Two types of simple recurrent networks (Jordan, 1986; Elman, 1988) were trained and compared on the task of adding two multi-digit numbers. Results showed that: (1) A manipulation of the training environment,called Combined Subset Training (CST), was found to be necessary to learn the large set of patterns used; (2) if the networks are viewed as learning simple programming constructs such as conditional branches, while-loops and sequences, then there is a clear way to demonstrate a capacity difference between the two types of networks studied. In particular, we found that there are programs that one type of network can perform that the other cannot. Finally, an analysis of the dynamics of one of the networks is described.

We describe a technique for automatically adapt- ing to the rate of an incoming signal. We first build a model of the signal using a recurrent net- work trained to predict the input at some delay, for a "typical" rate of the signal. Then, fixing the weights of this network, we adapt the time con- stant T of the network using gradient descent, adapting the delay appropriately as well. W e have found that on simple signals, the network adapts rapidly to new inputs varying in rate from twice as fast as the original signal, d o w n to ten times as slow. So far our results are based on linear rate changes. We discuss the possibilities of application to speech.