Two-Level Neural Network For Deterministic Logic Processing

A two-level neural network is proposed for the implementation of general deterministic logic (switching) functions. The network is potentially capable of implementing any set of binary switching functions of n variables. A cascade of two neural-like processor levels gives rise to a high-performance nonlinear functional memory. The first neural layer implements a linearly separable psuedorandom mapping that maps n dimensional binary input vectors into a higher m dimensional space of randomly scattered vectors, while the second neural layer implements a one-pass associative neural memory (ANM) that maps the output of the first layer into prerecorded target vectors. The interconnection weights of this layer are synthesized using a new and highly efficient recording technique[l]. The high fan-out of the first layer mapping and the highly distributed parallel architecture of the proposed network are ideal for optical implementation.