WorksEngineering
ASDL: A Fast Vector Approximation Algorithm for Multiplierless Designs of Convolutional Neural Networks
Abstract
Matrix multiplications are one of the main computational bottlenecks in neural networks. The calculation reduces to numerous sums of partial products, which scale with the total number of Signed Power of Two (SPT) values in the multiplier matrix. Minimizing the SPT count can significantly reduce the computational burden.In this context, we present a Sum Of (signed) Powers Of Two (SOPOT) vector approximation algorithm that limits SPT counts, namely the Accelerated Signed Digit Loading (ASDL) algorithm. This approximation algorithm converts ResNet50 times faster than its predecessor MPGBP, while being more precise. We apply this method to the pre-trained coefficients of different convolutional neural networks, achieving ImageNet accuracy down from , at an average of only active bits per coefficient. This is done without any retraining or calibration, which are typically required for the most aggressive quantization methods. Our approach offers a promising direction for neural network inference with higher energy efficiency and fewer hardware requirements: we anticipate up to times improvement over float16 inference.