Summary of Distilling Long-tailed Datasets, by Zhenghao Zhao et al.
Distilling Long-tailed Datasets
by Zhenghao Zhao, Haoxuan Wang, Yuzhang Shang, Kai Wang, Yan Yan
First submitted to arxiv on: 24 Aug 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: None
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The proposed method aims to address the issue of existing dataset distillation methods struggling with long-tailed datasets. Long-tailed datasets are common in real-world scenarios, and the current methods develop biased gradients when dealing with imbalanced data, leading to the synthesis of similarly imbalanced distilled datasets. This results in suboptimal performance on tail classes, which further hampers the quality of soft-label initialization. To overcome these limitations, the authors introduce Distribution-agnostic Matching and Expert Decoupling techniques. The former reduces the distance between the student and biased expert trajectories, while the latter improves distillation guidance by jointly matching decoupled backbone and classifier to enhance tail class performance. This work marks a pioneering effort in long-tailed dataset distillation. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Long-tailed datasets are a common challenge in machine learning. Existing methods for creating smaller, more focused datasets struggle with these types of datasets. When this happens, the resulting dataset can be biased towards certain classes, making it harder to learn from. The new method proposes two main solutions: first, avoid directly matching the expert’s trajectory by reducing the distance between the student and the biased expert trajectory. Second, improve distillation guidance by jointly matching the decoupled backbone and classifier to better handle tail class performance. This is an important step forward in creating high-quality datasets that are representative of real-world scenarios. |
Keywords
* Artificial intelligence * Distillation * Machine learning
