Integer Convolutional Neural Networks with Boolean Activations: The BoolHash Algorithm

Abstract

Improving the efficiency of convolutional neural networks (CNN) often relies on integer-only algorithms. Using boolean activations can bring further inference speed gain, and can make easier the design of CNN-specific ASICs. A convolutional algorithm called BoolHash that we propose here can additionally increase the inference speed several times, and permits functionalities that usually require more complex processing. A CNN model with 16-bit input weights, 8-bit filter weights and 1-bit activations was used to compare the speed of BoolHash to that of a classic weight-adder convolutional algorithm.