%%% File: ./footnotes/Transcribed_Brain_Discussion_with_Randy.tex
%
\footnote{%
%
\rawhtml
<a name="transcribed_brain_discussion_with_randy"></a>
\endrawhtml
%
Transcription of the ./Audio_Track_Brain_Discussion_with_Randy.m4a recorded on January 24, 2020 and transcribed by Google Pixel Audio Recorder:

But then I realized that nobody really understands it all that well. I think there are a few mysteries left to solve. You can probably guess the cartoon version of the model I have in mind; the reciprocal white-matter connections between the cortex and the striatum that work like a high speed parallel bus with gated registers or memory locations at each end and the prefrontal cortex / basal ganglia playing the role of the ALU (arithmetic logic unit) / CU (control unit) comprising the CPU (central processing unit). So what's wrong with this picture?

You can definitely summarize what the basal ganglia plus prefrontal cortex system is doing in terms of LSTM, and in fact, the original paper was really all motivated by LSTM from the computational level and so I think it is an exact fit with just a few key differences that you allude to. It's interesting to go back and read the original LSTM paper, they talk about the number of different units that might be controlled by a given set of gates ... we were recently looking at this to see if anybody in the current literature is ever using more than one ...

... but it seems like that's just what people do ... they use a single memory unit for each set of gates whereas in the biology, just based on the sizes of the respective layers there's many hundreds of neurons, maybe thousands and prefrontal cortex controlled by a single kind of gating signal coming from the basal ganglia and so that's really where this concept of stripes comes from.  If you look at the raw numbers, the globus pallidus in the basal ganglia where the final down select is happening - I can't remember the exact numbers so we have a big spreadsheet somewhere - but it's around a hundred thousand neurons in the human and then the prefrontal cortex is around five billion and so [...]

The numbers are pretty dramatic and you end up having [some idea of how many neurons] a given gating signal is definitely influencing, controlling or modulating - however you prefer to describe it - a large swath of what's happening in cortex and so that relationship with indicating signal and the corresponding kind of representations in the part of frontal cortex that are being influenced by that gating signal is a big question.

But other the [other] core idea about a gating signal that opens up updating of a working memory representation and then allows that to be maintained in a relatively unperturbed way for some indefinite amount of time and then the output is gated to drive the downstream behavior [which is] exactly what the LSTM gives you?

So if you want to just do a machine learning version of the frontal cortex basal ganglia system, just that's what LSTMs are designed for. So the striatum has a huge amount of information coming into it and it's broken up into chunks [what we've been calling stripe clusters] and at the other end of this there is the same number and size set of clusters so that when you gate you can transfer almost any striatal cluster onto any prefrontal cluster using the same offset in the striatum from which it originated ...

The same offset in the sense analogous to registers in the ALU  and they have same relationship [memory map] to one another, with the same strength carried over from the association cortex, but when it finally gets to the prefrontal cortex does it maintain that sort of structure [alignment / strength]?

Well so this is where it gets a little tricky, so we know and that anatomically each area of prefrontal cortex has bi-directional connections with a corresponding set of neurons in the stray them right and so that's each part of the court the free text yeah, so let's just take some you know, random chunk of supplementary motor area or dorsolateral prefrontal cortex anywhere, okay, and you know, again, we'll use this concept of stripe, [...]

Every chunk is whatever size it is that stripe is going to have bi-directional connectivity into a corresponding chunk of striata right and so in some sense that's the kind of registered concept of like this is the gate the striatum is the gate for that part of frontal cortex, but it's also receiving all this other, you know, kind of broader information and you can call this the kind of open loop signal or whatever you want to.

Call it coming in from association cortex that's helping to decide when you should update and maintain or whatever and that chunk of that stripe of frontal cortex. So definitely there is a computer-data-bus-like resemblance, anatomically determined [...] in the frontal cortex that says those guys are directly interconnected.

And then you get to all these modulatory influences from association cortex that say, okay, in this context we want to update here in this context, we don't want to update elsewher - I don't know if that helps clarifying what I'm getting at. I think so. It's not as though the striatum takes a snaphot and after that doesn't allow any transfer of information from the association area while the basal ganglia is operating on that snapshot until it gets Go signal. But rather, that it is continuously updating information in the striatum in a loop that includes the thalamus in determining whether to cycle from NoGo to Go and back.

Yes, absolutely so [...] the current contents of the register / stripe are always a factor in gating, and one thing you could think about that I think is true - although I might not be exactly remembering the details - but that kind of current self projection that the connections from the stripe that is the output is then also driving this input may have a bias to synapse on NoGo and so that would be the idea that if you already are maintaining something then it has a bias to not update and [...] 

[if it isn't] maintaining anything then it's more liable to be updated. I know we've implemented that idea in our models before so that's a sense in which you can you can have some kind of overall maintenance bias to not disturb what's already there unless of course, you know, the other signals come in and say well we want to update anyway ... What about input from the prefrontal cortex?

Does that mean if that's going to be [...] or modulatory signal? Yeah, it is it going to operate directly on the stripes that ended up getting the information from the from the association cortex or is it going to have its own stripes basically they're going to contribute what everybody's got their own stripes and then so that's the closed loop right so everybody's got their own kind of like local chunk of striatum and thalamus that's like, you know, they're they're little [...] if you want to use that term or whatever and then there's always this open loop module [...]

[...] for stuff coming in so if you're a stripe in the supplementary motor area then higher levels and the DLPFC\footnote{%
%
The dorsolateral prefrontal cortex (DLPFC or DL-PFC) is an area in the prefrontal cortex of the brain of humans and non-human primates. It is one of the most recently derived parts of the human brain. It undergoes a prolonged period of maturation which lasts until adulthood. The DLPFC is not an anatomical structure, but rather a functional one. It lies in the middle frontal gyrus of humans (i.e., lateral part of Brodmann's area (BA) 9 and 46). In macaque monkeys, it is around the principal sulcus (i.e., in Brodmann's area 46). ({\urlh{https://en.wikipedia.org/wiki/Dorsolateral_prefrontal_cortex}{SOURCE}})}
%
are going to be providing this kind of top-down modulatory input to your gating signals saying okay, according to my current master plan of you know, what we're going to do in this current sequence of motor actions.

Your transition to the Go signal, and so you can imagine that kind of top-down kind of modulatory signal saying now it's time for this particular sub action to be performed etc that sort of the timing aspect yes, yeah exactly yeah. I'm planning now and getting closer yeah you can go here go just yeah and so one thing that's also really important about the biology that I'm not sure again is also typically captured in the LSTM models these days is that you often want to have a very different control signals coming into the straitum that drive gating compared to the kind of content signals that you're going to maintain in your LSTM kind of active maintenance signal in the first place right and so you know, like in our ...

