Hessian-Aware Pruning and Optimal Neural Implant

TL;DR

This paper introduces Hessian-Aware Pruning (HAP) combined with Neural Implant, a novel structured pruning method that uses second-order sensitivity to prune neural networks effectively while maintaining accuracy.

Contribution

The paper proposes a new Hessian-aware pruning method with Neural Implant that improves pruning efficiency and accuracy preservation by using second-order sensitivity metrics.

Findings

Achieves less than 0.1%/0.5% accuracy degradation with over 70%/50% parameter pruning on ResNet models.

Outperforms gradient-based head pruning methods on transformer models with up to 0.8% accuracy gain at 60% pruning.

State-of-the-art results on computer vision and natural language tasks using the proposed method.

Abstract

Pruning is an effective method to reduce the memory footprint and FLOPs associated with neural network models. However, existing structured-pruning methods often result in significant accuracy degradation for moderate pruning levels. To address this problem, we introduce a new Hessian Aware Pruning (HAP) method coupled with a Neural Implant approach that uses second-order sensitivity as a metric for structured pruning. The basic idea is to prune insensitive components and to use a Neural Implant for moderately sensitive components, instead of completely pruning them. For the latter approach, the moderately sensitive components are replaced with with a low rank implant that is smaller and less computationally expensive than the original component. We use the relative Hessian trace to measure sensitivity, as opposed to the magnitude based sensitivity metric commonly used in the literature. We test HAP for both computer vision tasks and natural language tasks, and we achieve new state-of-the-art results. Specifically, HAP achieves less than $0.1\%$/$0.5\%$ degradation on PreResNet29/ResNet50 (CIFAR-10/ImageNet) with more than 70\%/50\% of parameters pruned. Meanwhile, HAP also achieves significantly better performance (up to 0.8\% with 60\% of parameters pruned) as compared to gradient based method for head pruning on transformer-based models. The framework has been open sourced and available online.