Abbott, Larry Columbia
<p>Apr 8, 2008 9:01 PM
 
Prof. Larry Abbott
Prof. of Theoretical Neuroscience
Columbia Univ.
 
Gave  Robt  Hofstadter Mem. Lectures  (Hofst. won Nobel in 1961 for discovering that proton and neutrons had internal structure)
 
My interview today:
 
Q Bob : Memory encoded by a few synapses OR  huge populations of neurons:
A Larry:  prob. large populations
 
Q:  Bob  What do you think of Jeff Hawkins works
A  Good;  also comes to it from another field;  consensus model;  accurate
 
Q:  IBM  Blue Brain    OR  C Cortex  Project in SF
A:  They are doing highly detailed modeling of anatomy and physiol.
My models are more limited.
 
Q:  How is info  transferred from  Hippocampus back to cortex for permanent storage
A:  My group is working on that;  have some ideas;  its a hard problem  (shifting a pattern)
(similar to Bill's  question of function of sleep; chaotic activity triggering patterns)
 
Q:  Do you hang out with Eric Kandel?
A:  Of  course;  he asks  do I know E.K.?  My A:  Only thru Stfd lectures.
(I mention my bkgrnd at MIT with Lettvin,  Minsky,  etc.)
 
*
 
Lecture 2:  HOw do you get accurate coding in presence of spontand, chaotic noise?
 
Who's Afraidof Chaotic Networks?  Models of Sensory and Motor Proc. in Face of Sponatn Neuronal Activity
 
Large, strongly coupled neural networks produce chaotic sponatna activity.  You'd think they'd be unreliable.  But NO.
Inputs can induce PHASE TRANSITIONS, leading to reponses uncontaminated by chaotic noise.   Likewise approp trained feedback units can controlthe chaos,  resulting in a wide variety of repeatable output patterns.
 
He shows  diagram of CAT looking at a target;  record from many units in  visual cortex.
It looks like you see a pattern if you look at the activity map in the cortex.
HOWEVER ...  if you TURN OUT the lights ...  the pattern is still there.
 
CONCLUSION:  Brain neurons fire on their own;  lots of spontan. activity without any sensory input.
 
My paper notes are in my file for  NEURAL NETS  (in neuro section).
 
Xi  =  total current in ith neuron;;   ri  is the firing rate of that neuron
 
Change in curruent = current +  SUM (Jij * rj    -   bi  (bias current)
 
Jij  is  strength of synapse connecting  jth neuron to neuron i.
 
 
typical time constant =  10 msec;
 
behavior of networks crucially depends on VARIANCE of  firing rates in network;
 
if  g =  normalized variance  is < 1 then network is inactive;  if  g is BIG;  > 1 then  there is spontan chaos.
 
Jij  matrix  is unusual in that columns are either ALL pos or  ALL neg  cuz  A GIVEN NEURON IS EITHER EXCIT. OR INHIB
BUT NEVER BOTH  (ie  even tho its inputs may be both,  its outputs are either all excit or all inhib..
 
Normalized  mean input to  0  ( so each row sums to 0).
 
Intuitive model is this:  networks are noisy;  if u supply big input signal then get  big  S/R ratio to over-ride noise.
That is wrong;  instead:  SIGNAL SUPPRESSES THE NOISE cuz the entire system undergoes a PHASE TRANSITION and
chaos shuts down.
 
If you transform  original  del (current) eqn;  it looks like  NEWTON;  law of motion;  ma = F =  del V/del x
 
Eqn has W shaped well;  stable point is in middle of W.
 
Get noise suppression around 5 to 9 hertz.
 
NOISE SUPPRESSION is prob. most sensative  index of coding of input rather than the system's actual direct coding response
to the input pattern.
 
That's sensory;  how abt  MOTOR;  can get noise suppression by feeding back from  output cell  (cortical pyramidal cell)
back to cells that give it its inputs.
 
 
 
 
 </p>