3. Basic principle of the neocognitron

Hierarchical feature extraction is the basic principle of the neocognitron. For easier understanding of this principle we will explain terms feature and feature extraction firstly.

Hierarchical feature extraction consists in distribution of extracted features to several stages. The simplest features (usually only rotated lines) are extracted in the first stage and in each of the following stages the more complex features are extracted. In this process it is important fact that only informations obtained in the previous stage are used for feature extraction in the certain stage.

Hierarchy of features which we can use for recognition of digit zero in the neocognitron is illustrated schematically in figure 3.1. Joining between features in adjoining stages represents what features from previous layer are used during extraction of the certain feature.

Fig. 3.1 - Principle of the hierarchical feature extraction