming Languages}},
        pages = {4:1-4:14},
      address = {Asilomar, CA, USA},
         year = {2017},
     abstract = {A powerful, but limited, way to view software is as source code alone. Treating a program as a sequence of instructions enables it to be formalized and makes it amenable to mathematical techniques such as abstract interpretation and model checking. A program consists of much more than a sequence of instructions. Developers make use of test cases, documentation, variable names, program structure, the version control repository, and more. I argue that it is time to take the blinders off of software analysis tools: tools should use all these artifacts to deduce more powerful and useful information about the program. Researchers are beginning to make progress towards this vision. This paper gives, as examples, four results that find bugs and generate code by applying natural language processing techniques to software artifacts. The four techniques use as input error messages, variable names, procedure documentation, and user questions. They use four different NLP techniques: document similarity, word semantics, parse trees, and neural networks. The initial results suggest that this is a promising avenue for future work.},
}

⁷⁵ Our current best guess for the beginning of life on Earth is 3.8 billion years ago. The oldest fossils of single-celled organisms date from about about 3.5 billion years ago. Viruses are present by 3 billion years ago, but may be as old as life itself. Oxygen from photosynthetic cyanobacteria starts to build up in the atmosphere around 2.4 billion years ago. Eukaryotic cells with internal organelles come into being 2.0 billion years ago. The eukaryotes divide into three groups: the ancestors of modern plants, fungi and animals 1.5 billion years ago. The first multicellular life develops around 900 million years ago, and the fossil evidence shows that animals were exploring the land by 500 million years ago. Humans diverge from their closest relatives, the chimpanzees and bonobos around 6 million years ago. (SOURCE)

⁷⁶ Andrew Ng, one of the cofounders of the Google Brain project, will be the new chairman of Woebot, a company that operates a chatbot of the same name.

The bot is designed to help people work on their own mental health issues using techniques that originate with cognitive behavioral therapy. It’s a sort of therapy that focuses on helping people manage their moods by developing personal coping strategies for mental illnesses like depression and anxiety.

A study from Stanford University showed that people who used Woebot reported a decrease in anxiety and depression symptoms after two weeks. Ng thinks that machine learning could provide a massive benefit in the mental health space, which is why he’s going to be working with Woebot.

For example, Woebot is available to talk when people are having problems at odd hours, during times when a therapist might not be awake to take a phone call. And while it’s not a replacement for a human therapist (the bot is quite clear about that), it can help augment human therapies by giving people an easier way to externalize their thoughts if nobody else is available.

Ng said in an interview that he’s still going to be working on other projects. His role at Woebot will be to serve on the company’s board and provide support for its work, not take on a full-time job at the startup. For example, he’s still working on finishing up his series of deep learning courses for Coursera, the online learning site that he cofounded.

Health care is one of the focuses of his current work, after leaving his post as Baidu’s head of machine learning. Earlier this year, a team at Stanford that he worked with released a paper showing that they had trained a machine learning system to read electrocardiograms better than a cardiologist could. (SOURCE)

⁷⁷ Defined concisely here as "awareness of an external object or something within oneself", and, with some additional detail and hyperbole by Schneider and Velmans [256] as "[a]nything that we are aware of at a given moment forms part of our consciousness, making conscious experience at once the most familiar and most mysterious aspect of our lives".

⁷⁸ This proposal is based on a regularization term which encourages the top-level representation (meant to be at the most abstract level) to be such that when a sparse attention mechanism focuses on a few elements of the state representation (factors, variables or concepts, i.e. a few axes or dimensions in the representation space), that small set of variables of which the agent is aware at a given moment can be combined to make a useful statement about reality or usefully condition an action or policy [25].

⁷⁹ Apropos our discussion of sharing data and knowledge in field of neuroscience, between 2012 and 2016 Google research scientist published over 700 papers, and 2017 promises to be another record year. ://www.technologyreview.com/s/603984/googles-ai-explosion-in-one-chart/(SOURCE)

⁸⁰ By way of comparison, The IBM TrueNorth chip contains 4,096 cores and 5.4 million transistors, while only drawing 70 milliwatts of power. The chip simulates a million neurons and 256 million synapses. TrueNorth is (very) roughly able to simulate the equivalent of a bee's brain. In contrast to the IBM and Intel focus on hardware, Qualcomm's Zeroth is a platform for brain-inspired computing from Qualcomm. It is based around a — thus far vaporware — neural processing unit (NPU) AI accelerator chip and a software API to interact with the platform. It is advertised as making deep learning available to mobile devices. The software operates locally rather than as a cloud application.

⁸¹ This 2012 article in Science Daily suggests that scientists now know how humans solve the cocktail party problem: "In the experiments, patients listened to two speech samples played to them simultaneously in which different phrases were spoken by different speakers. They were asked to identify the words they heard spoken by one of the two speakers.

The authors then applied new decoding methods to "reconstruct" what the subjects heard from analyzing their brain activity patterns. Strikingly, the authors found that neural responses in the auditory cortex only reflected those of the targeted speaker. They found that their decoding algorithm could predict which speaker and even what specific words the subject was listening to based on those neural patterns. In other words, they could tell when the listener's attention strayed to another speaker.

The authors claim that the algorithm worked so well that they could predict not only the correct responses, but also determine when the patients paid attention to the wrong word. [...] The new findings suggest that the representation of speech in the cortex does not just reflect the entire external acoustic environment but instead just what we really want or need to hear." — Nima Mesgarani, Edward F. Chang. Selective cortical representation of attended speaker in multi-talker speech perception. Nature, 2012.

⁸² Application of their work 3D covolutional networks trained for action recognition [125]:

@article{GyglietalCoRR-17,
        title = {Deep Value Networks Learn to Evaluate and Iteratively Refine Structured Outputs},
       author = {Michael Gygli and Mohammad Norouzi and Anelia Angelova},
      journal = {CoRR},
       volume = {arXiv:1703.04363},
         year = {2017},
     abstract = {We approach structured output prediction by optimizing a deep value network (DVN) to precisely estimate the task loss on different output configurations for a given input. Once the model is trained, we perform inference by gradient descent on the continuous relaxations of the output variables to find outputs with promising scores from the value network. When applied to image segmentation, the value network takes an image and a segmentation mask as inputs and predicts a scalar estimating the intersection over union between the input and ground truth masks. For multi-label classification, the DVN’s objective is to correctly predict the F1 score for any potential label configuration. The DVN framework achieves the state-of-the-art results on multi-label prediction and image segmentation benchmarks.}
}  

⁸³ Think of consciousness as the formation of low-dimensional combinations of simple concepts originating in an abstract rather than sensory space:

