Summary of Disentangling Masked Autoencoders For Unsupervised Domain Generalization, by An Zhang et al.
Disentangling Masked Autoencoders for Unsupervised Domain Generalization
by An Zhang, Han Wang, Xiang Wang, Tat-Seng Chua
First submitted to arxiv on: 10 Jul 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Artificial Intelligence (cs.AI)
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 Domain Generalization (DG) aims to learn invariance against domain shifts using sufficient supervision signals. However, the scarcity of labeled data has led to the rise of unsupervised domain generalization (UDG), a more challenging task where models are trained across diverse domains without labels and tested on unseen domains. To address this gap, we propose DisMAE, a novel learning framework for UDG that discovers disentangled representations revealing intrinsic features and superficial variations without class labels. The core is distilling domain-invariant semantic features, filtering out unstable and redundant variations. Co-training an asymmetric dual-branch architecture with semantic and lightweight variation encoders enables dynamic data manipulation and representation-level augmentation capabilities. We evaluate DisMAE on four benchmark datasets (DomainNet, PACS, VLCS, Colored MNIST) for both DG and UDG tasks, demonstrating competitive OOD performance compared to state-of-the-art DG and UDG baselines. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper is about making computer models learn new things even when they don’t have labeled data. This is a big challenge because most learning happens with labeled data. The researchers propose a new way to do this, called DisMAE. It helps the model discover what’s truly important and filter out what’s not. They test it on several different datasets and show that it can perform well even without labeled data. |
Keywords
» Artificial intelligence » Domain generalization » Unsupervised
