Eric Kauderer-Abrams
Theory: Spike Coding

Personal Background

I grew up near Boston, Massachusetts and in Englewood, New Jersey. I received my BA in Mathematics from the University of Pennsylvania. At UPenn, I focused primarily on math and physics. I'm currently a PhD candidate in Electrical Engineering at Stanford.

Eric Kauderer-Abrams.

For as long as I can remember, I've loved science. My childhood dreams were converted into a more serious commitment by several experiences, one of which was a high school physics class. The teacher inspired and shocked us (literally) with the power and mystery of physics. I wanted to learn as much as possible, as quickly as possible, but I found myself limited: every physics book was filled with the strange glyphs of higher math. I couldn't decipher the divs, grads, curls and all that, but their importance was evident. So I devoted the next several years to learning as much math as possible, to ensure that I'd never experience this limitation again.

While I enjoyed getting lost in proofs, by the end of college I was eager to work on more applied problems. I developed an interest in neuroscience with its seemingly infinite supply of mysteries, details, and relevance to human life. Having a solid foundation in math and physics, my primary goal in graduate school was to develop more applied, engineering skills, and if nothing else, to get my bearings in the vast world of neuroscience.

When I found the Brains in Silicon Lab led by Kwabena Boahen, I knew I had to join. The lab designs and fabricates novel computers that emulate the brain, from the transistors up. The lab's core philosophy is that building systems based on the brain's design principles requires a mix of different disciplines spanning neuroscience, circuit design, math, physics, and more. The lab's composition reflects this philosophy; analog circuit designers sit next to computer scientists, physicists and neuroscientists. The discussion in any lab meeting covers a remarkable range of topics (and only a modest number of communication breakdowns). Most importantly, working with Kwabena gave me the opportunity to learn how to solve problems in their entirety.

Research Goals

The brain processes an enormous amount of information on a very tight power budget. How does it do this? My goal is to develop computational paradigms that model how the brain represents information. Of all the questions to ask about how the brain works, I decided to start with first things first. When you see a scene, you perceive not the intensity and wavelength of the light hitting your eye, but the spikes generated by the retina's output neurons. What do these spikes mean? How do they encode information? Why is the code spike-based? My research addresses these questions. In addition, the computational paradigms that I develop are deeply connected to practical, hardware considerations. I've found the back-and-forth between theory and hardware work to be very fruitful.

Project Status

For the first few years of my PhD, I worked on and solved the problem of how to make neuromorphic chips robust and useable. This entailed two parts: calibration and temperature-compensation.

Calibrating the chips means allowing them to be programmed in an intuitive fashion by someone who knows nothing about the underlying circuits. For example, the lab's Neurogrid board is a tool for running biophysical simulations at a very large scale. In such simulations, it is necessary to program the simulation parameters such as the neurons' time constants, the synapse's reversal potentials, etc. However, we communicate with the chip by setting "raw" currents specified in arbitrary Digital-to-Analog converter (DAC) units. To make the chip usable, it's necessary to establish a mapping that converts the desired higher-level simulation parameters to the correct low-level DAC units.

Calibration is further complicated by the fact that even within a single chip, no two circuits are exactly like. The cause of this heterogeneity is "mismatch" between transistors, which is more significant in the subthreshold regime that our circuits operate in. Calibrating our chips entailed characterizing this mismatch and designing schemes to compensate for it so the circuits can be programmed reliably.

Temperature compensating the chips means making the operation of neuromorphic chips robust to temperature variation. I developed a novel approach to achieving temperature-robustness in neuromorphic systems. Existing approaches entailed designing elaborate temperature-compensating bias-circuits, or operating the chip within an energetically expensive temperature-controlled chamber. My solution entailed changing only how the chip is programmed.

Iím currently working on a spiking-neural-network algorithm that encodes analog, time-varying signals into spike trains with state-of-the-art efficiency. I'm currently exploring connections to information theory to compare the performance of spiking neural networks to traditional encoding and decoding techniques.


ID Article Full Text
E Kauderer-Abrams and K Boahen, Calibrating Silicon-Synapse Dynamics using Time-Encoding and Decoding Machines, IEEE International Symposium on Circuits and Systems (ISCAS), Baltimore MD, 2017.

Full Text
E Kauderer-Abrams, A Gilbert, A Voelker, B Benjamin, and T C Stewart, and K Boahen A Population-Level Approach to Temperature Robustness in Neuromorphic Systems, IEEE International Symposium on Circuits and Systems (ISCAS), Baltimore MD, 2017.
Full Text
Kauderer-Abrams, E. Quantifying translation-invariance in convolutional neural networks. arXiv preprint arXiv:1801.01450. 2017.
Full Text