Review of: "Enhancing Dataset Distillation via Label Inconsistency Elimination and Learning Pattern Refinement"
Alberto Fernández · 2024
Potential competing interests: No potential competing interests to declare.This paper presents a solution for the ECCV-2024 Data Distillation Challenge, which aims to create smaller synthetic datasets that enable machine learning models to perform comparably to models trained on the full datasets.The authors propose a modified version of the Difficulty-Aligned Trajectory Matching (DATM) method, named M-DATM, designed to address issues related to label inconsistency and the difficulty of learning hard patterns, especially on challenging datasets like Tiny ImageNet.M-DATM achieves high performance by removing soft labels to ensure label consistency and