Emergent on-line learning in min-max modular neural networks

This paper presents a novel on-line supervised learning model called emergent on-line learning for pattern classification. The model involves three mechanisms: decomposition of an on-line learning problem at each time step into a reasonable number of linearly separable problems; parallel learning of these linearly separable problems by using linear threshold gates; and integration of the trained linear threshold gates into a min-max modular network. Two simple emergent laws are used to control both the problem decomposition and solution integration. The advantages of the model are very fast learning speed, guaranteed convergence, high modularity, and parallelism.