Self-clustering recurrent networks

Recurrent neural networks have recently been shown to have the ability to learn finite state automata (FSA’s) from examples. In this paper it is shown, based on empirical analyses, that second-order networks which are trained to learn FSA’s tend to form discrete clusters as the state representation in the hidden unit activation space. This observation is used to define ‘self-clustering’ networks which automatically extract discrete state machines from the learned network. However, the problem of instability on long test strings is a factor in the generalization performance of recurrent networks - in essence, because of the analog nature of the state representation, the network gradually “forgets” where the individual state regions are. To address this problem a new network structure is introduced whereby the network uses quantization in the feedback path to force the learning of discrete states. Experimental results show that the new method learns FSA’s just as well as existing methods in the literature but with the significant advantage of being stable on test strings of arbitrary length