Domino Saliency Metrics: Improving Existing Channel Saliency Metrics with Structural Information

TL;DR

This paper introduces Domino saliency metrics that incorporate structural information to improve channel pruning in CNNs, especially in branched architectures, leading to better pruning efficiency and higher accuracy retention.

Contribution

The paper proposes Domino metrics that enhance existing channel saliency metrics by considering network structure, addressing limitations in pruning branched DNNs.

Findings

Domino metrics improved pruning rates in most networks.

Up to 25% accuracy improvement in AlexNet on CIFAR-10.

Enhanced pruning efficiency in branched architectures.

Abstract

Channel pruning is used to reduce the number of weights in a Convolutional Neural Network (CNN). Channel pruning removes slices of the weight tensor so that the convolution layer remains dense. The removal of these weight slices from a single layer causes mismatching number of feature maps between layers of the network. A simple solution is to force the number of feature map between layers to match through the removal of weight slices from subsequent layers. This additional constraint becomes more apparent in DNNs with branches where multiple channels need to be pruned together to keep the network dense. Popular pruning saliency metrics do not factor in the structural dependencies that arise in DNNs with branches. We propose Domino metrics (built on existing channel saliency metrics) to reflect these structural constraints. We test Domino saliency metrics against the baseline channel saliency metrics on multiple networks with branches. Domino saliency metrics improved pruning rates in most tested networks and up to 25% in AlexNet on CIFAR-10.