Data-driven methods for discovering the structure of neural and cognitive representations (Part 1...) Kenneth Whang Bryn Mawr College
Three main ideas hypothesis testing and model building population coding, distributed representations dimension reduction and individual differences modeling Outline Prologue: statement of the problem An approach: multidimensional scaling Compare/critique: what needs to be done
behavior/experience “implementation” what’s going on in here?
behavior/experience “implementation” Experimental Psychology develop model, build instrument, measure Example: attention Problems: limited by model’s ontology doesn’t scale well (how to capture complex behavior?) hypothesis formation itself can be a form of observer bias
behavior/experience “implementation” Physiology classify cells, characterize their properties Example: “oriented bars” Problems: unclear or speculative connection to function continuity of types– how many types are there? ad hoc
behavior/experience “implementation” Computer Science? concentrate on data and what they might represent; let data be primary driver of modeling and interpretation
About representations Computer Science representations and computational strategies closely linked what you can compute is related to what you can express Neurophysiology some examples indicative of function (orderly maps, tuning, graded responses) others much more muddled
Distributed or localized? “dense”“sparse”“grandmother cell” representational capacity highmediumlow learning/adaptationslowfast generalizationgood none fault tolerancehigh none energy usehighlow Examples: motor cortex (direction of reach); visual cortex (natural scenes) Except for grandmother cell, interested in relationships