# Approaching Test Time Augmentation in the Context of Uncertainty Calibration for Deep Neural Networks

**Authors:** Pedro Conde, Tiago Barros, Rui L. Lopes, Cristiano Premebida, Urbano J. Nunes  
**arXiv:** [2304.05104](https://arxiv.org/abs/2304.05104)  
**Date:** 2023-04-11  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.05104v1/

## Key points

- Deep Neural Networks have led to widespread use of machine learning systems in real-world applications
- Reliable models with accurate predictions and calibrated uncertainties are critical for decision making processes
- Researchers propose a novel technique based on test time augmentation to improve uncertainty calibration of deep models for image classification
- The technique comes in two variations named M-ATTA and V-ATTA and leverages an adaptive weighting system to improve uncertainty calibration without affecting the model's accuracy
- M/V-ATTA improves uncertainty calibration by adapting weights during testing rather than training, unlike other test time augmentation approaches
- The team evaluated the performance of the technique using different metrics of uncertainty calibration on CIFAR-10, CIFAR-100, and Aerial Image Dataset benchmarks
- Empirical results indicate that M/V-ATTA outperforms state-of-the-art calibration techniques while maintaining baseline classification performance
- The code for M/V-ATTA is available on GitHub at https://github.com/pedrormconde/MV-ATTA

## Summary

The rise of Deep Neural Networks has led to the widespread use of machine learning systems in various real-world applications, highlighting the need for highly reliable models. To address this issue, a team of researchers have proposed a novel technique based on test time augmentation to improve the uncertainty calibration of deep models for image classification. The technique comes in two variations named M-ATTA and V-ATTA and leverages an adaptive weighting system to improve uncertainty calibration without affecting the model's accuracy. Unlike other test time augmentation approaches, M/V-ATTA improves uncertainty calibration by adapting weights during testing rather than training. The team evaluated the performance of the technique using different metrics of uncertainty calibration on CIFAR-10, CIFAR-100 and Aerial Image Dataset benchmarks. Empirical results indicate that M/V-ATTA outperforms state-of-the-art calibration techniques while maintaining baseline classification performance. The code for M/V-ATTA is available on GitHub at https://github.com/pedrormconde/MV-ATTA. The proposed approach can be useful in improving the reliability of machine learning systems used in real-world applications where accurate predictions with calibrated uncertainties are critical for decision making processes.