Cornell Cognitive Studies Symposium

Statistical Learning across Cognition

Problems Arising in Unsupervised Learning of Structure

Shimon Edelman
Cornell University
se37@cornell.edu

 

It is logically impossible to form a principled structural description of a visual scene without prior knowledge of related scenes. Adapting an observation made by R. A. Fisher, such knowledge must, in the first instance, be statistical. Indeed, several recent studies showed that subjects are capable of unsupervised acquisition of statistical regularities (e.g., conditional probabilities of constituents) that can support structural interpretation of novel scenes composed of a few simple objects. Theoretical understanding of unsupervised statistical learning is, however, hindered by the following paradox: statistics can only be computed over a set of candidate primitive descriptors if the candidates are known to the system in advance (a related issue was raised by Horace Barlow in his work on efficient learning). I shall discuss possible ways of circumventing this problem, motivated by neurobiological and computational considerations.

Joint work with Benjamin Hiles and Nathan Intrator.

 

Back to main page