Summary of Partial Knowledge Distillation For Alleviating the Inherent Inter-class Discrepancy in Federated Learning, by Xiaoyu Gan et al.
Partial Knowledge Distillation for Alleviating the Inherent Inter-Class Discrepancy in Federated Learning
by Xiaoyu Gan, Xizi Chen, Jingyang Zhu, Xiaomeng Wang, Jingbo Jiang, Chi-Ying Tsui
First submitted to arxiv on: 23 Nov 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 In a recent study, researchers investigated the phenomenon of “weak classes” in federated learning, where certain classes consistently perform poorly even when class distribution is balanced. This issue affects not only minority classes but also inherently weak classes that exist independently of network structure or learning paradigm. The inherent accuracy discrepancy was found to be as high as 36.9% on FashionMNIST and CIFAR-10 datasets, despite balancing globally and locally. To address this issue, the authors proposed a class-specific partial knowledge distillation method that transfers knowledge upon specific misclassifications within weak classes. Experimental results show an improvement of 10.7% in weak class accuracy, effectively reducing the inter-class discrepancy. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Federated learning is a way for devices to work together and learn from each other’s data without sharing it. In this study, researchers found that some groups of things (called classes) tend to be really bad at getting the correct answer, even when all the groups are equally represented. This is not because there aren’t enough examples of those things in the data, but because they just don’t do well with learning. The researchers wanted to figure out why this happens and how to make it better. They came up with a way to teach the devices to learn from each other’s mistakes, especially when they’re getting something wrong that’s important for them to get right. This helped the devices get better at recognizing those things. |
Keywords
* Artificial intelligence * Federated learning * Knowledge distillation
