Information classification scheme of feedforward networks organised under unsupervised learning

Typical unsupervised learning, also called self-organisation, makes each neuron responsive to one of the input signals frequently presented. The prominent response of a neuron has been considered as a feature-detecting signal. It is found that a network organised under unsupervised learning has an information classification scheme rather than a mere ability to detect. Each bit of information is represented by an activity pattern of the set of neurons in the input layer and it is mapped into another activity pattern of the set of neurons in the output layer. Analysis of the activity pattern of multiple output neurons revealed the following tendencies beyond the feature detection by a single neuron. First, nearness and remoteness between pieces of information are emphasised by the mapping. Thus the neighbouring pieces are clustered. Secondly, the mapping emphasises the interrelationship implicit in the pieces of information and tends to categorise them. The point is discussed from a viewpoint of ultrametricity.