From Fewer Samples to Fewer Bits: Reframing Dataset Distillation as Joint Optimization of Precision and Compactness

TL;DR

This paper introduces QuADD, a novel dataset distillation framework that jointly optimizes dataset size and data precision, leading to more efficient training data representations with better accuracy per bit.

Contribution

It presents a unified, end-to-end approach integrating quantization into dataset distillation, enabling joint optimization of sample count and precision under fixed bit budgets.

Findings

QuADD outperforms existing methods in accuracy per bit.

Adaptive non-uniform quantization improves data representation.

Joint optimization enhances efficiency in image and communication tasks.

Abstract

Dataset Distillation (DD) compresses large datasets into compact synthetic ones that maintain training performance. However, current methods mainly target sample reduction, with limited consideration of data precision and its impact on efficiency. We propose Quantization-aware Dataset Distillation (QuADD), a unified framework that jointly optimizes dataset compactness and precision under fixed bit budgets. QuADD integrates a differentiable quantization module within the distillation loop, enabling end-to-end co-optimization of synthetic samples and quantization parameters. Guided by the rate-distortion perspective, we empirically analyze how bit allocation between sample count and precision influences learning performance. Our framework supports both uniform and adaptive non-uniform quantization, where the latter learns quantization levels from data to represent information-dense regions better. Experiments on image classification and 3GPP beam management tasks show that QuADD surpasses existing DD and post-quantized baselines in accuracy per bit, establishing a new standard for information-efficient dataset distillation.