DART1-a possible ideal building block of memory systems

In this paper, a bidirectional associative memory dual adaptive resonance theory 1 (DART1) consisting of two simplified ART1 (SART1) networks, which share the same category layer called the conceptual layer and use only top-down weight matrices, is studied. The DART1 can meet the most important requirement of the neural networks as the memory device, that is, the ability to store arbitrary patterns in networks and retrieve them correctly. Other advantages of this model include the capability of associating one pattern with another arbitrary pattern like the MLP (multilayer perceptron) networks, eliminating the input noise as a nearest-neighbor classifier in terms of the Hamming distance, capable of handling both spatial and temporal patterns, full memory capacity like the ART1 networks, fast construction of weights, easy to update the memory system, and self-organizing ability in response to arbitrary binary patterns. In addition, the DART1 can deal with the 1-to-many and many-to-1 mappings in the domains formed by trained patterns. These features make the DART1 very attractive as the building block of memory system. Therefore, it is possible to build the neural-based database or expert systems, which are the topics of future research