Efficient hardware acceleration for approximate inference of bitwise deep neural networks

Abstract

In recent years, Deep Neural Networks (DNNs) have been of special interest in the area of image processing and scene perception. Albeit being effective and accurate, DNNs demand challenging computational resources. Fortunately, dedicated low bitwidth accelerators enable efficient, real-time inference of DNNs. We present an approximate evaluation method and a specialized multiplierless accelerator for the recently proposed bitwise DNNs. Our approximate evaluation method is based on the speculative recomputation of selective parts of a bitwise neural network. The selection is based on the intermediate results of a previous input evaluation. In context with limited energy budgets, our method and accelerator enable a fast, power efficient, first decision. If necessary, a reliable and accurate output is available after reevaluating the input data multiple times in an approximate manner. Our experiments on the GTSRB and CIFAR-10 dataset show that this approach results in no loss of classification performance in comparison with floating-point evaluation. Our work contributes to efficient inference of neural networks on power-constrained embedded devices.