Associative digital neural network based on code and graph theories

An associative DNN (digital neural network) called SANNET, which is based on code and graph theories, is presented. Each neuron is an exclusive-OR unit. In the learning process, all node syndromes from all neurons are constrained to be zero, and a binary `current´ code that expresses a loop on each subnetwork is generated on the basis of the orthogonality of loops and cutsets. Only information on the tree is stored in each subnetwork. In the associative process, the incomplete code on the cotree of each subnetwork is corrected to the complete code according to the error-correcting capacity. SANNET has structural redundancy, sparsity, cellular structure, high concurrency, variable code length, robustness, testability, reliability, logical neurons, unions of logic and storage elements, no crosstalk, high speed, and a unique solution. The SANNET can be fabricated as a CMOS gate array. It can be applied to recognition and classification in which the recognized items are expressed by feature vectors with different lengths. Some simulation results for pattern completion are given