@article{BengioCoRR-17,
       author = {Yoshua Bengio},
        title = {The Consciousness Prior},
      journal = {CoRR},
       volume = {arxiv:1709.08568},
         year = {2017},
     abstract = {A new prior is proposed for representation learning, which can be combined with other priors in order to help disentangling abstract factors from each other. It is inspired by the phenomenon of consciousness seen as the formation of a low-dimensional combination of a few concepts constituting a conscious thought, i.e., consciousness as awareness at a particular time instant. This provides a powerful constraint on the representation in that such low-dimensional thought vectors can correspond to statements about reality which are either true, highly probable, or very useful for taking decisions. The fact that a few elements of the current state can be combined into such a predictive or useful statement is a strong constraint and deviates considerably from the maximum likelihood approaches to modeling data and how states unfold in the future based on an agent’s actions. Instead of making predictions in the sensory (e.g. pixel) space, the consciousness prior allow the agent to make predictions in the abstract space, with only a few dimensions of that space being involved in each of these predictions. The consciousness prior also makes it natural to map conscious states to natural language utterances or to express classical AI knowledge in the form of facts and rules, although the conscious states may be richer than what can be expressed easily in the form of a sentence, a fact or a rule.}
}

⁸⁴ By introducing the term cognitive apparatus I'm probably being too cautious here. The word "mind" seems pretentious and somewhat presumptive. "Brain" is close but I'm not sure that I want to divide up the autonomic (sympathetic and parasympathetic) and somatic (sensory and motor) nervous system along traditional lines.

⁸⁵ It takes an effort for me to listen to David Chalmers (EDGE) on augmented reality, consciousness or virtual worlds; Nick Bostrom (BIO) on what the chances are we are already living in a virtual world; Jaron Lanier on the myth of AI (EDGE). I'm only now learning to really listen carefully.

Historically, I've disagreed with all three of these philosophy / culture mavens, but we are always always stumbling on interesting lectures and interviews, and, while there are only so many minutes in a day, you could do worse than listen to a smart, thoughtful person talking about a subject he or she has spent a lifetime studying ... even if you disagree with a lot of what they say or the way in which they say it. Here's a suggestion: never actually look at the person when they're speaking and preferably avoid looking at them altogether, before, during or after their speaking. If you think appearances don't matter; think again.

Hidden biases arise from any number of factors besides physical appearance. It's not just whether you are male or female, young or old, your skin tone, race, religion, social standing, voice, etc. Humans instinctively and incorrectly believe that these characteristics tell you a lot about the person, their beliefs, the credence they attach to their claims, the strength of their conviction, the depth of their knowledge regarding whatever subject they are expostulating about, ... expose any of the markers we are programmed to believe reveal these character traits and you have been framed and your audience is now anchored in Kahneman and Tversky's sense of those terms.

⁸⁶ Imagine if you had a Fender Stratocaster physically integrated into your body. You need not be limited to six fingers and two hands but if we were to limit ourselves to a traditional Stratocaster with six strings, perhaps the most useful extension would be if you could press any six string-fret combinations on the fret-board as long as there was at most one combination per string.

You might have similar flexibility in picking with the other hand resulting in a level of dexterity that would put Chet Atkins and Jimi Hendrix to shame. The integration of a wah-wah pedal and tremolo bar could be controlled by twisting your torso or turning your head and you could control a room full of special-effects simply by directing your gaze at the icons in an image of a fully equipped studio sound-stage projected on your retina.

⁸⁷ Here are the BibTeX citations and abstracts from the three papers on learning how to write programs mentioned in the text [16, 15, 189]:

@inproceedings{LongandMartinICSE-16,
       author = {Long, Fan and Rinard, Martin},
        title = {An Analysis of the Search Spaces for Generate and Validate Patch Generation Systems},
    booktitle = {Proceedings of the 38th International Conference on Software Engineering},
     location = {Austin, Texas},
    publisher = {ACM},
      address = {New York, NY, USA},
         year = {2016},
        pages = {702-713},
     abstract = {We present the first systematic analysis of key characteristics of patch search spaces for automatic patch generation systems. We analyze sixteen different configurations of the patch search spaces of SPR and Prophet, two current state-of-the-art patch generation systems. The analysis shows that (i) correct patches are sparse in the search spaces (typically at most one correct patch per search space per defect), (ii) incorrect patches that nevertheless pass all of the test cases in the validation test suite are typically orders of magnitude more abundant, and (iii) leveraging information other than the test suite is therefore critical for enabling the system to successfully isolate correct patches.}
}
@inproceedings{BalogetalICLR-17,
       author = {Matej Balog and Alexander L. Gaunt and Marc Brockschmidt and Sebastian Nowozin and Daniel Tarlow},
        title = {{DeepCoder}: Learning to Write Programs},
    booktitle = {International Conference on Learning Representations},
         year = 2017,
     abstract = {We develop a first line of attack for solving programming competition-style problems from input-output examples using deep learning. The approach is to train a neural network to predict properties of the program that generated the outputs from the inputs. We use the neural network's predictions to augment search techniques from the programming languages community, including enumerative search and an SMT-based solver. Empirically, we show that our approach leads to an order of magnitude speedup over the strong non-augmented baselines and a Recurrent Neural Network approach, and that we are able to solve problems of difficulty comparable to the simplest problems on programming competition websites.}
}

⁸⁸ Here are the BibTeX citations and abstracts from three papers on the biology, psychology and related technology for theory-of-mind (ToM) reasoning [236, 34, 187, 87]:

