Modular feedforward neural network architecture with learning

Abstract

A first feedforward network receives an input vector signal and a plurality of weight signals and forms an output vector signal based thereupon. A second feedforward network, substantially identical to the first feedforward network, receives a first learning vector signal and the weight signals and forms a learning output vector based thereupon. A weight updating circuit generates the weight signals in accordance with a back propagation updating rule. The weight signals are updated based upon the first learning vector signal and a second learning vector signal in response to receiving a first predetermined layer signal, and are updated based upon the first learning vector signal, the second learning vector signal, and the learning output vector signal in response to receiving a second predetermined layer signal. A back propagation feedback network receives the second learning vector signal and the weight signals and generates a back propagated error vector signal based thereupon. The feedforward networks, the weight updating circuit, and the back propagation feedback network are implemented as analog circuits on a single integrated circuit.

Dependencies are inferred from the claim text.