Summary of Generalized Categories Discovery For Long-tailed Recognition, by Ziyun Li et al.
Generalized Categories Discovery for Long-tailed Recognition
by Ziyun Li, Christoph Meinel, Haojin Yang
First submitted to arxiv on: 4 Dec 2023
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Machine Learning (cs.LG)
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 This research paper presents a novel approach to Generalized Class Discovery (GCD), which enables the detection of both known and unknown categories in unlabeled datasets. The authors address the limitation of existing GCD methods, which assume an equitable distribution of category occurrences in the unlabeled data. In reality, many real-world datasets exhibit a long-tailed distribution, where common or frequent categories appear more often than rare ones. To bridge this gap, the researchers propose a Long-tailed Generalized Category Discovery (Long-tailed GCD) paradigm that mirrors these imbalances. They introduce a robust methodology comprising two strategic regularizations: reweighting to emphasize less-represented categories and class prior constraints aligned with anticipated distributions. Experimental results show that their proposed method outperforms previous state-of-the-art GCD methods by approximately 6-9% on ImageNet100 and achieves competitive performance on CIFAR100. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper helps machines discover new categories in pictures without labels. The problem is, current methods assume all categories appear equally often. But that’s not true for real-world datasets, where common things like sky or trees are much more frequent than rare things like rainbows or dinosaurs. The researchers propose a new way to find these hidden categories by adjusting how they look at the data and using two special tricks: making less common things count more and setting expectations based on what we know about pictures. Their method beats current best practices by 6-9% on one big dataset and does well on another. |
