Automated Abstraction of Dynamic Neural Systems for Natural Language Processing

Abstract

This paper presents a variant of the crystallizing substochastic sequential machine extractor (CrySSMEx), an algorithm capable of extracting finite state descriptions of dynamic systems, such as recurrent neural networks, without any regard to their topology or weights. The algorithm is applied to a network trained on a language prediction task. The extracted state machines provide a detailed view of the operations of the RNN by abstracting and discretizing its functional behaviour. Here we extend previous work and extract state machines in Moore, rather than in Mealy, format. This subtle difference opens up the rule extractor to more domains, including sensorimotor modelling of autonomous robotic systems. Experiments are also conducted on far more input symbols, providing a greater insight into the behaviour of the algorithm.