Second-order recurrent neural networks can learn regular grammars from noisy strings

Abstract

Recent work has shown that second-order recurrent neural networks (2ORNNs) may be used to infer deterministic finite automata (DFA) when trained with positive and negative string examples. This paper shows that 2ORNN can also learn DFA from samples consisting of pairs (W,μw) where W is a noisy string of input vectors describing the degree of resem blance of every input to the symbols in the alphabet, and μw is the degree of acceptance of the noisy string, computed with a DFA whose behavior has been extended to deal with noisy strings.