95-221 Dimitri Petritis

THERMODYNAMIC FORMALISM OF NEURAL COMPUTING (267K, 51 pages, uuencoded postscript) May 15, 95

Abstract , Paper (src), View paper (auto. generated ps), Index of related papers

Abstract. Neural networks are systems of interconnected processors mimicking some of the brain functions. After a rapid overview of neural computing, the thermodynamic formalism of the learning procedure is introduced. Besides its use in introducing efficient stochastic learning algorithms, it gives an insight in terms of information theory. Main emphasis is given in the information restitution process; stochastic evolution is used as the starting point for introducing statistical mechanics of associative memory. Instead of formulating problems in their most general setting, it is preferred stating precise results on specific models. In this report are mainly presented those features that are relevant when the neural net becomes very large. A survey of the most recent results is given and the main open problems are pointed out.

Files: 95-221.src( desc , 95-221.uu )