HARVARD VLSI GROUP

Neuromorphic Pulse Computation

Graduate Student: Jeff Miller

Faculty Advisor: Woodward Yang

Pulse computation is a novel analog/digital hybrid which uses the time domain to achieve large dynamic range, reasonable noise-insensitivity, and a new set of elementary operations with the possibility of efficiently performing many complex signal-processing tasks. Pulse streams, like digital signals, are binary-valued in voltage. Information is transmitted discretely by individual pulses. Pulse streams, however, are unclocked and pulses may arrive at arbitrary times; and pulses may be generated by analog processes within sensors, such as an integrate-and-fire driven by a continuously varying analog photocurrent.

We believe that, although both digital and analog computation and signal processing are well-established fields, there is a large middle ground of hybrid digital and analog techniques which has not yet begun to be explored. It is clear, for example, that the brain also has evolved a robust, relatively noise-immune system of computation; yet, the solution which has evolved is not a digital one, and on the contrary bears some similarities to our idea of traditional analog computation.

In particular, there has as yet been no systematic investigation of a new domain for information coding: the time domain. Our research has concentrated, first, on the development of useful techniques for encoding information in the time domain; second, on `neuromorphic' information processing with pulses -- in particular, through the construction of simple, generalizable abstractions of the behavior of neural fibre, which we can translate easily into devices on silicon. For an early example of this research, see the paper cited here.

VLSI Home Page.

Jeff Miller

miller@atlantis.harvard.edu