Self Similarity Matrix based CNN Filter Pruning

TL;DR

This paper introduces a novel CNN filter pruning method using Self-Similarity Matrices that reduces model complexity without fine-tuning, demonstrated on ResNet and VGG for CIFAR-10.

Contribution

The paper presents two new algorithms leveraging SSM for filter ranking and pruning, enabling simultaneous training and pruning without fine-tuning.

Findings

Effective filter pruning without fine-tuning

Significant reduction in model size and complexity

Maintained accuracy on CIFAR-10 dataset

Abstract

In recent years, most of the deep learning solutions are targeted to be deployed in mobile devices. This makes the need for development of lightweight models all the more imminent. Another solution is to optimize and prune regular deep learning models. In this paper, we tackle the problem of CNN model pruning with the help of Self-Similarity Matrix (SSM) computed from the 2D CNN filters. We propose two novel algorithms to rank and prune redundant filters which contribute similar activation maps to the output. One of the key features of our method is that there is no need of finetuning after training the model. Both the training and pruning process is completed simultaneously. We benchmark our method on two of the most popular CNN models - ResNet and VGG and record their performance on the CIFAR-10 dataset.