First-Order Recurrent Neural Networks and Deterministic Finite State Automata

We examine the correspondence between first-order recurrent neural networks and deterministic finite state automata. We begin with the problem of inducing deterministic finite state automata from finite training sets, that include both positive and negative examples, an NP-hard problem (Angluin and Smith 1983). We use a neural network architecture with two recurrent layers, which we argue can approximate any discrete-time, time-invariant dynamic system, with computation of the full gradient during learning. The networks are trained to classify strings as belonging or not belonging to the grammar. The training sets used contain only short strings, and the sets are constructed in a way that does not require a priori knowledge of the grammar. After training, the networks are tested using various test sets with strings of length up to 1000, and are often able to correctly classify all the test strings. These results are comparable to those obtained with second-order networks (Giles
et al
. 1992; Watrous and Kuhn 1992a; Zeng
et al
. 1993). We observe that the networks emulate finite state automata, confirming the results of other authors, and we use a vector quantization algorithm to extract deterministic finite state automata after training and during testing of the networks, obtaining a table listing the start state, accept states, reject states, all transitions from the states, as well as some useful statistics. We examine the correspondence between finite state automata and neural networks in detail, showing two major stages in the learning process. To this end, we use a graphics module, which graphically depicts the states of the network during the learning and testing phases. We examine the networks' performance when tested on strings much longer than those in the training set, noting a measure based on clustering that is correlated to the stability of the networks. Finally, we observe that with sufficiently long training times, neural networks can become true finite state automata, due to the attractor structure of their dynamics.