Neural networks with nonlinear weights for pattern classification

The adoption of nonlinear weights in artificial neural networks for pattern matching applications is studied. These weights laterally connect the processing elements of the output layers and force the output of the nondominant processing elements to converge to a low level. This facilitates the selection of the closest stored pattern. It is shown that the adoption of nonlinear weights in a Hamming net significantly improves performance and reduces complexity. A multistage Hamming net is also proposed. The memory capacity and training of this net are also studied.<>