@inbook{PynadathetalCAGI-14,
       author = {Pynadath, David V. and Rosenbloom, Paul S. and Marsella, Stacy C.},
       editor = {Goertzel, Ben and Orseau, Laurent and Snaider, Javier},
        title = {Reinforcement Learning for Adaptive Theory of Mind in the Sigma Cognitive Architecture},
    booktitle = {Proceedings of Artificial General Intelligence: 7th International Conference},
    publisher = {Springer International Publishing},
         year = {2014},
        pages = {143-154},
     abstract = {One of the most common applications of human intelligence is social interaction, where people must make effective decisions despite uncertainty about the potential behavior of others around them. Reinforcement learning (RL) provides one method for agents to acquire knowledge about such interactions. We investigate different methods of multiagent reinforcement learning within the Sigma cognitive architecture. We leverage Sigma's architectural mechanism for gradient descent to realize four different approaches to multiagent learning: (1) with no explicit model of the other agent, (2) with a model of the other agent as following an unknown stationary policy, (3) with prior knowledge of the other agent's possible reward functions, and (4) through inverse reinforcement learning (IRL) of the other agent's reward function. While the first three variations re-create existing approaches from the literature, the fourth represents a novel combination of RL and IRL for social decision-making. We show how all four styles of adaptive Theory of Mind are realized through Sigma's same gradient descent algorithm, and we illustrate their behavior within an abstract negotiation task.},
}
@article{BowmanetalDS-12,
       author = {Bowman, L. C. and Liu, D. and Meltzoff, A. N. and Wellman, H. M.},
        title = {Neural correlates of belief- and desire-reasoning in 7- and 8-year-old children: an event-related potential study},
      journal = {Development Science},
       volume = {15},
       number = {5},
         year = {2012},
        pages = {618-632},
     abstract = {Theory of mind requires belief- and desire-understanding. Event-related brain potential (ERP) research on belief- and desire-reasoning in adults found mid-frontal activations for both desires and beliefs, and selective right-posterior activations only for beliefs. Developmentally, children understand desires before beliefs; thus, a critical question concerns whether neural specialization for belief-reasoning exists in childhood or develops later. Neural activity was recorded as 7- and 8-year-olds (N = 18) performed the same diverse-desires, diverse-beliefs, and physical control tasks used in a previous adult ERP study. Like adults, mid-frontal scalp activations were found for belief- and desire-reasoning. Moreover, analyses using correct trials alone yielded selective right-posterior activations for belief-reasoning. Results suggest developmental links between increasingly accurate understanding of complex mental states and neural specialization supporting this understanding.}
}
@article{LiuetalCD-09,
       author = {Liu, D.  and Meltzoff, A. N.  and Wellman, H. M.},
        title = {Neural correlates of belief- and desire-reasoning},
      journal = {Child Development},
       volume = {80},
       number = {4},
         year = {2009},
        pages = {1163-1171},
     abstract = {Theory of mind requires an understanding of both desires and beliefs. Moreover, children understand desires before beliefs. Little is known about the mechanisms underlying this developmental lag. Additionally, previous neuroimaging and neurophysiological studies have neglected the direct comparison of these developmentally critical mental-state concepts. Event-related brain potentials were recorded as participants (N = 24; mean age = 22 years) reasoned about diverse-desires, diverse-beliefs, and parallel physical situations. A mid-frontal late slow wave (LSW) was associated with desire and belief judgments. A right-posterior LSW was only associated with belief judgments. These findings demonstrate neural overlap and critical differences in reasoning explicitly about desires and beliefs, and they suggest children recruit additional neural processes for belief judgments beyond a common, more general, mentalizing neural system.}
}
@article{EmeretalSN-06,
       author = {Ermer, E. and Guerin, S. A. and Cosmides, L. and Tooby, J. and Miller, M. B.},
        title = {Theory of mind broad and narrow: reasoning about social exchange engages {T}o{M} areas, precautionary reasoning does not},
      journal = {Society of Neuroscience},
       volume = {1},
       number = {3-4},
         year = {2006},
        pages = {196-219},
     abstract = {Baron-Cohen (1995) proposed that the theory of mind (ToM) inference system evolved to promote strategic social interaction. Social exchange--a form of co-operation for mutual benefit--involves strategic social interaction and requires ToM inferences about the contents of other individuals' mental states, especially their desires, goals, and intentions. There are behavioral and neuropsychological dissociations between reasoning about social exchange and reasoning about equivalent problems tapping other, more general content domains. It has therefore been proposed that social exchange behavior is regulated by social contract algorithms: a domain-specific inference system that is functionally specialized for reasoning about social exchange. We report an fMRI study using the Wason selection task that provides further support for this hypothesis. Precautionary rules share so many properties with social exchange rules--they are conditional, deontic, and involve subjective utilities--that most reasoning theories claim they are processed by the same neurocomputational machinery. Nevertheless, neuroimaging shows that reasoning about social exchange activates brain areas not activated by reasoning about precautionary rules, and vice versa. As predicted, neural correlates of ToM (anterior and posterior temporal cortex) were activated when subjects interpreted social exchange rules, but not precautionary rules (where ToM inferences are unnecessary). We argue that the interaction between ToM and social contract algorithms can be reciprocal: social contract algorithms requires ToM inferences, but their functional logic also allows ToM inferences to be made. By considering interactions between ToM in the narrower sense (belief-desire reasoning) and all the social inference systems that create the logic of human social interaction--ones that enable as well as use inferences about the content of mental states--a broader conception of ToM may emerge: a computational model embodying a Theory of Human Nature (ToHN).}
} 

⁸⁹ I may be delusional but I believe we can now build a complete, reasonably accurate facsimile of the human brain. If I'm right that means we can build an intelligence that is at least as capable of human beings in terms of its ability to interact with and perform computations similar to human beings but is also extensible allowing us to engineer artificial beings based on a silicon substrate that would be substantially more durable and computationally powerful than unaugmented human beings. I believe that all the clues are available for anyone to see. Clearly it will take substantial computational and engineering resources to deliver on the extrapolation from current technology I just made, but it implies that super-human level intelligence will scale with Moore's law and in general with respect to the law of accelerating exponential return from advanced technology.

In Issue 42 of The Wild and Crazy Guide to the Singularity, I mentioned we could build an AI today comparable in function to a human being. By that I mean engineers could sit down with cognitive and systems neuroscientists and figure out how to put together existing pieces of technology to produce a reasonable simulation of a human brain, I also mentioned that because of the law of accelerating returns we could extend this basic brain design by trading computational capacity and functional capability for power since we still don’t know how to achieve comparable efficiency. What's perhaps scary is that we could add as prosthetics the many powerful tools we use every day as technology for accelerating scientific progress and exploit these as add-ons to create a more powerful and capable machine than those designed by evolution and natural selection.

This claim does not imply we don't have a great deal to learn about the detailed organization, function and pathology of biological brains, crucial in diagnosing neurological disorders. Moreover, at the systems neuroscience level, advances in computer architecture and algorithm design will profit from a better understanding of the details of how biological brains work. It is also worth noting that humans encompass a huge variety of biological diversity in both form and function — including the molecular machinery comprising our immune systems and the extensive biota that populate our bodies and considerably expand our inherent assets, most of which is more complex than what is needed in order to replicate the basic function and architecture of the human brain, but that, as long as we inhabit flesh bodies, will be essential to our health and well-being.

⁹⁰ When it comes to research and development in a commercial setting, you have to establish a reasonable tradeoff between the two. If the problem is completely terra incognita and you have no backup plan, then you're likely to fail. If the problem is tractable with little or no research required — figure out how to put well-understood components together, make the resulting system scale and secure and improve the user interface, then it's "just engineering" as some managers disparagingly quip.

The line between research and development can be thin at times, and, in this case, the question of whether to "hack it" or "go boldly" is often left hanging. Whether a manager should bet on a project or not, depends on the risk profile of the people assessing the R&D tradeoff; a good technical manager is able to work with overly optimistic research scientists and overly pessimistic software engineers to produce a reasonably accurate assessment of risks and rewards.

This aside reminded me of the pitch I made at the beginning of launching an early personal assistant project — here's a video of a presentation I gave a senior VP in early 2014. The VP was definitely of the "hack it" persuasion and almost convinced the technical lead — who was already chomping at the bit to start hiring infrastructure talent — to create a music DJ demo modeled after some locally famous NYC radio personality whom none of us had ever heard of.

