We are studying neural/analog inspired computation. Neural/analog inspired
models are interesting for several reasons. First, these models are useful
because they lead toward fast, ultra low-power analog VLSI hardware.
Second, we study these models in order to better understand the
information processing tasks necessary for artificial and biological
computational systems. Finally (and ironically), these algorithms can lead
to efficient DSP-based algorithms that have a much wider applicability
than if they solely existed in analog hardware.
We are studying biologically-inspired analog signal processing techniques
in several application areas. Many real-world applications become feasible
because of the high-speed, compact-size, low-cost, and low-power
consumption of the analog chips under design. Special-purpose analog VLSI
hardware is currently under development for 3-D sound localization and
processing, analog speech synthesis, visual motion detection, wavelet
image enhancement, adaptive filters, and chaotic oscillators. Ultimately
this research will result in products for speech recognition
preprocessors, hearing aids, teleconferencing, real-time image processing,
and intelligent automobiles.
We are also building analog VLSI circuit models in order to gain a better
understanding of neurobiological computation. Since analog circuits
operate under many of the same power and communication restrictions
imposed upon their biological counterparts, it is hoped that the silicon
models will provide insights into these biological information processing
systems. We believe that we cannot easily model detailed biological
systems with analog VLSI since the two media differ significantly at their
lowest levels, however, these silicon systems can provide insights in
neurophysiologic organization--especially for higher level-brain functions
where models cannot realistically model every biological detail. A
indirect benefit of using analog VLSI to study biological computation, is
the development of devices for the handicapped such as improved hearing
aids for the deaf, artificial vision systems for the blind, and
collision-warning sensors for wheel-chairs.
Fully analog implementations are poorly matched for tasks requiring long-distance communication, long-term storage, high accuracy, flexibility, or general purpose computation. Thus, realistic systems are typically hybrid systems that utilize digital circuitry to deal with the above mentioned problems and analog hardware, at the very least, is required to interface to the mostly analog world. We believe that there is a larger role for analog systems than to serve as simple A/D interfaces. Two particularly promising areas we are investigating are front end sensor processing and dynamical systems modeling.