Recurrent competitive Hebbian learning

Abstract

Competitive Hebbian learning is extended to networks with trainable lateral connections, in addition to the trainable feedforward connections considered previously by the author (1991,1992). These recurrent systems are able to learn to respond to ordering in time of the input vectors. The theoretical framework for the extension of competitive Hebbian learning to recurrent systems is presented. This is followed by three demonstrations of recurrent competitive Hebbian learning, two unsupervised and one quasi-supervised. The first example is a system of two nodes which are trained on a set of Gaussian spots presented in a 10-by-10 input array. The second example shows the system learning to respond to vertical lines in a small, 4-by-4 input array. The final example is of a system trained to produce useful responses to a tiny Boolean algebra test, where the Boolean variables are the successive values of the single input variable. >