In the end, we collaborated with Google Play Music, but when NLP problems started to surface, the program manager focused on infrastructure, duplicating effort elsewhere in the organization, and neglecting the problem of continuous dialog that was core to the product. The PM was unwilling to consider the new recurrent neural network "chatbot" technology being developed in the Google Brain group at the time. Needless to say, the project was discontinued soon after.

⁹¹ Recently we discussed the prospects for building better retinal prosthetics. The problem is plagued by several techncial challenges, including (i) we don't know exactly how the retina works, (ii) technology for interfacing silicon and neural circuitry is in its infancy, and (iii) we don't know how the prosthetic should communicate with the visual system. Since the R&D tradeoffs are complicated and technical challenges daunting, it makes sense to consider simple low-tech alternatives.

I remember our experience with early BrainGate technology. Michael Black was using relatively sophisticated sensing and pattern recognition techniques for the BCI and in the end we found that back-to-basics kalman filtering with no fancy spike sorting or preprocessing was better. The lesson was clear, you provide the brain with meaningful information and it will sort it out and make use of it.

David Eagleman made this even clearer for the layperson in his self-promotional TED talk showing off the VEST (Versatile Extra-Sensory Transducer) technology now being developed at the kickstarter launched Neosensory.

Simple KISS engineering exploiting what the brain does best. The VEST doesn't seem adequate to handling complex sensory experience until you think about how many bits you could actually convey by multiplexing a visual or auditory signal. Think about how body-schema mapped regions of the somatosensory cortex increase in size / density of neurons in response to learning or shrink as in the case of the rat barrel cortex when you pluck or clip the rat's whiskers.

Disclaimer: I did not contribute directly to the early BrainGate technology. I recruited Michael when he was at Xerox PARC and eventually hired him at Brown University where I was Chair of the Computer Science Department and served on the Scientific Advisory Board of the Brown Institute for Brain Science during its formation . John Donoghue who co-founded Cyberkinetics was the Chair of the Neuroscience Department and also chaired the SAB. One of my former students, Arjun Bansal, did his PhD with Donoghue, then joined BrainGate and later worked at Qualcomm before founding Nervana Systems which was acquired by Intel in 2016.

⁹² Whenever we become aware of an unexpected piece of information, the brain suddenly seems to burst into a large-scale activity pattern. My colleagues and I have called this property global ignition. We were inspired by the Canadian neurophysiologist Donald Hebb, who first analyzed the behavior of collective assemblies of neurons in his 1949 best seller The Organization of Behavior. Hebb explained, in very intuitive terms, how a network of neurons that excite one another can quickly fall into a global pattern of synchronized activity—much as an audience, after the first few handclaps, suddenly bursts into broad applause. Like the enthusiastic spectators who stand up after a concert and contagiously spread the applause, the large pyramidal neurons in the upper layers of cortex broadcast their excitation to a large audience of receiving neurons.

Global ignition, my colleagues and I have suggested, occurs when this broadcast excitation exceeds a threshold and becomes self-reinforcing: some neurons excite others that, in turn, return the excitation. The net result is an explosion of activity: the neurons that are strongly interconnected burst into a self-sustained state of high-level activity, a reverberating "cell assembly," as Hebb called it.

This collective phenomenon resembles what physicists call a "phase transition," or mathematicians a "bifurcation": a sudden, nearly discontinuous change in the state of a physical system. Water that freezes into an ice cube epitomizes the phase transition from liquid to solid. Early on in our thinking about consciousness, my colleagues and I noted that the concept of phase transition captures many properties of conscious perception. Like freezing, consciousness exhibits a threshold: a brief stimulus remains subliminal, while an incrementally longer one becomes fully visible. Most physical self-amplifying systems possess a tipping point where global change happens or fails depending on minute impurities or noise. The brain, we reasoned, may be no exception.

⁹³ So how could we separate the genuine conscious code from its accompanying unconscious bells and whistles? In principle, the answer is easy. We need to search the brain for a decodable neural representation whose content correlates 100 percent with our subjective awareness. The conscious code that we are looking for should contain a full record of the subject’s experience, replete with exactly the same level of detail as the person perceives. It should be insensitive to features that she misses, even if they are physically present in the input. Conversely, it should encode the subjective content of conscious perception, even if that perception is an illusion or a hallucination. It should also preserve our subjective sense of perceived similarity: when we see a diamond and a square as two distinct shapes, rather than rotated versions of each other, so should the brain’s conscious representation.

The conscious code should also be highly invariant: it should stay put whenever we feel that the world is stable, but change as soon as we see it moving. This criterion strongly constrains the search for signatures of consciousness, because it almost certainly excludes all our early sensory areas. As we walk down a corridor, the walls project a constantly changing image on our retinas—but we are oblivious to this visual motion and perceive a stable room. Motion is omnipresent in our early visual areas but not in our awareness. Three or four times per second, our eyes jiggle around. As a result, on the retina as well as in most of our visual areas, the entire image of the world slips back and forth.

Fortunately, we remain oblivious to this nauseating swirling: our perception remains steady. Even when we gaze at a moving target, we do not perceive the background scenery gliding in the opposite direction. In the cortex, our conscious code must therefore be similarly stabilized. Somehow, thanks to the motion sensors in our inner ear and to predictions arising from our motor commands, we manage to subtract out our own motion and perceive our environment as an invariant entity. Only when these predictive motor signals are bypassed—for instance, when you move your eye by gently poking it with a finger—does the whole world seem moving.

⁹⁴ Forkhead box protein P2 (FOXP2) is a protein that, in humans, is encoded by the FOXP2 gene is required for proper development of speech and language.

⁹⁵ As an inspiration and relaxing interlude, you may find it worthwhile listening to the documentary entitled "Joshua Bell," produced in 1995 for BBC Television’s "Omnibus" series, on the life of former child prodigy Joshua Bell. At age 26, he was regarded as one of the greatest violinists in the world. The film looks back at Bell's childhood in Indiana, focusing on his early life and introduction to the violin. Despite his early interest in the violin, he led an otherwise normal childhood. His first teacher was rather taciturn and did little to inspire.

However, his second teacher Josef Gingold was warm, supportive and emanated such a deep love of music that Bell blossomed under his tutelage. Gingold introduced Bell to the work of Fritz Kreisler figuring so prominently in Gingold's career and provided Bell with additional inspiration and a deeper appreciation for the art and craft of a classical musician. It is interesting to speculate how Gingold's encouragement shaped the neural basis of Bell's emotional experience of listening to and playing classical music and the physical and intellectual challenges of playing the violin.