[...] science paper I think we have a diagram that shows that you know, you want to have updating determined by kind of structural kind of role type of signals that tell you oh this is the agent in the sentence but those gating control signals shouldn't have information about like the specific actual content of that, you know who the agent is it just is this kind of more abstract kind of structural role information and then but that the frontal area [...]

[REFERENCE:
@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.},
}]

[...] itself should get the details of the content who is it that is the agent of the sentence and that gives you this kind of critical kind of structure content dissociation that the gating signals are driven by kind of more structural information and then the frontal cortex is encoding kind of this the semantic content and so that you know, we have actually a recent paper that Jake Russin and my lab as has been doing that, you know is a kind of different take.

[REFERENCE:
@article{RussinetalCoRR-19,
       author = {Jake Russin and Jason Jo and Randall C. O'Reilly and Yoshua Bengio},
        title = {Compositional generalization in a deep seq2seq model by separating syntax and semantics},
      journal = {CoRR},
       volume = {arxiv:1904.09708},
         year = {2019},
     abstract = {Standard methods in deep learning for natural language processing fail to capture the compositional structure of human language that allows for systematic generalization outside of the training distribution. However, human learners readily generalize in this way, e.g. by applying known grammatical rules to novel words. Inspired by work in neuroscience suggesting separate brain systems for syntactic and semantic processing, we implement a modification to standard approaches in neural machine translation, imposing an analogous separation. The novel model, which we call Syntactic Attention, substantially outperforms standard methods in deep learning on the SCAN dataset, a compositional generalization task, without any hand-engineered features or additional supervision. Our work suggests that separating syntactic from semantic learning may be a useful heuristic for capturing compositional structure.}
}]

On that same idea but it's similar kind of emphasis on this idea that you really want to have ways of dissociating essentially the structure from the content and so having different inputs into the gating signal versus into the gating pathway versus the the kind of what's actually being maintained internal cortex pathway can can give you that so this is blinding study talking about basically that allows the strong cortex to define something yes, absolutely in the exactly yeah exactly in a kind of [...]

Structural way right in a functional way so that you're binding these kind of role filler binding exactly right it's exactly that it could be okay [...]  I just wanted to ask a quick clarifying question so I guess recapping our summarizing we've been talking about so far so also one of the, things that we've been discussing is a group is right for the PBWM model exactly how it maps to a LSTM, so I guess what you're saying here right is like.

The gaing system in an LSTM, right so you have the concept like an LSTM unit right you might have drawn a block but that unit has like mechanisms the idea mechanisms themselves in the units would be kind of the analog to basal ganglia totally yeah, okay cool so with the LTSM even though I guess is like computer scientists, we tend to think of it as one like quote-unquote units in terms of how that kind of splits into like the biology.

You kind of have like the thing that's being maintained the content yeah that the like so-called like registers or stripes right that are being obtained in frontal cortex and the actual gating mechanism itself is basal ganglia actually absolutely yeah and I think you know, it's it's kind of surprising to me that people haven't picked up on this idea because I think it's just a huge amount of opportunity there that you know, if you think about computers, right?

I mean the whole thing about a gate. In a computer is that you have a different signal coming in and the control, you know wire versus you know, what's being gated right and that's essential for it to do something interesting and so you know, [...] it seems to me that all the LSTM models are just having all the same inputs always going to the gate and to the to the kind of content and not really exploiting this ability to have a different signal coming into the control versus what's going through the content.

Yes running with the LSTM model where would the learning signals come from to learning yeah a great question this is where LSTM has a big advantage right so you just do backprop and so, you know the brain I don't think has the luxury of kind of you know, computing those gradients over all the time delays, you know, and so but it you know in a computer you can do that so.

Go for it and so you know, that's where you know, if you care about the science and you you want to really exactly understand how the brain works that's a problem, but if you just want to have a model that kind of, you know, maybe superhuman and does a little bit better than what the brain can do then just do backprop on those signals and you're you're probably going to do much better so that's what this coming really the LSTM just just just doing backprop on the gating signals is great.

It's like a cubist signal go for it. Say we did want to stick closer to the clients yeah are there I guess candidate the. Learning signals might work or alternative for this that are more biologically realistic absolutely so that the current idea which you know, we haven't actually published except in the kind of notes in the updated simulations for our textbook is that you have a kind of trace in the basal ganglia and so whenever you get a gating signal taking place administrator those synapses essentially flip a bit that says we have gated, okay.

And that bit stays kind of around until the next basic dopamine signal and so the idea is basically that you have this kind of trace memory in the synapses of who did gaiting and then later you get a dopamine signal that says, hey something good happened or something bad happened and the credit assignment can then kind of come in where the dopamine then you know has an effect only on the the synapses that have been kind of activated with this trace and there's actually really good biology.

Evidence for this kind of thing that Richard Morris originally was somebody who published work on this showing these kind of synaptic tags is what they're called and so that actually is pretty effective because and it's it's it's more or less like LSTM in the sense that you get to hold on to the gating signal kind of in indefinitely until you know, I mean, obviously there's some real biological time limit but it's probably pretty long for the synaptic tags like on the order of several minutes to maybe [...] an hour, something long enough that you'll get some kind of, you know consequent dopamine signal in that time frame so it does allow you to kind of spam that credit assignment gap and say this gating event that took earlier in time was resulted in either a positive and negative outcome.

[REFERENCE: Synaptic tagging, or the synaptic tagging hypothesis, was first proposed in 1997 by Uwe Frey and Richard G. Morris; it seeks to explain how neural signaling at a particular synapse creates a target for subsequent plasticity-related product (PRP) trafficking essential for sustained LTP and LTD SOURCE: https://en.wikipedia.org/wiki/Synaptic_tagging for MORE: do a Google search on "Richard Morris synaptic tagging and capture" ... see Moncada et al review paper for an up-to-date summary:

