Hinton DiCarlo Lewicki Mar 2 2010 Visual Object Invariance and Deep Belief Networks
<p>Mar 2, 2010 11:05 PM  (Jim Peterson also there  and Jim Strahorn)
 
Jay McC gave introduction
 
to quote him  VISION is an "emergent consequence of widespread neural activity"
he shows slide exactly like my slide with a microelectrode going from surface then down
1000 microns  showing  perhaps 1 hundred neural traces  (that's the RAW MATERIAL).
 
JIM DiCarlo ...  Assoc Prof  (just tenured)  superstar leads off ... now at MIT  (McGovern Insit)
MD, PhD at Hopkins.
fantastic presentation!  super sharp  (unfort no videos of him)
see slides in Brain under Dicarlo  (and Kreiman) ...  many of his slides.
 
Untangling Object Recognition.
He shows various objects (planes, people, cars, etc) at 6 frames per second ...
we can easily recognize each one.  critical to survival.
he does experiments in rhesus monkeys.  Ventral visual stream  (what pathway).
 
LAG  from  RGC retinal gang cell to IT is about 100 msecs.
 
Imagine N neurons in IT  N1  thru Nn  (like a hilbert space like Dileep George formulation)
where each object traces our a manifold in n space as it or our eyes/head move  then we
can divide those manifolds  (maybe) by a hyperplane in n dimensions.
 
He reocrds from V4 and from IT in monkeys... wait 100 msec after stimulus is turned on
then count spikes btwn  100 msec and 200 msec  (in a 100 msec window).
Try to decode identity of object.  IT is better than V4  (less tangled)
 
TEMPORAL CONTIGUITY is hugely important  (objects trace out a manifold (describe an
affine transformation)) with head/eye/object motion  as they tumble thru space.
change in size, position, and pose.  Also we see changes tghru  one  saccade.
 
UNsupervised temporal learning  (from smooth motion)
He is trying to constrain the space of brain algorithms that do this extraction of 
neural "tolerance"  to variation in size/positon/pose.
 
Calculates a Turing Ratio  =  machine performance /  humna perfomrance.
occasional  (for some brief presentaitons and objects, machine vision does better than human).
see PLOS 2009  Pinto, Doukin, et al.
 
Humans do much better or these tasks when there is large variation.
 
Q from Anthony Wagner:  how about role of medial temp lobe  (memory)
lesions of  perirhinal or entorhinal cortex  get rid of IT effect.
 
A:  just front=edge  (feed forward fast initl wave is good)  ie  it does NOT need to rattle
around in MTL (or to use memories/ Bayes thy) to get good performance.
 
Michale LEWICKI (very calm;  v smart...  v complex idea set... no videos on net)
He is NOW at CASE Westerm ... assoc prof.   he trained with Jay McC  and with Terry Sejnowski.
 
One underatndable part of his talk...  an analogy in the spirit of this quote from
Horace Barlow:  A WING would be most mystifying if we didn't know that birds flew.
ie  essential to understand function... what is the BRAIN's TASK.
He uses as an example the EVOLUTION of EYES  (see book by Michael Land and
Russell Fernald)
So,  the key is to ask the question <b>what is the optimal design of an eye giving the requirements</b>
<b>of seeing</b>?
Now Michael asks ...  similarly  what is the task that the brain is doing in object recognition that
it is trying to optimize?  (eg getting rid of noisy sensors etc)  optimal information capture.
 
A lot of his work is on understanding natural (very cluttered) scenes... trees, vegetation, etc.
encoding statistical variation.   Covariance matrix.  his models learn a compact set of
dictionary elements for images encountered in natural scenes.
 
 
 
 
 
Geoff Hinton ...  his talk resembled only was far more complex than his Goggle Tech Talk...
just watch that.
 
Great introduc anecdote:  tells when he ws in college, there were fine-grained distinctions
btwn leftist political parties  (Trotskyite, people's  manifesto party;  vs  Trotskyite people's
rights party  or whtaever)  and they would have violent disagreement. This is similar to
 his postion with respect to Michael Lewicki  (ie  99.9 %  agreement but .1% violent disagreement).
 
he discusses back prop,  RBMs  (restricted boltzmann machines),
visible units are the pixels (eg in the retina);  hidden units are activated by them.
Minimum  energy means reducing discrepancy btwn observed - expected.
Maximum likelihood for RBM.  Rather than going back and forth many times,
he only goes up and down TWICE  (just 2 iterations).
 
STUFF (objects in the world) ---->  Image  ----->  LABEL  (this is the conventional  AI/ cog sci view)
 
but rather he thinks it's like this
 
STUFF ----->  Image   
   |
   --------> LABEL     (ie they are concurrent)
 
Soldiers on a field analogy:  to get them to form a rectangular grid... can give each a precise
gps coordinate  (NO)  or just send them out anywhere and have them square up using their
R arm to postion themselves evely... ie use local relationships.
 
Q  about  FRUIT face.</p>