Learning Sequential Structure with the Real-Time Recurrent Learning Algorithm

Recurrent connections in neural networks potentially allow information about events occurring in the past to be preserved and used in current computations. How effectively this potential is realized depends on the power of the learning algorithm used. As an example of a task requiring recurrency, Servan-Schreiber, Cleeremans, and McClelland1 have applied a simple recurrent learning algorithm to the task of recognizing finite-state grammars of increasing difficulty. These nets showed considerable power and were able to learn fairly complex grammars by emulating the state machines that produced them. However, there was a limit to the difficulty of the grammars that could be learned. We have applied a more powerful recurrent learning procedure, called real-time recurrent learning2,6 (RTRL), to some of the same problems studied by Servan-Schreiber, Cleeremans, and McClelland. The RTRL algorithm solved more difficult forms of the task than the simple recurrent networks. The internal representations developed by RTRL networks revealed that they learn a rich set of internal states that represent more about the past than is required by the underlying grammar. The dynamics of the networks are determined by the state structure and are not chaotic.