| |
Abstract:
Most computational engineering based loosely on biology,
whether algorithmic or hardware, uses continuous variables and
sigmoid units. Most neurons communicate with action potentials, and
the engineering view is equivalent to using a rate-code for the
representation of information and computing. There are an
increasing number of examples in biology where this may not be the
case. Information can be represented using the timing of action
potentials, and computed with in this representation. The "analog
match" problem of olfaction is one of the simplest problems which
can be efficiently solved using action potential timing and an
underlying rhythm. By using units which adapt, and using the
adaptation to effect a fundamental change of representation of a
problem, the problem of recognizing words in connected speech with
uniform time-warp within a word can be mapped into the olfaction
problem. The architecture and preliminary results of such a system
(with Sam T. Roweis) will be described. Using the fast events of
biology in conjunction with an underlying rhythm gets away from the
limits of an event-driven view of computation. When the intrinsic
hardware is much faster than the time scale of change of inputs, it
can greatly increase the computation which can be done per unit
time on a given quantity of hardware.
|
