Digital neural network on a XLINK XC4000
Jesus Ocampo Tuazon, Ch. Prachetan Reddy, B. Orenstein · 2002
A new digital neural network, consisting of 16 layers with 128 neurons per layer was built. An XLINK XC4005 was used as a processor for the multilayer network of 128-input digital neural gates. The neural gate is functionally similar to that of an analog gate except that its weights are limited to +1, 0 and -1, and the digital inputs are +1 and -1. Using these discrete weight and input values greatly simplify the discrimination function of the gate and hence could be implemented in digital circuits. The gate consists of N-input, N-bit weight vector and a threshold with value equal to K. A digital K-of-N (K/N) neural gate will 'fire' if the number of bits that matches between the input and the weight vector is K or more. The weight vectors and the K threshold values are stored in an EPROM together with their layered connectivity. Depending upon the architecture several gates may be processed simultaneously, like a vector operation. The XLINK XC4005 microprogrammed control unit supervises the processing and sequencing of the instructions. The system is tested to recognized 90 hand printed digits.