@article{MoncadaetalNP-15,
       author = {Moncada, Diego and Ballarini, Fabricio and Viola, Hayd{\'{e}}e},
        title = {Behavioral Tagging: {A} Translation of the Synaptic Tagging and Capture Hypothesis},
      journal = {Neural Plasticity},
       volume = {215},
        issue = {650780},
         year = {2015},
        pages = {21},
     abstract = {Similar molecular machinery is activated in neurons following an electrical stimulus that induces synaptic changes and after learning sessions that trigger memory formation. Then, to achieve perdurability of these processes protein synthesis is required for the reinforcement of the changes induced in the network. The synaptic tagging and capture theory provided a strong framework to explain synaptic specificity and persistence of electrophysiological induced plastic changes. Ten years later, the behavioral tagging hypothesis (BT) made use of the same argument, applying it to learning and memory models. The hypothesis postulates that the formation of lasting memories relies on at least two processes: the setting of a learning tag and the synthesis of plasticity related proteins, which once captured at tagged sites allow memory consolidation. BT explains how weak events, only capable of inducing transient forms of memories, can result in lasting memories when occurring close in time with other behaviorally relevant experiences that provide proteins. In this review, we detail the findings supporting the existence of BT process in rodents, leading to the consolidation, persistence, and interference of a memory. We focus on the molecular machinery taking place in these processes and describe the experimental data supporting the BT in humans.}
}
@article{RussinetalCoRR-19,
       author = {Jake Russin and Jason Jo and Randall C. O'Reilly and Yoshua Bengio},
        title = {Compositional generalization in a deep seq2seq model by separating syntax and semantics},
      journal = {CoRR},
       volume = {arxiv:1904.09708},
         year = {2019},
     abstract = {Standard methods in deep learning for natural language processing fail to capture the compositional structure of human language that allows for systematic generalization outside of the training distribution. However, human learners readily generalize in this way, e.g. by applying known grammatical rules to novel words. Inspired by work in neuroscience suggesting separate brain systems for syntactic and semantic processing, we implement a modification to standard approaches in neural machine translation, imposing an analogous separation. The novel model, which we call Syntactic Attention, substantially outperforms standard methods in deep learning on the SCAN dataset, a compositional generalization task, without any hand-engineered features or additional supervision. Our work suggests that separating syntactic from semantic learning may be a useful heuristic for capturing compositional structure.}
}]

And so it's sort of like a poor man's version of the LSTM backup. And what parameter is getting changed is there the basal ganglia has a lot of circuits that are dedicated to sort out both storing these kinds of things and maintaining them in order to say make a change if I get a big dopamine first. In the simple version, you just have to Go and the NoGo inputs into the striatum and so basically you're encoding the synaptic trace on each of those synapses coming into and going.

You get the dopamine signal and it has these opposite effects on the Go and NoGo. So if you get for example a positive dopamine signal averse then you know, the Go synapses get increased and the NoGo synapses get decreased and vice versa for the dopamine dip, right?

And so it actually really goes at the tunable parameters which then will increase or decrease the chance of doing that same gating action according to those same kinds of inputs next time around. And I think that you said there's addition to make sure so that the Go NoGo that's really like a big multiplexer.

You can it's gonna it's gonna sort of suppress or go to some set some subset of the things that we're currently active. Yep, and it can do that independently for a range of different things. Absolutely, yeah. Variable by Go ahead. I before we move on from that like kind of so you talked about on basically this idea of a trace and dopamine providing some sort of reinforcement learning yeah.

I can't help but draw a like a connection to I guess RL models that we are using like our model. Yep, there is a lot of sample inefficiencies yes, so like I would expect I guess the brain this mechanism. Needs some way to solve that sample and efficiency problem.

Otherwise we're gonna need so many dopamine bursts over and over again just be training those like kind of synaptic traits. Yeah, it is it is absolutely a key kind of weakness that you know, we're depending on these, you know, really weak RL kind of links in the learning chain there.

And you know, so there's a couple answers to that potential answers to that. One is that you know to the extent that your cortical inputs are doing something more like backprop and they're developing kind of high level, you know optimized, you know, translations or transformations of the sensory inputs into the dimensions that are behaviorally relevant then you kind of make the gating decision in the basal ganglia easier.

So in other words, if you basically have in your cortex, really nice differential representations between things you need to Go to like, you know, positive reinforcing food and good things and you have other really distinct representations in your cortex for things that are bad that you want to avoid then the gating learning in the striatum becomes really simple.

And so, I think that's clearly what's happened in people is that we have you know, have this huge cortex that is developing all the right kinds of abstractions, you know encodings of the world to make the basal ganglia's job dead simple because it's so lame.

If all you have is the kind of reinforcements that signal ecologically and evolutionarily, that's what it started with and so you try to move as much of the job off into the cortex as we can and still use that kind of core basal ganglia as a kind of residual thing and it does give you a different take on .. it provides a different kind of emphasis on what decisions you're making and in some sense.

I think you know ecologically having it really tied to a dopamine signal probably is a good idea like you could think all kinds of interesting abstract thoughts, but at the end of the day if it doesn't pay the bills, so to speak or it's going to get you into trouble then you kind of do want that basal ganglia step to say well, maybe that wasn't such a good idea.

And so one of our most recent papers that we just put on archive that Seth Herd is the lead author on kind of takes this idea and shows that you can really put a lot more of the work into the cortex we call it the proposer in this model so we say that the cortex is developing these proposals for action and that's really where all the hard work is happening the smartness and then the basal ganglia is just this kind of final yeah, okay, that's okay, that's good or not.

[REFERENCE:
@article{HerdetalCoRR-19,
       author = {Seth Herd and Kai Krueger and Ananta Nair and Jessica Mollick and Randall OReilly},
        title = {Neural Mechanisms of Human Decision-Making},
      journal = {CoRR},
       volume = {arxiv:1912.07660},
         year = {2019},
     abstract = {We present a computational and theoretical model of the neural mechanisms underlying human decision-making. We propose a detailed model of the interaction between brain regions, under a proposer-predictor-actor-critic framework. Task-relevant areas of cortex propose a candidate plan using fast, model-free, parallel constraint-satisfaction computations. Other areas of cortex and medial temporal lobe can then predict likely outcomes of that plan in this situation. This step is optional. This prediction-(or model-) based computation produces better accuracy and generalization, at the expense of speed. Next, linked regions of basal ganglia act to accept or reject the proposed plan based on its reward history in similar contexts. Finally the reward-prediction system acts as a critic to determine the value of the outcome relative to expectations, and produce dopamine as a training signal for cortex and basal ganglia. This model gains many constraints from the hypothesis that the mechanisms of complex human decision-making are closely analogous to those that have been empirically studied in detail for animal action-selection. We argue that by operating sequentially and hierarchically, these same mechanisms are responsible for the most complex human plans and decisions. Finally, we use the computational model to generate novel hypotheses on causes of human risky decision-making, and compare this to other theories of human decision-making.}
}]

Maybe we shouldn't do that so it makes the the learning job for the basal ganglia much it's a much more you essentially have it you reduce the dimensionality of what that with the basal ganglia needs to learn and therefore the sample and efficiency becomes less of a problem so you say the proposer though, but the proposer is association cortex or free funding well, it's it's a you know, we we kind of leave it a little bit ambiguous because it's probably a lot of things you know, so in.

In you know in a motor action we think you know for example that the parietal lobe is gonna be representing an anticipated outcome of a motor action like, you know, you're about to reach for something in the parietal lobes is wait a second, you know, if you actually did that action here's where your hand would end up and maybe that would end up over the flame of the stove or something and so you'll get that kind of feedback representation of who this is going to be the outcome of that and that's part, you know, that's also we call that the predictor.

