On the Relatively Simple Statistical Mechanics of Neural-Network Acceptors

摘要

We define a neural network for sequence recognition, or simulation of abstract automata. This network can simulate any finite-state acceptor. We also define an energy function and show that the network minimizes its energy. Then we show how to derive a free-energy function, which has only boundary minima, and an expression of the average overlap for the network. These derivations are simpler than the analysis of the attractor network, from which the techniques for analysis have been adapted. This simplicity is intimatley connected to the network's design as a neural-network acceptor.