An ART2 trained by two-stage learning on circularly ordered data sequence

Adaptive Resonance Theory (ART), characterized by its built-in mechanism of handling the stability-plasticity dilemma and by fast adaptive learning without forgetting informations learned in the past, is based on an unsupervised template matching. We propose an improved two-stage learning algorithm for ART2: the original unsupervised learning followed by a new supervised learning. Each of the output nodes, after the unsupervised learning, is labeled according to the category informations of the feature vectors associated with the node. In the supervised learning, each feature vector is used to reinforce the template pattern associated with the target output node belonging to the same category as the feature vector. Another modification is a circular ordering of the training sequence, which is intended to prevent some dominant classes from exhausting a finite number of template patterns in ART2. The proposed learning algorithm has been shown to yield better accuracy than the original ART2, regardless of the size of the network. The hold-out recognition accuracy of the modified ART2 on the real data obtained from military ship images is 98.4%, and that of the original ART is 94.8%, when the size of the network is chosen reasonably in such a way that the size is minimized while maintaining the required accuracy