Pushing the Efficiency Limit Using Structured Sparse Convolutions

TL;DR

This paper introduces Structured Sparse Convolution (SSC), a method leveraging image structure to reduce convolutional parameters, enhancing efficiency and achieving state-of-the-art results across multiple image classification benchmarks.

Contribution

The paper proposes SSC, a novel structured sparsity approach that generalizes common convolutional layers, improving efficiency without extensive pruning procedures.

Findings

SSC achieves state-of-the-art accuracy on CIFAR-10 and CIFAR-100.

SSC outperforms existing pruning methods in efficiency.

Extensive experiments validate SSC's effectiveness across multiple datasets.

Abstract

Weight pruning is among the most popular approaches for compressing deep convolutional neural networks. Recent work suggests that in a randomly initialized deep neural network, there exist sparse subnetworks that achieve performance comparable to the original network. Unfortunately, finding these subnetworks involves iterative stages of training and pruning, which can be computationally expensive. We propose Structured Sparse Convolution (SSC), which leverages the inherent structure in images to reduce the parameters in the convolutional filter. This leads to improved efficiency of convolutional architectures compared to existing methods that perform pruning at initialization. We show that SSC is a generalization of commonly used layers (depthwise, groupwise and pointwise convolution) in ``efficient architectures.'' Extensive experiments on well-known CNN models and datasets show the effectiveness of the proposed method. Architectures based on SSC achieve state-of-the-art performance compared to baselines on CIFAR-10, CIFAR-100, Tiny-ImageNet, and ImageNet classification benchmarks.