Ask A Scientist© Biology Archive

Author:      jcolombe
Text:        I was wondering if anyone at Argonne was studying the brain or 
learning processes in which chaos (or tortuosity of state-space) plays a role 
in the "trial" part of trial-and-error learning.  I have thought that maybe 
the state-space of activity in the brain might tend to be more tortuous (and 
thus variable) in regions of state-space that are unfamiliar, or where 
experience has not taken the state of the system yet.  If experience is the 
tracking of system state through this space, and if Hebbian associative 
learning "convergences" experience paths through the space (making these paths 
themselves into attractors), then learning might "dechaoticize" the areas 
around these experienced paths, making them repeatable and reliable in 
activity.  So "trial" would only be necessary (and possible) in unfamiliar 
circumstances, and learning would "iron out" the trial in favor of giving the 
system the ability to converge on (or be attracted to) familiar states in 
familiar situations.  That might account for classical or perceptual 
conditioning, but what about operant or behavioral conditioning?  I think that 
if a central reward signal was available to the system, that the amount of 
association taking place should be proportional to reward.  Then the system 
will learn more and retain more function from favorable situations than 
unfavorable ones.

Response #:  1 of 1
Author:      psych
Text:        I am not at Argonne but.  .  .  It is an interesting idea (if I 
understand it correctly).  In an abstract way I suppose that is how neural 
networks are supposed to work - they go from a completely random state with 
random (Chaotic) output upon specified inputs to more specific output once 
learning has taken place.  I think I have seen this described as energy minima 
in the neural net and because of the present sophistication of neural nets 
that is most like classical conditioning.  In behavioral conditioning the 
brain learns by reward or punishment to a previous response.  Some people are 
looking at models that incorporate state dependent responses (that are 
contingent on a previous stimuli as well as the present one).  I have not seen 
this used to let the net program itself (enhance learning).  My guess is that 
the reason for this is in how much more computational power is required to 
adjust neural nets in their current form.  As some of my colleagues have noted 
this is one of the weaknesses of neural networks in their current form.  The 
brain does not appear to need to take up excessive amounts of processing 
capacity in order to adjust its net to new learned information.