Associative-Memory-Based Systems with Recognition and Learning Capability ～Automatic Reference-Pattern Learning and Optimization～

Pattern recognition and learning are basic functions, which are needed to build artificial systems with capabilities similar to the human brain. Their effective implementation in integrated circuits is therefore of great technical importance. Among the methods for achieving the pattern recognition and the learning functions, having been proposed so far, the neural-network approach is most widely used. However, the performance progress of hardware that uses neural-networks is much slower than expected initially. Because of this difficult situation a new method that includes the memory element, missing up to now, in a power-efficient LSI hardware is urgently needed. Presently, we are developing a new associative-memory architecture which achieves small area and high near-est-match speed [1-3]. The proposed architecture can search the reference pattern of minimum distance to the input pattern at high speed for different distance measures. Therefore, it is expected that this small-area and high-speed associative memory becomes the basis for a new method to construct systems with recognition and the learning capa-bility. Moreover, there is the advantage that it can be easily integrated by the use of present CMOS technology. Here we report 2 new associative-memory-based automatic-learning architectures for artificial intelligence systems that have recognition and learning capability. The developed associative memory that we have developed is used to imitate the long-term and the short-term memory of the human brain(Fig. 1) and only a simple adder and subtractor for learning the optimized reference patterns