Reference-pattern learning and optimization from an input-pattern stream for associative-memory-based pattern-recognition systems

A learning algorithm of the optimized data base for an associative-memory-based pattern recognition system from a time-serial input-pattern stream is presented. The algorithm covers learning of both, the reference patterns and the recognition thresholds being most suitable for a highly reliable recognition of new input patterns. Algorithm simulation verifies a better performance than the conventional k-means algorithm. In particular, learned-pattern quality is improved for a non-Gaussian distribution of the object patterns in the input-pattern stream. A VLSI-chip architecture for integration of this learning algorithm with a fully-parallel associative memory is proposed in addition.