Shift-add neural architecture

This article is focused on implementation of artificial neural networks in hardware. We give an overview of the shift-add neural arithmetics, which provide a complete set of functions suitable for fast perceptron and RBF network implementations. The set consists of logarithm, exponent, multiplication, square, square root, sigmoid-like and Gauss-like functions. All functions are linearly approximated to be easy implementable. Furthermore, we show the gate-level implementation of all functions provided by the shift-add arithmetics. Only adders and barrel shifters are necessary to accomplish all functions. The functions are optimized for very short propagation delay (a few nanoseconds). The shift-add architecture was evaluated by both software simulation and on-chip design. Results of the on-chip design are presented in the last section of this article