Towards a theory of neural computation for computing symbolic functions

SpeakerPaul Smolensky
AffiliationJohns Hopkins University
Date and Time April 4, 2013, 6 p.m. - 07:00 p.m.
LocationMcGovern Seminar Room: 3189
HostTomaso Poggio




I will discuss a framework for a theory of neural computation from several perspectives. From the viewpoint of automata theory, the machines in question are formal neural networks (connectionist nets) supporting distributed representations of symbol structures and simultaneously performing stochastic global optimization as well as quantization to discrete symbolic states. From the perspective of recursive function theory, recursive equations for symbol-mapping functions become recursive equations for the (weight) matrices of linear mappings (networks). Finally, a language-theoretic approach uses optimization over symbol structures to define formal languages, and to support optimization-based approaches to natural languages.

There are no comments yet

Post a new comment...