As we'll see later, Graziano's theory relies on the idea of attention schema [107]. He writes that "[a] schema is a complex information structure that is used to represent something. One of the first schemas to be extensively studied in psychology and neuroscience was the body schema, first proposed by Head and Holmes in 1911. The body schema is an internal model — an organized set of information that represents the shape, structure, and movement of the body, that distinguishes between objects belonging to the body and objects that are foreign." You might find it interesting to learn about how the body schema changes when someone learns to ride a bike or play a musical instrument [28].

⁹⁶ Attention, physiological attention as it is understood by neuroscientists, is a procedure. It is a competition between signals in the brain. It occurs because of the specific way that neurons interact with each other. One set of signals, carried by one set of neurons, rises in strength and suppresses other, competing signals carried by other neurons. For example, the visual system builds informational models of the objects in a scene. If you are looking at a cluttered scene, which informational model in your brain will win the competition of the moment, rise in signal strength, suppress other models, and dominate the brain’s computations? This competition among signals—the process by which one signal wins and dominates for a moment, then sinks down as another signal dominates—is attention. Attention may be complicated, but it is not mysterious. It is physically understandable. It could be built into a circuit [107].

⁹⁷ The correspondence between awareness and attention is close. In the previous chapter I outlined a list of similarities, and in later chapters I will discuss the relationship in greater detail. The two are so similar that it is tempting to think they might be the same thing with no distinction at all. But there is a fundamental difference. Attention, the competition among signals and the enhancement of signals in the brain, is a mechanistic process. It is not explicit knowledge. Awareness, in contrast, is accessible as explicit knowledge. The brain does attention but knows awareness. The relationship between attention and awareness is therefore exactly the relationship between [...] between a real thing and a representation of it in the brain that is cognitively accessible. Awareness, in this view, is a description, a useful if physically inaccurate sketch of what it means for the brain to focus its attention [107].

⁹⁸ Much has been learned recently about the neuronal basis of decision-making, especially in the relatively simple case of visual motion. Suppose that you are looking at a blurry or flickering image and are asked to decide its direction of motion. It can drift either to the left or the right, but because of the noisy quality of the image, you have trouble determining the direction. By making the task difficult in this way, neuroscientists can slow down the decision process, thereby making it easier to study.

This decision process appears to work as follows. First, the machinery in the visual system constructs signals that represent the motion of the image. Because the visual image is noisy, it may result in conflicting signals indicating motion in a variety of directions. Second, those signals are received elsewhere in the brain by decision integrators. The decision integrators determine which motion signal is consistent enough or strong enough to cross a threshold. Once the threshold is crossed, a response is triggered. In this way, the system decides which direction the image is likely to be moving.

Strictly speaking, the system is not deciding whether the visual image is moving to the right or left. It is deciding between two information streams in the brain: is the left or right motion signal stronger? The decision can even be manipulated by inserting a fine electrode into the brain, into a particular part of the visual system, passing a very small electrical current and thereby boosting one or another of the motion signals.

Since neuroscientists have some notion of the brain’s machinery for decision-making, what can be inferred about awareness? As noted above, you can decide whether you have, inside of you, an awareness of something. Therefore awareness—or at least whatever it is you are deciding you have when you say you have awareness—can be fed into a decision integrator. We can make the task even more comparable to the visual-motion task. By making the images extremely brief or dim, we can make the task difficult. You may struggle for a moment, trying to decide whether any awareness of a stimulus is present. The decision machinery is engaged. This insight that conscious report depends on the machinery of decision-making has been pointed out before.

A crucial property of decision-making is that not only is the decision itself a manipulation of data, but the decision machine depends on data as input. It does not take any other input. Feeding in some res cogitans will not work on this machine. [...] You can’t feed it ectoplasm. You can’t feed it an intangible, ineffable, indescribable thing. You can’t feed it an emergent property that transcends information. You can only feed it information.

In introspecting, in asking yourself whether you have an awareness of something, and in making the decision that you have it, what you are deciding on, what you are assessing, the actual stuff your decision engine is collecting, weighing, and concluding that you have, is information. Strictly speaking, the neuronal machinery is deciding that certain information is present in your brain at a signal strength above a threshold [107].

⁹⁹ Simple pleasant and unpleasant feelings come from an ongoing process inside you called interoception. Interoception is your brain’s representation of all sensations from your internal organs and tissues, the hormones in your blood, and your immune system. Your insides are in motion. Your heart sends blood rushing through your veins and arteries. Your lungs fill and empty. Your stomach digests food. This interoceptive activity produces the spectrum of basic feeling [19].

¹⁰⁰ [W]hen doing insight practices, the only useful gold standard for reality is your own sensate experience. From the conventional point of view, things are usually thought to be there even when you can no longer experience them, and are thus assumed with only circumstantial evidence to be somewhat stable entities. Predictability is used to assume continuity of existence. For our day-to-day lives, this assumption is adequate and often very useful.

However, when doing insight practices, it just happens to be much more useful to assume that things are only there when you experience them and not there when you don’t. Thus, the gold standard for reality when doing insight practices is the sensations that make up your reality in that instant. Sensations not there at that time do not exist, and thus only the sensations arising in that instant do exist. In short, the vast majority of what you usually think of as making up your universe doesn’t exist the vast majority of the time, from a pure sensate point of view [198].

¹⁰¹ Your nervous system sends sensory input to your brain when your muscles or joints are injured, or your body tissues are damaged by excessive heat or cold, or in response to a chemical irritation like a pinch of pepper in your eye. This process is called nociception. Pain is an experience that occurs not only from physical damage but also when your brain predicts damage is imminent.

You might feel the pain even before the nurse pricks your arm with a needle. Your predictions are then corrected by actual nociceptive input from the body—the injection occurs—and once any prediction errors are dealt with, you have categorized the nociception sensations and made them meaningful. The pain you experience as coming from the needle is really in your brain. [19].

¹⁰² Your brain’s 86 billion neurons, which are connected into massive networks, never lie dormant awaiting a jump-start. Your neurons are always stimulating each other, sometimes millions at a time. Given enough oxygen and nutrients, these huge cascades of stimulation, known as intrinsic brain activity, continue from birth until death. This activity is nothing like a reaction triggered by the outside world. It’s more like breathing, a process that requires no external catalyst [21].

¹⁰³ Excerpts from Daniel Ingram's Mastering the Core Teachings of the Buddha, An Unusually Hardcore Dharma Book [198] relating to the experience of arising and passing (A&P) during insight meditation — also called the "4th insight stage" or the "2nd Vipassana jhana" in texts on advanced meditative practice:

Perceptual Thresholds — My favorite of the criteria, particularly found in technical and skilled meditators but also found in many others: people during the arising and passing may have the ability to perceive sensations with a speed, precision, and consistency that may be radically beyond what they were capable of before, such that they may perceive sensations up to maybe 40 times/second arising and vanishing during certain peak perceptual moments, particularly during the middle phase of the breath and in the center of wherever they place their attention. The phase characteristics of the A&P borrows from the 2nd jhana in general and involves the ability to perceive the arising and passing clearly of phenomena in a way that can feel quite effortless, and any vibrations noticed tend to be harmonically simple and change in frequency sinusoidally.

