# A Survey of Uncertainty in Deep Neural Networks

**Authors:** Jakob Gawlikowski, Cedrique Rovile Njieutcheu Tassi, Mohsin Ali, Jongseok Lee, Matthias Humt, Jianxiang Feng, Anna Kruspe, Rudolph Triebel, Peter Jung, Ribana Roscher, Muhammad Shahzad, Wen Yang, Richard Bamler, Xiao Xiang Zhu  
**arXiv:** [2107.03342](https://arxiv.org/abs/2107.03342)  
**Date:** 2021-07-07  
**Summary page:** https://www.summarizepaper.com/en/paper/2107.03342v1/

## Key points

- Importance of uncertainty estimation in neural network predictions
- Issues with over or under confidence in basic neural networks
- Active research on understanding and quantifying uncertainty in neural network predictions
- Identification of various types and sources of uncertainty
- Development of different approaches to measure and quantify uncertainty in neural networks
- Separation of uncertainties into reducible model uncertainty and not reducible data uncertainty
- Modeling techniques using deterministic neural networks, Bayesian neural networks, ensemble methods, and test-time data augmentation approaches
- Need for a standardized evaluation protocol for comparing different methods against established benchmarks on real-world datasets
- Systematic comparison of existing baseline methods for uncertainty estimation to determine the best approach for specific applications
- Challenge of lack of uncertainty ground truths in validating existing methods, calling for further investigation into data generation process and sources of uncertainty
- Emphasis on explainability and physical models as crucial components in obtaining reliable uncertainty estimates

## Summary

The importance of uncertainty estimation in neural network predictions has been increasingly recognized due to their widespread use. Basic neural networks lack the ability to provide certainty estimates, leading to issues of over or under confidence. As a result, researchers have been actively working on understanding and quantifying uncertainty in neural network predictions. Various types and sources of uncertainty have been identified, leading to the development of different approaches to measure and quantify uncertainty in neural networks. This comprehensive overview delves into the various aspects of uncertainty estimation in neural networks, highlighting recent advances, current challenges, and potential research opportunities in the field. The separation of uncertainties into reducible model uncertainty and not reducible data uncertainty is discussed, along with modeling techniques using deterministic neural networks, Bayesian neural networks, ensemble methods, and test-time data augmentation approaches. Furthermore, the summary emphasizes the need for a standardized evaluation protocol for comparing different methods against established benchmarks on real-world datasets. A systematic comparison of existing baseline methods for uncertainty estimation is deemed necessary to determine which approach performs best for specific applications. The lack of uncertainty ground truths poses a challenge in validating existing methods, calling for further investigation into the data generation process and sources of uncertainty. Additionally, explainability and physical models are highlighted as crucial components in obtaining reliable uncertainty estimates. Various works providing insights into Bayesian neural networks and uncertainty quantification in deep learning are referenced for further exploration. Overall, this refined summary underscores the significance of addressing uncertainties in practical applications through improved evaluation protocols and systematic comparisons among different research domains.