跳到正文
研踪PaperTrail
← 研踪 PaperTrail

arXiv:2609.34299 · v2

Difficulty-Aware Statistical Matching for Dataset Distillation

首次提交 2026-09-28 12:43
版本更新 2026-11-18 09:45
雷达观测 2026-11-19 19:08
arXiv ↗PDF ↗

摘要

Dataset distillation (DD) condenses a large original set into a small distilled set with high training utility. Decoupled statistical matching methods substantially reduce distillation time and memory overhead while achieving strong performance. However, they typically supervise all distilled data with running statistics estimated from the entire original set. These statistics mainly capture the average feature distribution while ignoring the difficulty differences among original data, making them insufficient for characterizing the difficulty structure of the original data. To address this issue, we propose Difficulty-aware Statistical Matching (DSM) for Dataset Distillation. After pretraining, DSM uses the Global Difficulty Score (GDS) to evaluate image difficulty, sorts the original data according to the scores, and then partitions it into IPC (Images per Class of distilled data) difficulty groups. During distillation, Statistics Updated Again (SUA) updates the teacher's BN running statistics using original data from each group via forward passes, providing difficulty-specific supervision for each corresponding distilled batch. Meanwhile, Initial Sample Screening (ISS) initializes distilled data with original images from the corresponding group, providing an effective starting point for precise matching. Experiments across multiple datasets and model architectures demonstrate the effectiveness of DSM, showing that it broadens the difficulty range of distilled samples and further improves the final performance. Code will be released.

AI 论文解读

基于摘要 · 非全文审阅

此版本尚未生成 AI 解读。解读由后台任务生成并按论文版本保存,登录后共享已有结果。

版本差异

本站首次收录为 v2,暂无更早版本可对照。

观测到的版本

相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。

v2 · 2026-11-18 09:45 · 最新观测版本

Difficulty-Aware Statistical Matching for Dataset Distillation

Dataset distillation (DD) condenses a large original set into a small distilled set with high training utility. Decoupled statistical matching methods substantially reduce distillation time and memory overhead while achieving strong performance. However, they typically supervise all distilled data with running statistics estimated from the entire original set. These statistics mainly capture the average feature distribution while ignoring the difficulty differences among original data, making them insufficient for characterizing the difficulty structure of the original data. To address this issue, we propose Difficulty-aware Statistical Matching (DSM) for Dataset Distillation. After pretraining, DSM uses the Global Difficulty Score (GDS) to evaluate image difficulty, sorts the original data according to the scores, and then partitions it into IPC (Images per Class of distilled data) difficulty groups. During distillation, Statistics Updated Again (SUA) updates the teacher's BN running statistics using original data from each group via forward passes, providing difficulty-specific supervision for each corresponding distilled batch. Meanwhile, Initial Sample Screening (ISS) initializes distilled data with original images from the corresponding group, providing an effective starting point for precise matching. Experiments across multiple datasets and model architectures demonstrate the effectiveness of DSM, showing that it broadens the difficulty range of distilled samples and further improves the final performance. Code will be released.

查看此版本解读 →查看此版本原文 ↗