Gluck Lab Online Mark Gluck
<p>Jul 5, 2010 9:59 PM
 
Mark Gluck
Prof of Neurosci at Rutgers, NJ.
 
Gateway to Memory ... intro to Neural Network Models of HC and Memory.
 
and undergrad text on Learning and Memory  see node.
 
Publications are great!  and all are online!
 
Also did fasc study on sleep (which assists HC based (observational) learning)  with
Bob Stickgold at Harvard/Deaconess.
 
Category Learning... which is it... a Manet or a Monet
a Claret or a Burgundy?  gradually learn to discriminate.
 
Basal Ganglia learning  deals with consequences of behavior... reinforcement learning.
Hippocampus is used to build NEW representations  (which things can you compress and
ignore  as in all states beginning with I  (Iowa, Indiana, Idaho) to a jaded New Yorker
as opposed to Thai restaurants on 9th avenue in Manhattan.
 
Aristote i  384 to 322 BS  talked abt associations as the basis for learning.
Contiguity  (closeness in space and time)  and  frequency of assoc.
this was later modeled in detail by Clark HULL in early 20th cent.
 
Rescorla and Wagner  rule  models it well.
Learning  =  beta *  (Actual Outcome - Expected Outcome)
as below.
 
There is a cue to an outcome,  and then the outcome
 
Say the cue predicts the outcome will not occur;  but the outcome DOES OCCUR...
that is a POSITIVE surprise....  dopaminergic neurons in the Substancia Nigra fire
and active neurons in the Striatum  (caudate and putamen) that strengthens the association.
 
If cue predicts outcome and the outcome does occur ... then nothing happens,
you expected the outcome, and it did happen (big deal)  nothing happens.
 
If cue predicts outcome, and the outcome does NOT OCCUR.. .  
then u r disappoited,  dopamine neurons in Substan Nigra  fire less than usual
and  less activation of Striatum in Basal Ganglia  and the link btwn the cue and the outcome
is weakened.
 
So, reinforcement learning  (Pavlovian conditioning)  depends on Basal Ganglia.
 
This also underlies  CHAINING OF LEARNING...
Mad scientist stays in his lab... he's learned that working hard gets him
publishing in science.  These pubs get him NIH NSF grants... these get him
summer money  to buy cars to get girls.
 
He tests that and shows that thats how learning works with sequential room navigation
in a vidego game maze.  go thru red door and then thru green door etc to get to the gold.
 
Mr Potato Head...  different outfits predict  whether he likes chocolate or vanilla.
if given feedback, then works by BG.
if observational then works by HC.
 
Subject is like a Baskin Robbins clerk trying to guess what customer will order.
 
Edward Tolman was an early cognitivist  1886 to 1956... he studied cognitive maps...
rats make these just for fun...  no reinforcement required.... called latent learning
cuz its done without reinforecemtn.
 
study this say in rabbits... linke tone with eye puff/blink.
 
Latent inhibition is when rabbit learns first that cue is irrelevant  eg
if cue occurs randomly or if cue and puff occur randonly  with no order.
Once u learn to ignore the tone cuz its random,  its harder to pair it with
uncondit. stimulus (shock or puff).
 
damaged HC pts do better at this task cuz they can ignore the earlier trial in which
tone was random.
 
What is function of HC?
 
WWWJD  What would Wm James Do?
1896  Prin of Psychol book...
 
James said "  is all things that were cold were also wet we wud ignore
wetness or coldness and just lump them into one catgory:
similarly for sharpness (pungent) and harp ... lump into one category.
Gluck calls that  redundancy compression ... ie we fail to make a distinction btwn those features.
On the other hand, some features we learn to v. carefully distinguish.
 
His example is  the New Yorker cover map  with all states beg. with I  (indiana etc0 merged together
by jaded New Yorker but every store  and restaurant on 9th avenue is carefully memorized.
That's what HC does...  underlies  the creation of brand new representations of the world  that
are then xfered to the cortex...  predictive differentiation...  what exact features distinguish
two categories ... the Manet and the Monet... that kind of learning requires REM sleep.
 
So,  without HC,  u have rigid, unchanging view of the world... unable to add in novelty.
 
HC  enables predicitive differentiation... he models it using an auto encoder that subserves
novelty detection and new representations.</p>