Recurrent networks for learning stochastic sequences

Some experiments exploring the ability of networks to learn the underlying statistics of artificially generated temporal data are described. In one experiment, data generated by two simple Markov models were fed into a multilayer perceptron. The desired output was an indication of whether a transition out of one of the models had been made. The network produced a close approximation to the probability that a transition had just been made. In another experiment, hidden Markov models were used to generate the data. This made the determination of whether a transition had occurred much more difficult, and the network produced a much poorer approximation to the correct probability