A design method for multilayer feedforward neural networks for simple hardware implementation

A method for designing a multiplierless multilayer feedforward neural network for continuous input-output mapping is presented. This method uses the simplified sigmoid activation functions at the weights in the output layer, 3-level discrete quantization functions at the hidden neurons, and single powers-of-two weights in the input layer. When tested with noisy vectors, the multiplierless network can achieve high recall accuracy, while having increased computational speed in practical applications and reduced hardware cost in digital implementation