Minimum Variance Unbiased N:M Sparsity for the Neural Gradients

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Chmiel, Brian, Hubara, Itay, Banner, Ron, Soudry, Daniel
Format: Preprint
Published: 2022
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866913382624395264
author Chmiel, Brian
Hubara, Itay
Banner, Ron
Soudry, Daniel
author_facet Chmiel, Brian
Hubara, Itay
Banner, Ron
Soudry, Daniel
contents In deep learning, fine-grained N:M sparsity reduces the data footprint and bandwidth of a General Matrix multiply (GEMM) up to x2, and doubles throughput by skipping computation of zero values. So far, it was mainly only used to prune weights to accelerate the forward and backward phases. We examine how this method can be used also for the neural gradients (i.e., loss gradients with respect to the intermediate neural layer outputs). To this end, we first establish a tensor-level optimality criteria. Previous works aimed to minimize the mean-square-error (MSE) of each pruned block. We show that while minimization of the MSE works fine for pruning the weights and activations, it catastrophically fails for the neural gradients. Instead, we show that accurate pruning of the neural gradients requires an unbiased minimum-variance pruning mask. We design such specialized masks, and find that in most cases, 1:2 sparsity is sufficient for training, and 2:4 sparsity is usually enough when this is not the case. Further, we suggest combining several such methods together in order to potentially speed up training even more.
format Preprint
id arxiv_https___arxiv_org_abs_2203_10991
institution arXiv
publishDate 2022
record_format arxiv
spellingShingle Minimum Variance Unbiased N:M Sparsity for the Neural Gradients
Chmiel, Brian
Hubara, Itay
Banner, Ron
Soudry, Daniel
Machine Learning
Artificial Intelligence
In deep learning, fine-grained N:M sparsity reduces the data footprint and bandwidth of a General Matrix multiply (GEMM) up to x2, and doubles throughput by skipping computation of zero values. So far, it was mainly only used to prune weights to accelerate the forward and backward phases. We examine how this method can be used also for the neural gradients (i.e., loss gradients with respect to the intermediate neural layer outputs). To this end, we first establish a tensor-level optimality criteria. Previous works aimed to minimize the mean-square-error (MSE) of each pruned block. We show that while minimization of the MSE works fine for pruning the weights and activations, it catastrophically fails for the neural gradients. Instead, we show that accurate pruning of the neural gradients requires an unbiased minimum-variance pruning mask. We design such specialized masks, and find that in most cases, 1:2 sparsity is sufficient for training, and 2:4 sparsity is usually enough when this is not the case. Further, we suggest combining several such methods together in order to potentially speed up training even more.
title Minimum Variance Unbiased N:M Sparsity for the Neural Gradients
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2203.10991