Compact Deep Convolutional Neural Networks With Coarse Pruning

TL;DR

This paper introduces a coarse pruning method at feature map and kernel levels to create compact deep convolutional neural networks, significantly reducing computational costs while maintaining accuracy.

Contribution

It presents a novel, simple strategy for coarse-grained pruning that is hardware-friendly and effectively reduces network complexity without substantial accuracy loss.

Findings

Achieved over 85% sparsity in convolution layers

Less than 1% increase in misclassification rate

Effective pruning strategy for real-time inference

Abstract

The learning capability of a neural network improves with increasing depth at higher computational costs. Wider layers with dense kernel connectivity patterns furhter increase this cost and may hinder real-time inference. We propose feature map and kernel level pruning for reducing the computational complexity of a deep convolutional neural network. Pruning feature maps reduces the width of a layer and hence does not need any sparse representation. Further, kernel pruning converts the dense connectivity pattern into a sparse one. Due to coarse nature, these pruning granularities can be exploited by GPUs and VLSI based implementations. We propose a simple and generic strategy to choose the least adversarial pruning masks for both granularities. The pruned networks are retrained which compensates the loss in accuracy. We obtain the best pruning ratios when we prune a network with both granularities. Experiments with the CIFAR-10 dataset show that more than 85% sparsity can be induced in the convolution layers with less than 1% increase in the missclassification rate of the baseline network.