ABSTRACT
This paper describes a connectionist architecture designed to learn temporal relationships between input events in a relatively small number of training trials, and presents results of simulations applying the network to an A→B+vs. B- serial conditioning task. The proposed architecture employs a cascade of hidden layers with memory which feed into a single output layer. The hidden layers were trained using unsupervised Hebbian learning while the output layer was trained using supervised gradient descent. The network was given the task of learning that the sequence A→B was a predictor of C, while B alone predicted C’s absence. After as few as 20 training trials, 4 of 7 randomly initialized networks were making the required distinction. This represents a significant improvement in learning rate over networks trained on similar tasks using a recurrent backpropagation algorithm, and more closely models results obtained in animal conditioning experiments.
