Learning capabilities of recurrent neural networks

The author relates the power of recurrent neural networks to those of other conventional models of computation like Turing machines and finite automata, and proves results about their learning capabilities. Specifically, it is shown that (a) probabilistic recurrent networks and probabilistic Turing machine models are equivalent; (b) probabilistic recurrent networks with bounded error probabilities are not more powerful than deterministic finite automata: (c) deterministic recurrent networks have the capability of learning P-complete language problems; and (d) restricting the weight-threshold relationship in deterministic recurrent networks may allow the network to learn only weaker classes of languages