And you know there's kind of this a little bit of a slippery slope there where we think that in fact as you're proposing your motor actions, you're evaluating them through these cortical loops and at least certainly for one step kind of outcomes you're probably not choosing or kind of vetoing actions that lead to that outcomes just right there in the cortex before you even really kind of say, hey basal ganglia, this is what I came up with and so that's again where you can use, you know, this kind of notion of constraint satisfaction.

Processing. So you've got ideas bubbling up for motor actions that you might want to take you've got other cortical areas that are kind of reflecting oh well if you did that here's what might happen and then that deeds back and suggests well maybe we should try some other actions and so this kind of proposer predictor dynamics in the cortex can be pretty smart and you know, like a hot field network do this kind of constraint satisfaction process so that you only end up sending down to the basal ganglia something that you know is pretty reasonable.

And so that's another way in which that you know, it it's a lot smarter than just kind of pure dumb RL basically yeah trying to go ahead sorry. I just want to summarize it sounds like you're answered to the question is fundamentally about transfer because we have these deep dense representations in the cortex yeah, it's very rudimentary signals in the basal ganglia a lot for basically yep.

Yep. and relative to the boosters action selection or perception action cycle yes the in the background it's always just basically trying to so the the the gradient descent stuff is working there all the time and as soon as as long as you're in this space, you're filling it in all the time different there's no particular action, you're gonna take in some of the states state you're getting you can really capture that state if you need it yeah, yeah.

That twister stuff is very influential. I mean, I think that kind of pervades everybody's thinking about about PFC and the hierarchy, you know of these different levels of abstraction within those loops and the stuff you mentioned, you know, the David Badre and Etienne Koechlin and all these other people, you know, talking about these hierarchies.

I think that's very appealing and almost certainly has to be true that you know, that's how you do more complex kind of outer loop stuff and, Ask that within these interloops etc so so I think that almost everybody, you know is on that same page in perspective that so one of the big problems because it's a partially observable system is that you're often going to have to anticipate what you're going to do and then for some perceptual act yeah, yeah happen is that in the in the motor cortex or is it yeah, so that's again where we think that really is in the in the parietal lobe and so yeah, that's so that given that recent paper where we have this notion of a predictor along with the proposer.

Tries to kind of articulate that idea a little bit further that and you know, one of the main things I've been working on recently is this kind of predictive learning framework in the cortex more generally and so the idea is that, you know, the the circuits in the cortex are constantly trying to predict what's going to happen and in particular parietal lobe is really well situated to get you know signals from motor cortex about you know, a potential actions that you're considering and through prior experience, you know, has basically learned.

To predict if you. This action here's what would happen right and so you you just constantly getting that loop between possible courses of action in the frontal cortex, we're driving predictions of outcomes in in parietal lobe and so you always have available essentially in your decision-making like here's what would happen, you know, at least a one-step prediction of what would happen if I took this action, you know, and there's a question of like okay now if I want to do multi-step, I'm gonna have to you know, run a simulation in my head that's that's more elaborate or whatever but I think certainly for for you know, kind of immediate outcomes of any given action those are kind of automatically being activated and parietal lobe and then that speeding into the overall GO NoGo decisions just kind of automatically yeah, so I think that's a critical element and again something for which there's a lot of support and then that actually ties right into the cerebellum, so you know, the the cerebellum is generally thought to be this kind of forward model idea and what I just described.

Was kind of prediction of outcomes of motor actions is exactly a forward model right and so again in the motor control literature everybody believes that forward models are essential for for doing any kind of realistic motor action and so so so that the surveillance can be doing more I think more fine-grained quicker time scale forward model representations and learning and then the cortex may be is operating on a slightly slower time scale longer time scale.

But also, you know more high-dimensional kind of more elaborated representation but those two really do work together and there's a huge projection from parietal lobe into cerebellum huge projection from cerebellum out to motor cortex to basically provide training signal we think in in the thalamus in particular so that the projections that come up from the cerebellum synapse onto the the thalamus which then drives a kind of learning signal.

We think in the motor cortex and so essentially the the cerebellum can by itself anticipate a outcome of a potential motor action and send a kind of corrective signal up to motor cortexing yeah you should probably not move your hands so quickly because you'll probably want to you know, that's going to lead to an overshoot so you want to move it more slowly so here's the here's the better motor action that you should be driving right now.

And that corrective signal can then actually drive learning in the cortex so that the cortex, you know, soaks up that same kind of corrective signal so essentially the idea is that the cerebellum the forward model in the cerebellum teaches the the the motor plan in the cortex to kind of correct for anticipated errors that explains a lot so essentially if I understand you're saying, Be the motor cortex, you know, the premotor supplementary eight more primary that's basically going to be the the the.

Source of the of the programs of you know program is going on those programs automatically by itself yeah and get feedback to say to say stop or go or correct or whatever come either these little ganglia or that cerebellum yeah and actually so what I would really say to distinguish those to is that the cerebellum tweaks the motor program, okay, so it says well, you said, you know exert this amount, of course, you know, so the original cortical program was like exert this amount of force and then the cerebral.

Um can it goes well you know you might want to drop that by that right so it's it's tuning the actual kind of vector motor signal whereas the basal ganglia because of this really huge down select in the in the number of neurons in the in the globus pallidus is a much more coarse-grained kind of like do it or don't do it decision-making kind of signal and so I really do think those have complementary roles these of you the cortex so that the basal ganglia is really this kind of go no-go decision-making process, whereas the cerebellum is like the the little editor in there just.

Like, you know, copy editor fixing all the bugs in your motor program. So it is the the what buckets gated out of the basal ganglia into the motor cortex is it is it more like some command or is it really sort in some sense a fixed point or a set point of a you know, in the sense of you know, here's the state you want to go for go yes see this is where I think it's really important also to think about the LSTM analogy if we think about these as gaiting signals then they're not content right they're multipliers, they're they're literally gates right they're multiplicative and therefore they're not any content they're not they.

Beg ye is not saying here's what to do, the basal ganglia is essentially saying when to do something right, so it's basically again that kind of decision-making step of now it's time. Cortex, you know pfc motor cortex do whatever it is you want to do but now is the right time and so it doesn't it doesn't have again because of this massive down select in the GP it doesn't have the bandwidth to communicate any meaningful signals up to PFC it can I can only provide this kind of modulatory dating multiplicative type signal that's me does it does it make sense to say that the input to the basal ganglia is kind of large scale large scale.

