Summary of Decoupled Prototype Learning For Reliable Test-time Adaptation, by Guowei Wang et al.
Decoupled Prototype Learning for Reliable Test-Time Adaptation
by Guowei Wang, Changxing Ding, Wentao Tan, Mingkui Tan
First submitted to arxiv on: 15 Jan 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 This paper proposes a novel approach to test-time adaptation (TTA), a task that adapts pre-trained models to new domains during inference. Current approaches fine-tune models using cross-entropy loss and pseudo-labels, but are vulnerable to noisy label noise. The authors introduce Decoupled Prototype Learning (DPL), which optimizes class prototypes in a contrastive manner, reducing overfitting to noisy labels. DPL also incorporates memory-based updates and consistency regularization to leverage unconfident pseudo-labels. Experiments show that DPL achieves state-of-the-art performance on domain generalization benchmarks and improves self-training-based methods on image corruption benchmarks. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper helps us learn from machines in new situations. Right now, we can only train models once and then use them for a specific job. But sometimes we need to teach the model to do something new without retraining it completely. This is called test-time adaptation (TTA). The problem is that if there are mistakes in what the machine thinks it sees or hears, these mistakes will get copied into the new training. To fix this, scientists created a new way of training models called Decoupled Prototype Learning (DPL). DPL makes sure the model doesn’t copy mistakes and can still learn from the new data. It’s like creating a backup system to make sure the machine learns correctly. |
Keywords
* Artificial intelligence * Cross entropy * Domain generalization * Inference * Overfitting * Regularization * Self training
