RESEARCH

And the Bit Goes Down: Revisiting the Quantization of Neural Networks

January 20, 2020

Abstract

In this paper, we address the problem of reducing the memory footprint of convolutional network architectures. We introduce a vector quantization method that aims at preserving the quality of the reconstruction of the network outputs rather than its weights. The principle of our approach is that it minimizes the loss reconstruction error for in-domain inputs. Our method only requires a set of unlabelled data at quantization time and allows for efficient inference on CPU by using byte-aligned codebooks to store the compressed weights. We validate our approach by quantizing a high performing ResNet-50 model to a memory size of 5MB (20x compression factor) while preserving a top-1 accuracy of 76.1% on ImageNet object classification and by compressing a Mask R-CNN with a 26x factor.

Download the Paper

AUTHORS

Written by

Pierre Stock

Armand Joulin

Benjamin Graham

Hervé Jegou

Rémi Gribonval

Publisher

ICLR

Related Publications

December 15, 2021

RESEARCH

Sample-and-threshold differential privacy: Histograms and applications

Akash Bharadwaj, Graham Cormode

December 15, 2021

January 09, 2021

RESEARCH

COMPUTER VISION

Tarsier: Evolving Noise Injection in Super-Resolution GANs

Baptiste Rozière, Camille Couprie, Olivier Teytaud, Andry Rasoanaivo, Hanhe Lin, Nathanaël Carraz Rakotonirina, Vlad Hosu

January 09, 2021

January 09, 2021

RESEARCH

Improved Sample Complexity for Incremental Autonomous Exploration in MDPs

Jean Tarbouriech, Alessandro Lazaric, Matteo Pirotta, Michal Valko

January 09, 2021

December 07, 2020

RESEARCH

COMPUTER VISION

Labelling unlabelled videos from scratch with multi-modal self-supervision

Mandela Patrick

December 07, 2020