Context that gets kind of master modulated like you said by the basal ganglia such that kind of the the output of the basal ganglia is also context but it's not it's not content that basically generating. Content goes to kind of select or kick off and action or something along those lines or is it more like ganglia receives like like a single like I don't know move your hand forward commands or something along those lines and then it either inhibits or enhances that.

Yeah, I mean again, I think I really do think it's it's it is inhibited enhance is the right terminology for this kind of you know, it's it's fundamentally modulatory on what's happening and in cortex and so and you know, that's that's actually biologically if you look at the connectivity this disinhibition of thalamus is actually you know, mathematically perfect for like a multiplicative effect so so, you know, it's like whatever if you're disinhibited, whatever you're gonna do.

Otherwise just kind of happens, but if you're inhibited you're kind of basically multiplying by a zero and you're turning it down, so it really is in that role of kind of gain modulation or whatever you want to call it just a multiplicative role and so and so yeah that makes sense in terms of like not providing content, but providing that kind of modulation or orgating.

So so whatever it is that cortex is thinking about doing is all kind of within the cortex within the prefrontal cortex areas and so the basal ganglia really is just is fundamentally not I think changing that content it's just multiplying it by this getting signal but it's really it's really like a puppet yep yeah yeah, it's pulling the strings exactly yeah, but you conductor arms is another oh you do it yes first to do yeah, yeah, the conductor is really the best metaphor because it's like, you know, The conductor is not playing the instrument, you know, it's not actually doing conductors not doing anything in particular, but they are saying now now and how much you know, that kind of stuff okay, yeah.

Yep. See what's next on your list here. Sorry every and I guess like you mentioned before also it's just a recap like it really is kind of like the the frontal frontal cortex proposes a proposes a commands yeah our roses like an action yeah, it's necessarily the case that.

Basal ganglia kind of receives and then inhibits and like by so inhibiting kicks off an action that's yeah, yeah cool. So so this is just to clarification. I think you've already said this but in some sense it's not as though you know, you can define a variable and then set it basically it's that you just have.

The system is divided into you know, the slot fillers for just a lot of different kinds of slots and combinations, you know, yeah and you're just gonna use those and the other system when in order to make use of it it's just gonna have to know that I have to I have to use that register basically right input or output for figuring out what the variable by exactly and that's you know, it really is like if you think about you know, assembly language kind of code on a on a chip, you know, it's like you have.

To know that you stored the address to go to in this register and not that registered, right? And so you've got all these registers but what role those registers are playing and any given situation essentially has to be established by convention and then you know, if everybody's following the rules then you stick the information into the right register and the other guy who reads it out of that register and you're good to go, right?

So so if you have an operator basically, it's gonna note say the division operator even those were the dividends gonna be exactly yeah all. Has to do that register yeah. I know if you have to do that in in the ALU or not, but probably you do at the very end.

I imagine yeah. Yeah.

So this business about abstract thinking what well loaded word abstract, right? Yeah, then you and then you see you know again Padre all of them have a different like it was doing you know, in terms of that different track. Yeah. One thing that strikes us is you know we keep turning back to this application, you know, automatic programming.

Yeah, and that obviously we don't have evolution hasn't caught up with us yet, you know, whatever, you know, wherever that stored wherever that the bits are stored whether the parameters are stored it's it's not new it's some old keys yep so so lots of things could happen, you know, not only you could do programming you could become a a chef, you know, right and now you gotta do all the cooking things, you know, yeah where did You what do you look for is you know, like are there some additional?

I mean all I can't imagine that you have a enough of these big chunks of stuff that you go allocate a new one each time so it must be that you sort of say, I'm gonna put it here and if I screw things up a little here, it's not so bad yeah yeah and I think you know, I mean, This is where it it it gets really challenging because I mean intuitively I think we have fairly general purpose kinds of you know, sequencing representations and control structures that that we develop and so like if you're following instructions to cook to do a recipe right that's a sequence of actions and if you're following instructions to write a program, you know, whether they're external layer internally generated.

That's a that's a sequence of actions as well and so if you have some kind of general purpose again abstract, whatever that means representations that emerge that allow you to kind of control a arbitrary sequence of actions and then there's this kind of binding process that says well the particular actions that I'm doing in this particular case are cooking actions, but I could just as well have done, you know, the programming actions using those same, you know, the same hardware then then you get.

To that kind of. Essentially like, you know, touring machine kind of sequential flexible combinatorial kind of system in your brain that can basically allow you to do all sorts of different kinds of sequences of motor actions or cognitive actions and that gets you into that kind of, you know, reusable general purpose system that can be done can be used for any different kind of domain and I really do think we have that we are to some to some degree of approximation, you know.

A general purpose symbol processing system in the outer loop but the inner loop is this kind of amazing parallel processing system right and so you've got kind of the best of both worlds you've got this really robust flexible kind of perceptual motor system, but then it also has this general purpose turning machine like out or loop where you can can kind of you know mix and match any of those different inner loop elements with each other and try to get new new behavior to take place right so if you know, X number of recipes, you know doing a new one is not that hard because you just mix and match the elements that you've already learned from previous ones, you know, so that kind of classic commonatorial, you know, generativity type of issues about how do we how do we go beyond what we have slavishly learned, you know by trial and error really I think does have to depend on that kind of sequential, you know, recombining the existing elements to get new new behavior stuff.

And so that's so that makes sense to intuitively but how does it actually work the challenge right so instead of thinking about I'm gonna get a bunch of parallel circuits for doing these high level abstract things I'm gonna get done big thing yeah we handle all kinds of things and switch yeah different tasks exactly and so for for example, you know, you wonder with all the machinery is for doing relational, you know, sort of a graph next kind of thing yeah that must be distributed around a bunch of stuff, you know, yeah places could do that, but then on the other hand maybe some of the you know, this this the, Kind of imagination-based planning stuff you know where you have to do prediction and you have anticipate things you've never seen before maybe that'd be general too, wouldn't it?

I think so. I mean, again, it's like this mix of what's general and what specific so you the the ability to orchestrate these these specific things kind of sequentially I think is the key right so that you can you know, you've got lots of different domain knowledge that is kind of hard-fought, right?

I mean, you have to you have to get that domain knowledge through trial. And error and you know experience etc but once you get it if you have enough of this more general purpose kind of outer loop stuff going on where you can now say okay, well let's take this new concept that I've just learned about like calculus or something and try to match that into this other, you know space that I have learned about you know, so you know that ability to kind of take information and sort of, you know, move it flexibly around the the to these different kind of learned.

Pathways in the brain I think is what makes us flexible but it's also again you get all these binding problems coming up always right and in computers, you know, the binding problems are just not there because you don't you know, you just address things by memory and it there's instant instant binding so I think yeah, so anyway, that's where I always get kind of stuck is you know, they the idea sounds sounds very compelling but you have this general purpose kind of cognitive abilities at some level but actually getting those.