Insights into Selflessness — Some may perceive that all phenomena are arising and passing away, and wherever they turn their attention may notice the transience of sensations. Extrapolating from this clear perception, they may realize: "Ah, this means that there is no permanent self." Further, unitive experiences may have the same effect, and further, in some strange intuitive way the same basic notion of something having changed in the basic notion of self-hood may shift to something less solid. Also, the fact that the A&P experiences tend to happen in a way that is seemingly effortless or even unbidden, this experience of natural occurrences can also reinforce the notion that there is less control of things than one initially suspected, adding to the sense of there being somehow less of a self in things.

Cognitive Abilities — People who are in and have crossed the A&P tend to have an easier time with processes variously called things like "vision logic", "metacognitive processing" and the like. For those who are prone to such things, they will tend to have philosophical talents beyond those who have not crossed the A&P, realizing that things like age, underlying intelligence however defined, exposure to philosophical and related branches of thinking, and education level can significantly effect how this presents. They will also tend to have an increased ability to understand and navigate in terminology that may reluctantly be termed "spiritual", though this may show in other ways, such as an increased appreciation of things like the profundity and beauty of differential equations, the implications of modern physics for questions of Subject-Object non-duality, debates of free will vs super-determinism, and the like.

¹⁰⁴ Here is a speculative evolutionary timeline of consciousness from Consciousness and the Social Brain by Michael S. A. Graziano [107]:

The theory at a glance: from selective signal enhancement to consciousness. About half a billion years ago, nervous systems evolved an ability to enhance the most pressing of incoming signals. Gradually, this attentional focus came under top-down control. To effectively predict and deploy its own attentional focus, the brain needed a constantly updated simulation of attention. This model of attention was schematic and lacking in detail. Instead of attributing a complex neuronal machinery to the self, the model attributed to the self an experience of X — the property of being conscious of something.

Just as the brain could direct attention to external signals or to internal signals, that model of attention could attribute to the self a consciousness of external events or of internal events. As that model increased in sophistication, it came to be used not only to guide one’s own attention, but for a variety of other purposes including understanding other beings. Now, in humans, consciousness is a key part of what makes us socially capable.

In this theory, consciousness emerged first with a specific function related to the control of attention and continues to evolve and expand its cognitive role. The theory explains why a brain attributes the property of consciousness to itself, and why we humans are so prone to attribute consciousness to the people and objects around us.

Timeline: Hydras evolve approximately 550 million years ago (MYA) with no selective signal enhancement; animals that do show selective signal enhancement diverge from each other approximately 530 MYA; animals that show sophisticated top-down control of attention diverge from each other approximately 350 MYA; primates first appear approximately 65 MYA; hominids appear approximately 6 MYA; Homo sapiens appear approximately 0.2 MYA

¹⁰⁵ Notes and Excerpts on Language and Sensory Association Areas:

"I argue that language comprehension in sighted people might best be thought of as a kind of code-directed scene comprehension that draws heavily upon specifically visual, and probably largely prelinguistic processing constraints. The key processes of word-recognition and the assembly of visual word meaning patterns into interacting chains, however, may be mediated in part by species-specific activity patterns in secondary auditory cortex similar to those generated by uninterpreted speech-sound sequences." — (SOURCE: [117] in  [269])

"This tendency to ignore the structure of the brain is quite unfortunate in light of the recent progress made in primate neurobiology. Most current texts of physiological psychology, neuropsychology, and cognitive neuroscience (e.g., Caplan, 1987; Ellis and Young, 1988) still implicitly employ a model of the organization and evolution of the cortex that dates to the associationists of the late nineteenth century. In this way of thinking, ’primitive’ mammals like rats start out with primary visual, auditory, and somatosensory areas almost touching.

Next up the rung of an essentially pre-evolutionary scala natura come animals like cats, which have a small amount of ’uncommitted’ space in between. Finally, at the top, are primates and especially humans, where we find a great deal of uncommitted ’association’ cortex, properly situated to integrate and associate the modality-specific information presented to it by visual, auditory and somatosensory cortices (see e.g., Fodor, 1983; Ellis and Young, 1988, on the ’semantic system’ postulated in most models of word processing; Damasio, 1989)." — (SOURCE: [258])

¹⁰⁶ Von Economo neurons (VENs) (spindle cells) are not ubiquitous across animal brains. Only humans, great apes, certain whales, and elephants have been shown to possess VENs thus far. The relative number of cells varies by species, with humans having by far the greatest amount — over twice the number of our nearest evolutionary neighbor. One argument is that some aspect of VEN function is giving rise to the cognitive abilities that make us uniquely human. SOURCE

¹⁰⁷ Notes and Excerpts on Mirror Neurons and Physical Intuition:

Related mimetic concepts, acronyms, terminology and pseudo technical babble: Mirror Neural Network; Reflective Artificial Intelligence Dynamical System; Conscious Embodied and Virtually Embedded Models; association areas & sensory integration; revisiting couch potato; body-centric analogical machines, embodied cognition [28], brain systems supporting allostasis and interoception in humans [162].

"Mirror neurons represent a distinctive class of neurons that discharge both when an individual executes a motor act and when he observes another individual performing the same or similar motor act", "In humans, brain activity consistent with that of mirror neurons has been found in the premotor cortex, the supplementary motor area, the primary so much a sensory cortex, and the inferior parietal cortex", "a recent experiment [...] Shown that, in both humans and monkeys, the mere system also responds to the sound of actions. Functional magnetic resonance imaging (fMRI) can examine the entire brain at once and suggest that a much wider network of brain areas shows mere properties in humans than previously thought. These additional areas include the somatosensory and are thought to make the observer feel what it feels like to move in the observed", "the researchers found a small number of neurons that fired or showed their greatest activity both when the individual performed a task and when he observed a task. Other neurons had anti-mirror properties, that is, they responded when the participant saw an action but were inhibited when the participant perform that action. The mirror neurons found were located in the supplementary motor area and medial temporal cortex."

Mirror neurons are associated with one of the most intriguing aspect of our complex thought process, that is "Intention understanding". There are two distinct processes of information that one can get observing an action done by another individual.

"Thus, the activation of IPL action­constrained mirror neurons give information not only about, but also on why grasping is done (grasping­ for­ eating or grasping­ for placing). This specificity allowed the observer not only to recognize the observed motor act, but also to code what will be the next motor act of the not—yet—observed action, in other words to understand the intentions of the action's agent."

"The observation of an individual grasping an apple is immediately understood because it evokes the same motor representation in the parieto­ frontal mirror system of the observer. On the basis of this fundamental property of mirror neurons and the fact that the observation of actions like hand grasping activates the caudal part of IFG (Broca's area), neuroscientists proposed that the mirror mechanism is the basic mechanism from which language evolved."

