Lecture � computational neuroscience, attention + STM

Greg Detre

6/2/02

Prof. Rolls

 

Fuster � monkey STM, neurons keep firing during delay intervals while in memory (Goldman + Rakic) remembering where to move eyes (dorsolateral prefrontal???)

Rolls proposed separate AAs for regions of eye movements to remember � regional or directional???

 

prefrontal STM model

IT AA offloads into PF AA when it needs to compare memory against new samples � all setup with associative learning and probabilistic diluted connectivity

replicates Miller�s neurophysiology data

 

covert visual search � looking for an object in a complex scene without moving your eyes � spotlight of attention

back projections down the dorsal stream biases bits of V1 + V2 which project up the ventral stream � attention

OR you can do it the other way round, i.e. spatial search with object cue

this works without requiring that IT be topologically-mapped

symmetrical forward/back projections

 

argues that there�s nothing serial about �serial� processing � it�s just that the constraints are more difficult to meet in the parallel dynamical system, linearly increases with the number of distractors

spiking nets give you actual timescale predictions

 

Questions/misc

ch 10 of NN + BF � architecture of cerebral cortex

how do the AAs form special new attractors immediately???

mean field theory

how do you get position-invariant from V1, and preserve that information (symmetry)

biased competition