A MAC-less Neural Inference Processor Supporting Compressed, Variable\n Precision Weights

2020/12/10 by Vincenzo Liguori, Liguori, Vincenzo
Computer Science · #Advanced Neural Network Applications #Adversarial Robustness in Machine Learning #B.0 #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Human Pose and Action Recognition #Neural Networks and Applications

paper · pdf · doi:10.48550/arxiv.2012.06018

openalex publication_date 2020/12/10 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28

Abstract

This paper introduces two architectures for the inference of convolutional\nneural networks (CNNs). Both architectures exploit weight sparsity and\ncompression to reduce computational complexity and bandwidth. The first\narchitecture uses multiply-accumulators (MACs) but avoids unnecessary\nmultiplications by skipping zero weights. The second architecture exploits\nweight sparsity at the level of their bit representation by substituting\nresource-intensive MACs with much smaller Bit Layer Multiply Accumulators\n(BLMACs). The use of BLMACs also allows variable precision weights as variable\nsize integers and even floating points. Some details of an implementation of\nthe second architecture are given. Weight compression with arithmetic coding is\nalso discussed as well as bandwidth implications. Finally, some implementation\nresults for a pathfinder design and various technologies are presented.\n

Citations

• YOLOv3: An Incremental Improvement

Related

• ShiftCNN: Generalized Low-Precision Architecture for Inference of Convolutional Neural Networks
• Pyramid Vector Quantization and Bit Level Sparsity in Weights for Efficient Neural Networks Inference
• Deep Convolutional Neural Network Inference with Floating-point Weights and Fixed-point Activations
• Hardware-Efficient CNNs: Interleaved Approximate FP32 Multipliers for Kernel Computation
• UCNN: Exploiting Computational Reuse in Deep Neural Networks via Weight\n Repetition