Inducing Semi-Structured Sparsity by Masking for Efficient Model Inference in Convolutional Networks

TL;DR

This paper introduces a method to learn semi-structured sparsity in convolutional neural networks through masking, significantly accelerating inference while maintaining model accuracy and allowing easy updates.

Contribution

It presents a novel masking technique to induce semi-structured sparsity, enabling hardware acceleration without altering original model weights or structure.

Findings

Over two-fold inference speedup achieved.

Model performance remains unchanged.

Stability bounds for predictions under maskings are derived.

Abstract

The crucial role of convolutional models, both as standalone vision models and backbones in foundation models, necessitates effective acceleration techniques. This paper proposes a novel method to learn semi-structured sparsity patterns for convolution kernels in the form of maskings enabling the utilization of readily available hardware accelerations. The approach accelerates convolutional models more than two-fold during inference without decreasing model performance. At the same time, the original model weights and structure remain unchanged keeping the model thus easily updatable. Beyond the immediate practical use, the effect of maskings on prediction is easily quantifiable. Therefore, guarantees on model predictions under maskings are derived showing stability bounds for learned maskings even after updating the original underlying model.