Enhancing Dataset Distillation with Difficulty-Aware Guidance

カテゴリ：部門大会

論文No：OS4-2-5

グループ名：【C】2025年電気学会電子・情報・システム部門大会

発行日：2025/8/20

タイトル(英語)：Enhancing Dataset Distillation with Difficulty-Aware Guidance

著者名：Zou Yawen（富山大学）,Cai Wenqi（富山大学）,Wang Zi（新潟大学）,Gu Chunzhi（豊橋技術科学大学）,Zhang Chao（富山大学）

著者名(英語)： Yawen Zou (University of Toyama),Wenqi Cai (University of Toyama),Zi Wang (Niigata University),Chunzhi Gu (Toyohashi University of Technology),Chao Zhang (University of Toyama)

キーワード：Dataset Distillation,Difficulty-Aware,Diffusion Model

要約(日本語)：Dataset distillation aims to synthesize a compact yet informative dataset that enables training models with performance comparable to those trained on full data. However, existing methods often suffer from limited diversity, particularly lacking a balanced distribution of easy and hard instances. We propose a detector-guided distillation framework that leverages a detector trained on the original dataset to assess sample difficulty. During synthesis, difficulty scores guide the diffusion model to ensure both easy and hard examples are included. This promotes richer supervision signals and improves generalization. Experiments on benchmark datasets show that our approach improves the performance and robustness of models trained on distilled data, especially under limited data regimes.

本誌掲載ページ：964-968p

原稿種別：英語

PDFファイルサイズ：343Kバイト