Distilled Pruning: Using Synthetic Data to Win the Lottery

TL;DR

This paper presents a new pruning method using distilled data to efficiently identify sparse, trainable subnetworks, significantly reducing pruning time while maintaining performance.

Contribution

It introduces a novel data-driven pruning approach leveraging distilled datasets to find lottery tickets faster than traditional methods.

Findings

Achieves up to 5x faster pruning on CIFAR-10

Maintains comparable sparsity and accuracy to iterative pruning

Demonstrates potential for resource-efficient model compression

Abstract

This work introduces a novel approach to pruning deep learning models by using distilled data. Unlike conventional strategies which primarily focus on architectural or algorithmic optimization, our method reconsiders the role of data in these scenarios. Distilled datasets capture essential patterns from larger datasets, and we demonstrate how to leverage this capability to enable a computationally efficient pruning process. Our approach can find sparse, trainable subnetworks (a.k.a. Lottery Tickets) up to 5x faster than Iterative Magnitude Pruning at comparable sparsity on CIFAR-10. The experimental results highlight the potential of using distilled data for resource-efficient neural network pruning, model compression, and neural architecture search.