To work and in actual neurons is always the challenge and the giving take something like. Sort of syntactic stuff, you know, like yeah language you need syntactic stuff you eat, you know, you yeah do it mathematics as well yeah and but probably have to do it from lots and lots of things it's just that we don't have pain for them yes exactly.

I mean all of motor actions and you know, if you think about schemas everything has a kind of syntactic level right these kind of. You know structures or roles or you know, if you if you look at you know, in what syntax really is in language a lot of it is very verb centered and you know, what is a verb well, it's really motor action and what is what is the syntax well it's about what are the arguments to that verb like who gave what to whom, you know, those are the argument structures for an actual kind of motor action and so, you know, I do think that all this stuff grounds out in in real kind of.

Physical you know sensory motor kind of representations that we learn and we just learned to encode the kind of general kind of roles in those in those motor actions about you know, giver receiver those are roles and somehow we develop the ability to encode that as an abstract syntactic kind of knowledge structure and then we can bind any thing in the giver and the receiver role and then you know your gold right and so if you have enough of that vocabulary of different kinds of structure-like representations.

Like that and you have that ability to do the binding then you know, presumably that gets you enough enough flexibility to do these kinds of you know, more general purpose cognitive functions, so yeah the so there's so I like that idea that somehow outside of the of the the federal cortex anyway there's storage for for procedural information and declarative information and we can just keep that sort of separate because it's just knowledge.

Yeah and but then in the cortex we put it all together and then it is complicated because now we have to account not just for specific things like knowledge of x and y and z we have to account for every little minutiae of you know, our life and our interaction with the world etc and now you can have problems wherever you try to start learning you get because you want to do transfer learning but you have you know, Forgetting and interference and all those kinds of things occurring there as well, yeah, yeah.

So the cerebellum and the and the and the hippocampal complex they look a lot they look like they they came from the same mold and then yeah a little bit yeah exactly and that was the amazing thing that Mar hit on, you know, the series of papers that is really this one idea that kind of applies can you make that analogy?

I mean, you just saw this a great story about what the cerebellum does yeah, yeah is there a like kind of story you can tell about how the the hippocampal complex? To tends all this and you know what so there's two there's two dimensions to it so one is like, you know this kind of forward model predictive learning idea and that people have definitely tried to take that approach for the hippocampus.

I'm not a hundred percent convinced about that but it's it's certainly a possibility but the other dimension to it is just really the same insight as the SVM right which is that problems that are hard to solve in a low-dimensional. State space become so much easier in a high dimensional space and so basically if you just blow things up into a huge height eventual space it's easy to kind of slice off these hyperplanes and divide out the different categories of things that you care about and then you know, you you compress it back down after you've done the carving up and this side dimensional space and now you've got you know, your your behaviorally relevant categories nicely kind of separated for you and so I really think that essential insight of Let's put everything into a super high dimensional space is common between the cerebellum and the and the hippocampus and that is essentially what what Mars core insight was, you know, is that both of these things have these granules cells and there's so many of these granule cells that it really produces this crazy high dimensional representation and so that's really what it's common is the that both depend on this, you know, blowing up into a super high dimensional space.

It's interesting that you can blow away your cerebellum and you can still manage pretty well but this is the thing that's where I think the cerebellum fundamentally and you know, there's an analogy there too, which is you know, the cerebellum is fundamentally teaching the motor cortex like to fix the bugs in your motor program, it's your editor, right and I I find it's myself.

I've submitted enough papers that I now fix many grammatical errors even before I submit the paper because I've seen enough of the editor comments coming back to say, oh yeah well that. Stuff. I should probably just fix that right and so you incorporate that those corrections into your original production and so the same thing happens in what the cortex it eventually learns kind of what the cerebellum has to tell it and so yeah if you're an adult and you lose your cerebellum, not that big a deal.

If you're a kid and you lose your servo and you're screwed and that's kind of a big difference because then you don't get all those nice training signals and then with the hippocampus same kind of idea right you you have the system that can learn episodic information really quickly and then theoretically it can be teaching the cortex, you know, hey, here's here's some facts we've learned, you know, learn to learn those yourself, you know, and so eventually the cortex kind of soaks up some of those facts so set that kind of teacher student relationship between both the Cerebellum and the hippocampus as teachers to the cortex is very common as well and so the only difference at that level then is you know, the hippocampus may be teaching about you know, episodic facts whereas the the cerebellum is teaching, you know, anticipated motor errors basically so the content of what they're teaching is different and kind of, you know specific to different parts of the brain and different functionality, but that same kind of relationship is still present but things in perspective we can say go ahead.

I will. Hope to point out I mean we've talked a lot about the similarities between cerebellum and people campus both having kind of patterns decorations. I mean, I would say that you know, the cerebellum has even more dramatic separator dynamics by having the smallest and largest cells but yeah, it will campus also has a tractor dynamic which yep a cerebellum does and then differentiator totally yeah, absolutely because you know, when you want to retrieve those memories you need to the the attractor dynamics and I don't think there's an equivalent of retrieve the memory.

In cerebellum, so it doesn't need that so yeah, that's another key difference, yeah. and yeah, I mean we sympathize with HM, you know because he's lost all of his episodic memory yeah ideas, you know, he's an adult and so a lot of stuff all programmed in and he'll exactly a long time, you know, yeah wake up every morning and think it's the same the different yeah a morning all over again, yeah yeah, it's not so bad yeah pretty bad it's amazing it's amazing how much of cognition is intact in the absence of the hippocampus are truly it's yeah.

Yeah kind of relating back to what we're talking about a little bit earlier about applying these positive architecture to kind of pass like abstract thoughts or whatever, you know agent or artificial intelligent application they have one challenge that I've always seen with these architectures is that they kind of provide a high level structural framework for like what should connect to what but not a great way to kind of.

Control or infect the fine-grained dynamic. Are like running through these very systems. And on other words, you have this like macro architecture disconnect to this this and like, I don't know how. You'd deal with that because it seems like if we're trying to apply this in an agent, we need to be able to inspect our control finer grain dynamics.

Needs dynamics, you mean the dynamics of the neural system right the activation the right like what what? The software. Running on the wires, yeah. Yeah, I think that's the biggest the biggest challenge actually is you know, if you if you look at what people actually do with models these days, I mean, it's the reason everything is so kind of, you know narrow AI centric, you know is because you can't go in and kind of control the microdynamics of what's happening in these models in a flexible way, right?

