A VLSI neural network design approach using a priori knowledge

Analyzing neural network (NN) algorithms, a priori knowledge can be often obtained from: the number of classes in pattern classification; the patterns in content-addressable memory: the coefficients in the element, e.g., weight value; and the adapting patterns in layered networks. The passive procedures to embody such knowledge can be implemented by lookup table technology. A NN design approach using such knowledge is described. Three typical structures, including output, input, and the learning model, are discussed. Their NN designs are accompanied by the corresponding lookup table technology. Compared with the traditional NN implementations, it is shown that this approach reduces VLSI complexity, while maintaining the same high performance. Examples of several applications are used to illustrate the effectiveness of the approach