Mix-and-Match Pruning: Globally Guided Layer-Wise Sparsification of DNNs

TL;DR

Mix-and-Match Pruning is a globally guided, layer-wise sparsification framework that improves DNN compression by leveraging sensitivity scores and architectural rules to generate diverse, high-quality pruning strategies, enhancing deployment on edge devices.

Contribution

The paper introduces a novel framework that systematically combines multiple pruning signals and architectural considerations to produce diverse, high-quality sparsification strategies for DNNs.

Findings

Achieves Pareto-optimal accuracy-sparsity trade-offs on CNNs and Vision Transformers.

Reduces accuracy degradation by 40% on Swin-Tiny compared to standard pruning.

Demonstrates that combining existing signals yields more reliable compressed models.

Abstract

Deploying deep neural networks (DNNs) on edge devices requires strong compression with minimal accuracy loss. This paper introduces Mix-and-Match Pruning, a globally guided, layer-wise sparsification framework that leverages sensitivity scores and simple architectural rules to generate diverse, high-quality pruning configurations. The framework addresses a key limitation that different layers and architectures respond differently to pruning, making single-strategy approaches suboptimal. Mix-and-Match derives architecture-aware sparsity ranges, e.g., preserving normalization layers while pruning classifiers more aggressively, and systematically samples these ranges to produce ten strategies per sensitivity signal (magnitude, gradient, or their combination). This eliminates repeated pruning runs while offering deployment-ready accuracy-sparsity trade-offs. Experiments on CNNs and Vision Transformers demonstrate Pareto-optimal results, with Mix-and-Match reducing accuracy degradation on Swin-Tiny by 40% relative to standard single-criterion pruning. These findings show that coordinating existing pruning signals enables more reliable and efficient compressed models than introducing new criteria.