I mean, they just you you've got an input pattern and it goes through a bunch of layers and it ends up being transformed into an output pattern. And boom okay it does that now if you want to have that same thing do something else, you know, it's a whole retraining and then it's lost the ability to do the first thing, you know, catastrophic interference etc and so you know, again this idea that we developed this kind of ability to dynamically reconfigure and kind of control the flow of activity in these parallel networks and our brain is really easy to state it's intuitively, you know, sort of appealing but actually, Getting that to work in a real neural hardware kind of framework.

I think is the challenge and it's just nobody nobody's doing it because it's such a pain in the ass. I mean, so you know, if you if you try to make a network even just a basic neural network fully recurrent so that you know activity in one place can flow in and influence over here and then in the same network, you know, depending on what you're doing, this could be the input and this could be the output in other words true fully by directional processing nobody.

Does that in the current models like you know your classic type models because just even that step makes these things kind of chaotic and so that's that's that's kind of indicative of the the the kind of challenge there that yeah having control over the dynamics of the model is the problem.

Yeah, you're on the right trail but. Yeah, I'm gonna say like when you say that's the biggest challenges or like what the challenges are the frontiers. I I guess I'm interested in the context of how your research kind of what what you're facing right now totally that's it. I mean, that's a hundred percent yeah.

So you know again you can you can get you can train any network to do anything but getting that to happen in a way that is kind of dynamically reconfigurable and controllable with you know, some kind of contextual top-down control signal telling you oh in this situation I should be doing X Y and Z and this other situation I should do these other things and having the system kind of have that dynamic kind of reconfigurability and and everything is just yeah, very hard.

So, I mean to imagine. A build a robot. I I don't want to have to like take this robot through all the life cycles and let it. I need to. I'd say cooking a recipe as you were yeah. Yeah. Yeah. Yeah, totally totally sorry okay. I was just saying I think I followed the year whenever you're talking about like I need to program like like cooking.

Sorry, how does that follow? So let's say you wanted to build one of these agents you use the cognitive architecture and you're trying to get it ultimately to say be able to be like a Christmas profile. What you have to do is you would have to. Train this road.

Likely in some multitasking whatever but now when you actually want it to say put an egg you don't really have any control over what it's doing like, maybe the most. Condition on the task, so like I mean that seems to be like a dominant paradigm right now, we're like given the task you like condition in some way but like that's not really fine grained control over the dynamics.

That's exactly right. If it's at the task level - the whole task, then it's not going work. You need you need to be able to work with smaller components of activity in order to achieve a finer grained level of control and then having that control, you can mix and match different motor actions in ways that make sense to produce coherent behavior [...] that's the challenge.

Any ideas? So to me, the number one thing that we're working on and that I think is missing - of course there are many missing pieces - is the development of a kind of goal representation, essentially an encoding within the system of the desired outcome. 

To me this is the key difference between what we see in all the existing AI models and what we feel is so essential to how we behave. We do things because [...] and this feeds into issues relating to consciousness, we know what we want to do; we're doing it because we want some particular outcome, we're not doing it just because we happen to have done it before, like a zombie slave to a habitual system that it can't resist, we're doing it because we have an internal representation of a desired outcome - a goal that we want to achieve, and that by having that kind of goal-level representation [...]

I think that's one critical element for getting the system to be able to use a goal representation to drive behavior [...] so essentially you learn a bunch of actions by exploring your environment, randomly babbling or whatever, doing exploratory learning and through that process you're learning that if I did execute this sequence of actions, then I would get to this goal, so I guess that I want that goal.

I'm now going to be able to drive this sequence of motor actions to get to that goal and if I don't get that goal, I know what I was trying to get and I know what I got and now I can do a lot more, for example, intelligent learning about the difference between where I thought I was going to get to and where I actually got and where I went wrong. I think this opens up a lot of possibilities for smarter [...] learning than dumb reinforcement learning in which you proceed by trial and error, all the time bumping into walls, [...] that's just not characteristic of how people learn. So one research route I've seen people take involves a form of unsupervised segmentation of activity traces.

Watching an animal perform a task, I probably wouldn't believe that it is pursuing a goal and planning a sequence of actions to perform - for the most part, animals are just dynamic reaction machines. I think the reason why we have this intuition for really nice crisp goals is because of language. Not only does language provide a way for us to segment our experience early on but it also provides a common ground for other agents to trigger dynamics in us. So we're seeing two problems here, not only would you be having goals in your agent like [...] through the representation of [...] but you also have a way for us to now control [...]

I absolutely agree that language is central but I will disagree in in the appraisal of other, you know animals that I think they, Generally speaking when we look at in like rats running mazes, you see this really clear signature that you know before they start the maze the basal ganglia gating happens at the very start prefrontal representations, there are you know, rudimentary prefrontal cortex in a rat and it kind of loads up a plan and the rat, you know, basically follows that plan and so, you know, the evidence suggests that they they're they are far more teleological, you know, ends driven gold driven then then I think, We tend to give them credit for you know, and in and you can see it, you know, if you have pets or whatever dogs or particularly, you know attuned to kind of, you know thinking about what you're trying to do and how they can get what they want yeah yeah, so I think there is a lot more of this kind of kind of top-down gold-driven type of stuff even in you know, simpler brains we take care of a lot of feral cats and you know those cats know which windows to go to in order to yeah.

Right yeah yeah and when we feed them so so Joe will sometimes get you know, really fancy food those little tiny things yeah that you know and the next time we feed them friskies and they don't come back for a week that's planning yeah. I have one more question that and just we don't have much time and you're already going for a long time, so um, We're trying to we're thinking about implementing a version of boosters hierarchy yeah, maybe it's a shallow hierarchy but sort of learn it yeah when we get down to the bits and we look at those reciprocal connections this is something you refer to before or we don't know how to do that recurrent thing in a completely totally and no and as frustrating because it's should be simple, you know, yeah, they're like predictive coding so we predict forward yeah and then maybe we're hoping that somehow we can get the inverse out some other way but it's, The easiest it's not be pretty.

I don't think yeah, okay yeah, all right I got that what about what about development do you think we're gonna have to recapitulate, you know childhood development and in order to program some of these things right but I I, you know, I yes, I it certainly, you know, maybe we can find shortcuts but I mean, I do think there there has to be you know for something that learns its own goal representations and, Learns its own control representations, if you're not gonna kind of bake it in which you know, I mean, there's maybe you can bake it in but it certainly one of the magic things that we do is you know have this open-ended vocabulary of goals where we kind of discover and develop our own internal goal room presentations through you know, experience.

And to me that open-endedness is critical like that's what allows it to kind of be flexible and adapt to new situations etc. So at least you know, if you want to capture that element of the flexibility, I do think that's what development is all about is kind of, you know, sort of putting in the the grunt work to develop those control representations kind of from the ground up.

