[Martin Taylor 970220 14:45]
Peter Cariani (970220.1130)]
Feedback can be quite fast. There are neural loops in the auditory
system that
provide efferent control over the cochlea within tens of milliseconds.
There
is also the example of the ability to repeat speech that one has heard
with
(if I remember correctly) only a few hundred millisecond delays.
Working (like you) from memory, there's a training shift in shadowing.
At first the delay is a few hundred msec, but I seem to remember that
it can drop below 100 msec when people are well trained. Anyone got the
real data here?
There is the example ... (These kinds of fast mechanisms arguably take
advantage of the precision of spike times to process
information with only a handful of precisely timed spikes.)
Did you notice the article by Markram et al. in the Jan 10 issue of "Science"?
It's about spike timing in Hebbian learning. If the postsynaptic action
potential (AP; read "spike") follows the synaptic potential pulse (EPSP) by
10 msec, the synaptic strength increases, but if the AP preceded the EPSP
by 10 msec the synaptic strength decreses by about the same amount. At 100
msec differential time, there's no change in synaptic strength either way.
I'm particularly interested in this result, because in a 1973 paper(*) I
argued that such joint Hebbian and "anti-Hebbian" learning, based on this
kind of differential timing effect, would cause a single layer of a
neural net with initially random cross-connections to change its structure
until its outputs came to represent a principal components transformation
of the inputs--an informationally optimum structure. In plainer language,
it would cause the net to produce outputs corresponding to those (statistically
orthogonal) features most strongly represented in the input, and the output
would be stronger (i.e. from most neurons) for the more prominent features
than for the weaker ones.
More recently, I've argued on CSGnet that the same structure (with stronger
cross-link weights) would give rise to perceptions of associations and
categories. It's very gratifying to see that there is, after nearly a
quarter century, neurophysiological evidence to suggest that my old
speculation might have some merit, after all.
Martin
(*) South African J. Psychology, 1973, 3, 23-45.