A Hierarchical ART Network Model for General Pattern Recognition

A new neural network architecture is introduced which may be used for fault-tolerant general pattern recognition. Images are learned by extracting features at each layer. These same images may later be recognized by extracting features which are then used to constrain a search for additional features to validate one of a set of chosen image representation candidates. Unsupervised learning of feature patterns at each layer is accomplished using adaptive resonance theory (ART1) networks. recognition phase in which any previously seen pattern may be identified. The paper begins with an overview of the ART1 algorithm along with modifications to make it suitable for a hierarchical structure building block. A description of the model follows in addition to discussion of the learning and recognition algorithms.