So that they are not you know hand wired and externally driven but really something that the system itself kind of organically created and therefore. Can expand kind of you know, again be this open-ended a set of things you can do. And that's like, oh yeah since because we're sophisticated human beings and we can you know pass on to preacher generations exactly what you should be doing at different times.

Yeah instincts play a huge role but there's another role, you know, start sort of introducing them to things very gradually, you know. Yeah, yeah the shaping kind of property and all those things in it, you know, but I mean you one of the things as a parent, You know talking about development it's remarkable how little the kids will actually, you know directly absorbed from here parental wisdom, everybody's got to kind of learn it on their own and I do think that is kind of a part and parcel of the fact that these are internal goal representations and you can't just export them, you know into somebody else's brain.

They everybody's got to kind of develop their own sense of like what what they want and what they don't want and what works and for them, you know, etc. So that kind of, you know singular, Kind of you know, independent aspect of learning that kids have. I think you know, maybe not accidental.

Yeah, maybe essential yeah then it means it's really hard to just like you can't just probe you know, if you could program your kid you take a lot easier but it wouldn't really work that well because then they wouldn't end up, you know, having the flexibility they needed, right?

The first 18 months looked like you know there's this business this brain that's just not even consembled yet, you know. Yeah, there's a lot of there's a lot of really basic stuff that has to happen in that state, right? You could probably bypass some of that right but yeah, it's a question.

I mean how much how much of it is just an accident of you know, what when the baby's born and is informed too early and you know the first canal and all these other questions versus you know, is that what you do you really need that kind of early really globular kind of learning to really.

Shape the whole system in the right way. I don't know. Yeah. Any last questions here and we should let Randy go and have dinner or whatever. Okay, what are your thoughts concerning the who, what, where and why of consciousness?

I mean if we didn't have language. Would we be that much different than those feral cats? Right. I mean evidence says, you know a little bit but not much, right? And so, you know, I think this idea that we program ourselves in natural language is really compelling and so the ability to follow instructions to create instructions to think about instructions to do something complicated, you absolutely have to kind of verbalize it to remind yourself what you're doing. You know that verbal play-by-play is essential, right? Another challenge is how far can you get without trying to do that or do you try to take on a really hard additional problem. But I do think it is playing a critical role.

You have arbitrary sort of ways of thinking you have on top is that you can just borrow and bring in yeah exactly that's basically they they could teach a lot they can and they know how to hunt and do you still yeah these kinds of things. Yeah, but that could be just instinctual.

Yeah, yeah. No, I think they open ended this obviously the generativity is very much direct with the language. So and but you know in relation to consciousness per se I think you know, this is paper by Victor lame that really makes this compelling case that the the singular kind of key element of consciousness is the bidirectional authenticity that when you have a brain where bi-directional activity is flowing kind of across large chunks of the brain.

Then your conscious and when you don't have that you're not conscious and so at least it's a pretty nice neural correlate and then it makes sense in terms of like, you know this classic idea of as attractor right you've got all these different parts of the brain if you want to have all the different parts of the brain kind of focusing on a single kind of coordinated thought that's consciousness and it is fundamentally the fact that you can kind of get all the different pieces kind of working on the same problem and reflecting different aspects of this same kind of core whatever it is that you're currently thinking about. And then it moves on and that's this kind of sequencing problem but I do think that you know, this notion that we have a kind of flexible ability to kind of combine together elemental cognitive operations is part and parcel of consciousness consciousness is that kind of plenary open interchange of information from one time step to the next where each moment in time you're kind of integrating whatever you have whatever happened.

Last time what? With whatever you're currently. The experiencing etc and then you now kind of can say okay well that's that thought and now I'm gonna take some new idea and mix it in with what I was just thinking and it just keeps kind of integrating and moving along but it's not a bunch of little separate things all flying off in different directions, it's staying coordinated so that information can be shared across all these different kind of parts of your brain and that I think is key for the binding problem and so I do think that consciousness is really fundamentally about this kind of, you know, It's about the dynamics this fundamental core problem of the dynamics and a system that it that has consciousness has that ability to essentially experience dynamic kind of control because information is going and all which ways in the brain and therefore some piece of some fact over here can kind of influence some you know, updated object representation over here, etc you got that you got the stuff flying all over the place and therefore it can kind of be shared and controlled and those the two things that we've really identified as one.

Like that's what we need to make it kind of flexible so yeah I think that's all that's all part and parcel the same thing and that you think that is that is something that animals don't have at all no. I think they I think they have it and again it's like, you know, the other definition of consciousness is what it's like to actually be a specific human being in other words we only ever have, you know access to our own subjective conscious state and we evaluate whether something else is consciousness basically in proportion to how much it's exactly like our consciousness.

But take that aside, you know this essential kind of shared reverberating dynamic activity states. I think those are gonna exist in anything with a cortex and one thing that's really interesting is you know, hippocampus and cortex are they only two brain areas that have those bi-directional affections right and so if you're a lizard you don't have bi-directional connectivity and therefore by this definition, you do not have consciousness.

So so there is a kind of neural dividing line that you can you can draw according to that kind of definition and language basically is a prosthetic, you know, so we still think yes alt right on if and totally yeah just plugs in there and it gives structure and form to all those kind of juggling states in our brain right it gives them that kind of symbolic nitas to kind of nucleate around and kind of discretizes to some extent that whole kind of reverberating state and the other.

That you mentioned in your original lecture two years ago is we know how to teach people few things that are really difficult things that if they had if they didn't have to they wouldn't do yeah so hard and so frustrating, but once you look know how to do that.

All of a sudden got a tool that nothing else exactly absolutely yeah. This is this has been really great we've done yeah cool so so you guys are I guess I I'm not a hundred percent sure you're gonna try to actually make a model that does something in the space of the booster control hierarchy but not related to the programming stuff or yes, but yeah probably looking at now this is yeah, okay a motor kind of thing okay cool really have the motor cortex and what we also want sort of topographical maps and alignment and those kinds of things yep team basically to.

Do a lot of the structure that you can exploit and also you know when the kills very interested in in robotics and we think that you know embodied is it's a key aspect yeah right yeah. Yeah I used to be interested in more interested in language that I feel like working and robotics is more concrete you're dealing with more concrete circuits.

I felt like, you know, there's a step there's a higher PC you need to start with an entry point but yeah, yeah, well I think it all is sensory motor grounded so robotics is the right the right starting point. I think you know, again that developmental question that's where that's where it all grounds out and the language almost certainly builds on the semantics that you learn in the sensory motor world.

And the sensor motor grounding is a pretty big yeah for sure yeah, so it's starting there makes a lot of sense. Well, thank you for clearing up a lot pieces of the puzzle and sharing your speculation about the parts that remain murky.  We'll do it again sometime soon and keep you posted how things are going, okay?}