Summary of Is Large-scale Pretraining the Secret to Good Domain Generalization?, by Piotr Teterwak et al.
Is Large-Scale Pretraining the Secret to Good Domain Generalization?
by Piotr Teterwak, Kuniaki Saito, Theodoros Tsiligkaridis, Bryan A. Plummer, Kate Saenko
First submitted to arxiv on: 3 Dec 2024
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 The paper presents a study on Multi-Source Domain Generalization (DG), focusing on improving classification performance on unseen target domains by leveraging robust features from web-scale pretrained backbones and new features learned from source data. The authors investigate whether recent DG finetuning methods are becoming better over time or if improved benchmark results are an artifact of stronger pre-training. They propose the Alignment Hypothesis, which suggests that high DG performance is achieved when image and class label text embeddings are highly aligned. Experiments on DomainBed datasets confirm this hypothesis and reveal that all evaluated DG methods struggle with generalizing beyond pretraining alignment. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Domain Generalization (DG) is a task where models learn to classify images from multiple source domains and perform well on unseen target domains. Recent methods combine strong features from web-scale pretrained backbones with new features learned from source data, leading to improved results. But is this improvement due to better DG finetuning or stronger pre-training? The authors investigate these questions and propose a new idea called the Alignment Hypothesis. They find that high performance requires good alignment of image and class label text embeddings. This research shows how important it is for DG methods to generalize beyond what they learned during training. |
Keywords
» Artificial intelligence » Alignment » Classification » Domain generalization » Pretraining
