# Fully Test-time Adaptation by Entropy Minimization

**Authors:** Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, Trevor Darrell  
**arXiv:** [2006.10726](https://arxiv.org/abs/2006.10726)  
**Date:** 2020-06-18  
**Summary page:** https://www.summarizepaper.com/en/paper/2006.10726v1/

## Key points

- Machine learning models are trained on labeled data to make predictions on new, unseen data.
- Models must adapt themselves to accurately classify new samples when faced with different test data distributions.
- A team of researchers proposed an entropy minimization approach for fully test-time adaptation.
- The proposed approach involves taking the model's confidence as an objective measure of its performance and modulating its representation with affine transformations to minimize entropy during testing.
- Experiments were conducted to evaluate the approach's effectiveness in improving robustness to corruptions for image classification on CIFAR-10/100 and ILSVRC datasets, as well as target-only domain adaptation for digit classification on MNIST and SVHN datasets.
- The research presents an innovative solution to the problem of fully test-time adaptation in machine learning models without requiring additional labeled data or fine-tuning steps.

## Summary

In the field of machine learning, models are often trained on labeled data to make predictions on new, unseen data. However, when faced with new and different data during testing, a model must adapt itself to accurately classify the new samples. This is particularly challenging when the test data comes from a different distribution than the training data and there is no labeled test set available for fine-tuning. To address this challenge, a team of researchers including Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell have proposed an entropy minimization approach for fully test-time adaptation. In their paper titled "Fully Test-time Adaptation by Entropy Minimization," they describe a method in which a supervised model confronts unlabeled test data from a different distribution without any help from its labeled training data. The proposed approach involves taking the model's confidence as an objective measure of its performance as measured by the entropy of its predictions. During testing, the model is adapted by modulating its representation with affine transformations to minimize entropy. The researchers conducted experiments to evaluate their approach's effectiveness in improving robustness to corruptions for image classification on CIFAR-10/100 and ILSVRC datasets. They also demonstrated that their method can be used for target-only domain adaptation for digit classification on MNIST and SVHN datasets. Overall, this research presents an innovative solution to the problem of fully test-time adaptation in machine learning models. By minimizing entropy during testing through affine transformations of the model's representation, it shows promise in improving classification accuracy on previously unseen distributions without requiring any additional labeled data or fine-tuning steps.