Learning mechanism for the identification of hidden structures in signal sets

A neural network is proposed which has the ability to identify signals by recognizing the presence or absence of various signal components, rather than by recognizing input signals as whole units. The presence of rich component structures in naturally-occurring signals suggests that this type of learning is common in human systems. The neural network is built of Hebbian-type neurons with the modification introduced by Bienenstock, Cooper, and Munro (1982), which causes each neuron to become selective for one type of input pattern. These neurons are formed into groups and subgroups. Each group is responsible for detecting one type of input pattern. Every subgroup in a group is responsible for detecting the same input pattern with slightly different parameters.