Okay, this is just ripped off of the hompage that my advisor semi-insisted that I create. It's interesting stuff, and it perhaps deserves more of a writeup than this, but for now I guess this is okay. Hope you enjoy.
So, my research interests, in the broad sense are related to
neural modeling. More specifically, I work in the medium of
Neuromorphic Analog VLSI. It's a big fancy term,
I know, but what it amounts to is the modeling of neural systems in
silicon hardware. We used a
mixed analog-digital system, much like real
neurons. The basic idea is to design circuits which function in a manner as analagous as possible (or neccessary, depending on your point of view) to real neurons. So, the big question, of course, is "Why analog VLSI?" Why not
computer simulations of neurons, why not use
DSP to do these computations? (we get questioned by the biologists and the engineers). So, I'll try my best to explain what I think about neuromorphic engineering, and why I'm interested in it.
Why not computer models?
So, one thing is that I think modeling in silicon gets at a lot of things that are often left out of computer simulations, for better or worse (or perhaps better and worse). Silicon systems are real
hardware, just like the
brain. This means that they have to deal with constraints based on
noise and
available space. This can be both good and bad. It does limit what we can do in one sense, but it limits us in a biologically realistic way. Our circuits have to be designed to be
robust in the face of these limitations, and I think that by understanding that, we perhaps gain some insight into
biology. That's all well and good, but I think the real advantage of neuromorphic engineering is that it is
fundamentally more real than a computer simulation. Circuits run in real time, and if they're controlling a
robot, they interact with the real world. First,
real time. Because our circuits run in real time we can do things like adjust
parameters as the system runs, which is not typical of
computer simulations, and gives an added
insight into the workings of the system. Also, we can model several relatively large systems, and have them communicate with one another, in real time. To simulate a
sensory system, internal computation, and a resultant
motor output that interacts with the world, in the presence of noise, all on a computer would be a staggering task. The program would take forever to run, and you could only adjust it properly between runs. And then how accurate would this real world be? Likely not very accurate. I think that neuromorphic engineering gives us a wonderful tool in that it allows us to put multiple systems together into a
coherent unit. Which is not to say that computer modeling isn't a very powerful tool. I'm not advocating this as a replacement for such things, just as a new and different approach, which has a different set of tradeoffs than other methods.
Why not just use digital processing like any sane engineer?
Well, one, becaue we're interested in studying biology, and biology isn't
digital. There's a valid question here, though. From an engineering standpoint, what sort of advantages do you get by running
transistors in
subthreshold as
analog devices? Well, there are some standard answers for this. First of all, if your processing task doesn't require more than four or five
bits of precision, analog processing will save you a ton of space over digital. Adding two values in analog (if they're
currents) only requires a wire. One wire, and it can be as small as you like. In digital, adding two values requires significantly more hardware. Also, if biology can do it, then it can (debatably) be done with five bits of precision or less. There's also a power advantage to analog processing. We work with currents that range from pico
amps to nanoamps, which is quite small in comparison to digital. This is all good stuff, but I honestly think the biggest advantage of
neuromorphic VLSI is that it draws on biological systems. In the fifties, everyone was promised that they would someday have an
automatic vacumcleaning-floormopping-dishwashing-dogfeeding-tvrepairman-mailfetching amazing wonderful gizmobot. Here it is fifty years later, and I don't see any robots running around. Why is that? It's because
making sense of the real world and navigating in it turned out to be a hell of a lot harder than we expected. If we want to build systems that can
make sense of the world, it seems to me that a good place to look for inspiration is where such systems already exist, and in fact have existed and been refined for millions of years, in biology.
Back to
How your brain works.