Soft Label Pruning and Quantization for Large-Scale Dataset Distillation

TL;DR

This paper introduces LPQLD, a method that significantly reduces soft label storage in large-scale dataset distillation while improving accuracy, by enhancing image and supervision diversity through pruning and quantization techniques.

Contribution

The paper proposes LPQLD, a novel approach combining label pruning and quantization to address label size issues and improve performance in large-scale dataset distillation.

Findings

Reduced soft label storage by 78x on ImageNet-1K and 500x on ImageNet-21K.

Achieved up to 7.2% and 2.8% accuracy improvements on ImageNet-1K and ImageNet-21K.

Validated effectiveness across different architectures and distillation methods.

Abstract

Large-scale dataset distillation requires storing auxiliary soft labels that can be 30-40x larger on ImageNet-1K and 200x larger on ImageNet-21K than the condensed images, undermining the goal of dataset compression. We identify two fundamental issues necessitating such extensive labels: (1) insufficient image diversity, where high within-class similarity in synthetic images requires extensive augmentation, and (2) insufficient supervision diversity, where limited variety in supervisory signals during training leads to performance degradation at high compression rates. To address these challenges, we propose Label Pruning and Quantization for Large-scale Distillation (LPQLD). We enhance image diversity via class-wise batching and batch-normalization supervision during synthesis. For supervision diversity, we introduce Label Pruning with Dynamic Knowledge Reuse to improve label-per-augmentation diversity, and Label Quantization with Calibrated Student-Teacher Alignment to improve augmentation-per-image diversity. Our approach reduces soft label storage by 78x on ImageNet-1K and 500x on ImageNet-21K while improving accuracy by up to 7.2% and 2.8%, respectively. Extensive experiments validate the superiority of LPQLD across different network architectures and dataset distillation methods. Code is available at https://github.com/he-y/soft-label-pruning-quantization-for-dataset-distillation.