On 10/21/07, Edward W. Porter <[EMAIL PROTECTED]> wrote: > > Vladimir, > > Yes, the deleted point "FIVE" mentioned that I had assumed (perhaps > incorrectly) that Valiant was looking for enough interconnect to do > traditinal Hebbian learning, which as normally defined would require > synapses from either A and/or B cell assemblies firing directly on each > other. But I hypothesized that with a loosening of the hebbian timing > requirements, or through the hebbian timing requirements being more driven > by a sychronisity pattern, rather than precise phase matching, it would be > possible for indirect connections between A and B through the associated > nodes of each to produce learned associations. >
Edward, There is no problem with using indirect association, but it's unclear how intermediate neurons can learn to consistently fire when they observe A and not in other cases. When neuron finds regularities, it's easier (as all features comprising regularities are transmitted towards neuron), but in our case it has no information about destination. Also probability of finding such intermediate neurons is the same as probability of finding feature-detecting neurons even under phase matching requirement. I don't see an issue in phase matching requirement itself, other than a constant factor for various estimations. Phase matching allows one to unify classification of scenes and temporal schemes, and also provides a way to compute is a continuous manner, without 'tick-by-tick' global synchronization, which simplifies the model. Problem with association-making is in number of incoming synapses and it's independent on use of timing. If concept consists of M neurons, it still has only M*F incoming synapses, which would be enough to detect activation of any other concept only if brain had on the order of (M*F)^2 synapses, S=(M*F)^2, so that M=(S^(1/2))/F=(3*10^7)/10^4=3*10^3=30000 neurons (!). And this is to make association from any one of A's neurons to any one of B's neurons, which requires that that single random neuron in B is enough to reawaken all 30000 of B's neurons, which is ridiculous. -- Vladimir Nesov mailto:[EMAIL PROTECTED] ----- This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244&id_secret=56117834-bfb748