"It is obvious that the mirror mechanism does not explain by itself the enormous complexity of speech. but, it solves one of the fundamental difficulties for understanding language evolution, that is, how and what is valid for the sender of a message become valid also for the receiver. Hypotheses and speculations on the various steps that have led from the monkey mirror system to language have been recently advanced."

"The ability to make consistent connections across different senses may have initially evolved in lower primates, but it went on developing in a more sophisticated manner in humans through remapping of mirror neurons which then became co­opted for other kinds of abstraction that humans excel in, like reasoning metaphors."

"Michael Corballis, an eminent cognitive neuroscientist, argues that what distinguishes us in the animal kingdom is our capacity for recursion, which is the ability to embed our thoughts within other thoughts. "I think, therefore I am" is an example of recursive thought, because the thinker has inserted himself into his thought. Recursion enables us to conceive of our own minds and the minds of others. It also gives us the power of mental "time travel" that is the ability to insert past experiences, or imagined future ones, into present consciousness. Corballis demonstrates how these recursive structures led to the emergence of language and speech, which ultimately enabled us to share our thoughts, plan with others, and reshape our environment to better reflect our creative imaginations. Mirror neurons shape the power of recursive embedding."

"This theory suggests that humans can construct a model in their brains of the thoughts and intentions of others. We can predict the thoughts, actions of others. The theory holds that humans anticipate and make sense of the behavior of others by activating mental processes that, if carried into action, would produce similar behavior. This includes intentional behavior as well as the expression of emotions. The theory states that children use their own emotions to predict what others will do. Therefore, we project our own mental states onto others. Mirror neurons are activated both when actions are executed, and the actions are observed. This unique function of mirror neurons may explain how people recognize and understand the states of others; mirroring observed action in the brain as if they conducted the observed action."

¹⁰⁸ Oliver Selfridge and I. J. Good — often disparaging called Turing's statistician but actually one of the most prolific scientists of his generation — were the only two Bletchley Park veterans I knew personally. Each of them was unique in their particular brand of genius. Each was iconoclastic and scientifically engaged throughout their long professional careers. Good wrote a huge amount about AI [100] and influenced my, Stuart Russell's and Eric Horvitz's research on bounded rationality [101].

¹⁰⁹ One important contribution of Stanislas Dehaene's work is that, building on the philosophical and cognitive-psychology work of Daniel Dennett, Paul and Patricia Churchland and other more enlightened thinkers, the clarity of Dehaene's theories and experiments will put to rest the arguments of Thomas Nagel ("What is it like to be a bat?), Ned Block ("Inverted Earth") and, to a lesser degree, David Chalmers' equivocating ("Facing Up to the Problem of Consciousness") or at the very least put them on a shelf labeled "Qualia and other made-up phenomena that are practically useless, intellectually barren and psychologically disturbing". See this entry in the Stanford Encyclopedia of Philosophy for a discussion of qualia, or, better yet, read Chapter 12 entitled "Qualia Disqualified" in Dan Dennett's Consciousness Explained [76].

¹¹⁰ If you want a quick review of the theories surrounding consciousness, I suggest you watch this 30 minute documentary featuring: Eben Alexander, Susan Blackmore, David Chalmers, Deepak Chopra, Patricia Churchland, Stanislas Dehaene, Daniel Dennett, Stuart Hameroff, Dean Radin, John Searle and Rupert Sheldrake, and, if you can't dismiss all but Dehaene for their lack of precision and all the rest except Dennett, Churchland and Blackmore for their lack of clarity and a profound state of befuddlement, then you have a lot of remedial reading and thinking to do.

¹¹¹ See here for an exercise that might help you to better understand consciousness both viscerally and computationally. Note that, however, while a purely computational account may turn out to be relatively straightforward for a computer scientist to understand intellectually, the implications of such an account could prove uncomfortable if they undermine the stories we tell about ourselves, reflect negatively on the special status we accord our species, or seem at odds with our everyday experience.

¹¹² Here is a small selection of papers published between 2000 and 2010 relating to language development and the evolution of ToM reasoning. The theories are like all scientific theories, subject revision based on subsequent contravening evidence, but the theories represented in this list are particular vulnerable due to the paucity of the evidence for and against at the time of their publications. Unfortunately, at the time of this writing — September 2017 — none of these theories have definitively been shown to be false or convincingly demonstrated to being even contingently true:

@article{PyersandSenghasPSYCHOLOGICAL-SCIENCE-09,
       author = {Pyers, Jennie E. and Senghas, Ann},
        title = {Language Promotes False-Belief Understanding: Evidence From Learners of a New Sign Language},
      journal = {Psychological Science},
       volume = {20},
        issue = {7},
         year = {2009},
        pages = {805-812},
     abstract = {Developmental studies have identified a strong correlation in the timing of language development and false-belief understanding. However, the nature of this relationship remains unresolved. Does language promote false-belief understanding, or does it merely facilitate development that could occur independently, albeit on a delayed timescale? We examined language development and false-belief understanding in deaf learners of an emerging sign language in Nicaragua. The use of mental-state vocabulary and performance on a low-verbal false-belief task were assessed, over 2 years, in adult and adolescent users of Nicaraguan Sign Language. Results show that those adults who acquired a nascent form of the language during childhood produce few mental-state signs and fail to exhibit false-belief understanding. Furthermore, those whose language developed over the period of the study correspondingly developed in false-belief understanding. Thus, language learning, over and above social experience, drives the development of a mature theory of mind.},
}
@article{BednyetalPNAS-09,
       author = {Bedny, Marina and Pascual-Leone, Alvaro and Saxe, Rebecca R.},
        title = {Growing up blind does not change the neural bases of Theory of Mind},
      journal = {Proceedings of the National Academy of Sciences},
       volume = {106},
       number = {27},
        pages = {11312-11317},
         year = {2009},
     abstract = {Humans reason about the mental states of others; this capacity is called Theory of Mind (ToM). In typically developing adults, ToM is supported by a consistent group of brain regions: the bilateral temporoparietal junction (TPJ), medial prefrontal cortex (MPFC), precuneus (PC), and anterior temporal sulci (aSTS). How experience and intrinsic biological factors interact to produce this adult functional profile is not known. In the current study we investigate the role of visual experience in the development of the ToM network by studying congenitally blind adults. In experiment 1, participants listened to stories and answered true/false questions about them. The stories were either about mental or physical representations of reality (e.g., photographs). In experiment 2, participants listened to stories about people's beliefs based on seeing or hearing; people's bodily sensations (e.g., hunger); and control stories without people. Participants judged whether each story had positive or negative valance. We find that ToM brain regions of sighted and congenitally blind adults are similarly localized and functionally specific. In congenitally blind adults, reasoning about mental states leads to activity in bilateral TPJ, MPFC, PC, and aSTS. These brain regions responded more to passages about beliefs than passages about nonbelief representations or passages about bodily sensations. Reasoning about mental states that are based on seeing is furthermore similar in congenitally blind and sighted individuals. Despite their different developmental experience, congenitally blind adults have a typical ToM network. We conclude that the development of neural mechanisms for ToM depends on innate factors and on experiences represented at an abstract level, amodally.},
}
@article{AnanthaswamyNEW-SCIENTIST-09,
        title = {Language may be key to theory of mind},
       author = {Anil Ananthaswamy},
      journal = {New Scientist},
         year = 2009,
      comment = {This is a general-science survey article that references the two, more-rigorous papers mentioned above.}, 
     abstract = {How blind and deaf people approach a cognitive test regarded as a milestone in human development has provided clues to how we deduce what others are thinking. Understanding another person's perspective, and realising that it can differ from our own, is known as theory of mind.},
}
@article{deVilliersLINGUA-07,
       author = {de Villiers, Jill},
        title = {The Interface of Language and Theory of Mind},
      journal = {Lingua: International Review of General Linguistics},
         year = {2007},
       volume = {117},
        issue = {11},
        pages = {1858-1878},
     abstract = {The proposal is made that the interface between language and theory of mind is bidirectional. It seems probable that the conceptual developments of early Theory of Mind form an essential basis for helping to fix at least word reference. In development from two to four years, no basis exists in research for conclusions about the direction of influence between language and Theory of Mind. At the stage of false belief reasoning, after age four, the role of the mastery of syntactic complementation is highlighted as a representational tool, that is, language development assists reasoning. The paper presents a brief summary of Theory of Mind, ranging from its earliest beginnings in infancy to the appreciation around age four years that others might hold false beliefs and act according to them. For each development, the parallel language developments are described, and questions are raised about the interface between the two. In particular, research that might determine the direction of influence from one to the other is discussed. More work is called for, especially with nonverbal tasks, good experimental linguistic work, and other special populations, that might allow a more precise delineation of how language and Theory of Mind interrelate at the interface.},
}
@incollection{MalleLANGUAGE-THEORY-OF-MIND-02,
       author = {Malle, B. F.},
        title = {The relation between language and theory of mind in development and evolution},
       editor = {T. Giv\'{o}n and B. F. Malle},
    booktitle = {The evolution of language out of pre-language},
    publisher = {Benjamins},
      address = {Amsterdam},
         year = {2002},
        pages = {265-284},
     abstract = {There is reason to believe that language and theory of mind have co- evolved, given their close relation in development and their tight connection in social behavior. However, they are clearly not inseparable—neurologically, cognitively, or functionally. So the question becomes, "What is the exact relation between language and theory of mind, in evolution, development, and social behavior?" To answer this question is a daunting task; I will try merely to clear a path toward an answer. I will consider several possible relations between the two faculties, bring conceptual arguments and empirical evidence to bear on them, and end up arguing for an escalation process in which language and theory of mind have fueled each other's evolution.}
}
@article{CharmanetalCOGNITIVE-DEVELOPMENT-00,
        title = {Testing joint attention, imitation, and play as infancy precursors to language and theory of mind},
      journal = {Cognitive Development},
       volume = {15},
       number = {4},
        pages = {481-498},
         year = {2000},
       author = {Tony Charman and Simon Baron-Cohen and John Swettenham and Gillian Baird and Antony Cox and Auriol Drew},
     abstract = {Various theoretical accounts propose that an important developmental relation exists between joint attention, play, and imitation abilities, and later theory of mind ability. However, very little direct empirical evidence supports these claims for putative "precursor" theory of mind status. A small sample (N=13) of infants, for whom measures of play, joint attention, and imitation had been collected at 20 months of age, was followed-up longitudinally at 44 months and a battery of theory of mind measures was conducted. Language and IQ were measured at both timepoints. Imitation ability at 20 months was longitudinally associated with expressive, but not receptive, language ability at 44 months. In contrast, only the joint attention behaviours of gaze switches between an adult and an active toy and looking to an adult during an ambiguous goal detection task at 20 months were longitudinally associated with theory of mind ability at 44 months. It is argued that joint attention, play, and imitation, and language and theory of mind, might form part of a shared social–communicative representational system in infancy that becomes increasingly specialised and differentiated as development progresses.}
}
@article{FranketalCOGNITION-08,
       author = {Frank, M. C.  and Everett, D. L.  and Fedorenko, E.  and Gibson, E. },
        title = {Number as a cognitive technology: evidence from {P}irahã language and cognition},
      journal = {Cognition},
         year = {2008},
       volume = {108},
       number = {3},
        pages = {819-824},
     abstract = {Does speaking a language without number words change the way speakers of that language perceive exact quantities? The Pirahã are an Amazonian tribe who have been previously studied for their limited numerical system [Gordon, P. (2004). Numerical cognition without words: Evidence from Amazonia. Science 306, 496-499]. We show that the Pirahã have no linguistic method whatsoever for expressing exact quantity, not even "one." Despite this lack, when retested on the matching tasks used by Gordon, Pirahã speakers were able to perform exact matches with large numbers of objects perfectly but, as previously reported, they were inaccurate on matching tasks involving memory. These results suggest that language for exact number is a cultural invention rather than a linguistic universal, and that number words do not change our underlying representations of number but instead are a cognitive technology for keeping track of the cardinality of large sets across time, space, and changes in modality.}
}
@article{HesposandSpelkeNATURE-04,
       author = {Hespos, S. J. and Spelke, E. S.},
        title = {Conceptual precursors to language},
      journal = {Nature},
         year = {2004},
       volume = {430},
       number = {6998},
        pages = {453-456},
     abstract = {Because human languages vary in sound and meaning, children must learn which distinctions their language uses. For speech perception, this learning is selective: initially infants are sensitive to most acoustic distinctions used in any language, and this sensitivity reflects basic properties of the auditory system rather than mechanisms specific to language; however, infants' sensitivity to non-native sound distinctions declines over the course of the first year. Here we ask whether a similar process governs learning of word meanings. We investigated the sensitivity of 5-month-old infants in an English-speaking environment to a conceptual distinction that is marked in Korean but not English; that is, the distinction between 'tight' and 'loose' fit of one object to another. Like adult Korean speakers but unlike adult English speakers, these infants detected this distinction and divided a continuum of motion-into-contact actions into tight- and loose-fit categories. Infants' sensitivity to this distinction is linked to representations of object mechanics that are shared by non-human animals. Language learning therefore seems to develop by linking linguistic forms to universal, pre-existing representations of sound and meaning.}
}
@book{SchallerandSacks1991,
        title = {A Man Without Words},
       author = {Schaller, S. and Sacks, O.W.},
    publisher = {University of California Press},
         year = {1991},
}