# Test-Time Adaptation via Self-Training with Nearest Neighbor Information

**Authors:** Minguk Jang, Sae-Young Chung  
**arXiv:** [2207.10792](https://arxiv.org/abs/2207.10792)  
**Date:** 2022-07-08  
**Summary page:** https://www.summarizepaper.com/en/paper/2207.10792v1/

## Key points

- Adapting trained classifiers using only online test data is a challenging task in machine learning
- Self-training is a popular approach for test-time adaptation, but it has limitations under test-time domain shift
- Test-time Adaptation via Self-Training with nearest neighbor information (TAST) is a novel method proposed by Minguk Jang and Sae-Young Chung to overcome this limitation
- TAST leverages the idea that a test data and its nearest neighbors in an embedding space of the trained classifier are more likely to have the same label
- TAST generates pseudo labels for each new test data point using its nearest neighbors from a set composed of previous test data and then fine-tunes it with these generated pseudo labels
- The authors conducted experiments on two standard benchmarks - domain generalization and image corruption benchmarks - to evaluate TAST's performance against current state-of-the-art methods
- The results showed that TAST outperforms existing methods in both benchmarks
- TAST provides an effective solution for adapting trained classifiers using only online test data by leveraging nearest neighbor information to generate accurate pseudo labels
- This method has shown promising results in improving classification accuracy under domain shift scenarios and can be applied in various real-world applications where access to training or future testing data is limited.

## Summary

In the field of machine learning, adapting trained classifiers using only online test data is a challenging task. One popular approach for test-time adaptation is self-training which fine-tunes the trained classifiers using the classifier predictions of the test data as pseudo labels. However, under the test-time domain shift, self-training methods have a limitation that learning with inaccurate pseudo labels greatly degrades the performance of the adapted classifiers. To overcome this limitation, Minguk Jang and Sae-Young Chung propose a novel test-time adaptation method called Test-time Adaptation via Self-Training with nearest neighbor information (TAST). This method leverages the idea that a test data and its nearest neighbors in an embedding space of the trained classifier are more likely to have the same label. TAST adapts the trained classifier by first generating pseudo labels for each new test data point using its nearest neighbors from a set composed of previous test data and then fine tuning it with these generated pseudo labels. The authors conducted experiments on two standard benchmarks - domain generalization and image corruption benchmarks - to evaluate TAST's performance against current state-of-the-art methods. The results showed that TAST outperforms existing methods in both benchmarks. In conclusion, TAST provides an effective solution for adapting trained classifiers using only online test data by leveraging nearest neighbor information to generate accurate pseudo labels. This method has shown promising results in improving classification accuracy under domain shift scenarios and can be applied in various real world applications where access to training or future testing data is limited.