# Addressing Randomness in Evaluation Protocols for Out-of-Distribution Detection

**Authors:** Konstantin Kirchheim, Tim Gonschorek, Frank Ortmeier  
**arXiv:** [2203.00382](https://arxiv.org/abs/2203.00382)  
**Date:** 2022-03-01  
**Summary page:** https://www.summarizepaper.com/en/paper/2203.00382v1/

## Key points

- Deep neural networks are popular for classification tasks but can behave unpredictably with inputs outside of the training distribution
- Out-of-distribution (OOD) detection mechanisms are needed to address this issue
- Evaluating OOD detection methods on unseen data is challenging due to lack of prior information on OOD data
- Contemporary evaluation protocols rely on open set simulations, which involve splitting a dataset into in- and out-of-distribution samples and averaging performance over several synthetic random splits
- Open set simulation frameworks include several sources of randomness that can significantly impact results
- Too few simulations may fail to provide reliable estimates of expected performance as outcomes largely depend on chance
- The authors propose treating OOD method evaluation in open set simulations as a fundamentally probabilistic process and estimating expected performance using a Monte Carlo approach to draw more reliable conclusions
- Hypothesis tests demonstrated even a considerable number of simulations was insufficient to infer statistically significant performance differences between compared methods
- Future work should investigate other OOD evaluation protocols not based on open set simulations and include different types of data such as natural language, sound, or video.
- Studying individual sources of randomness might help quantify their contribution to fluctuations and enable better experiment design.
- This study aimed to demonstrate the brittleness of current best practice evaluation protocols rather than evaluate specific methods. Different experimental setups or hyperparameters could lead to different conclusions regarding method performance; however, these experiments might also be subject to inherent randomness and should address it accordingly.
- Overall, this study highlights the importance of addressing randomness in evaluating OOD detection methods and proposes a new approach to estimate expected performance more reliably by treating OOD method evaluation in open set simulations as a probabilistic process through Monte Carlo approaches for drawing more reliable conclusions about method performances while accounting for different sources of randomness when designing experiments.

## Summary

The use of deep neural networks for classification tasks has become increasingly popular, but these models can behave unpredictably when presented with inputs that do not come from the training distribution. This has led to a need for out-of-distribution (OOD) detection mechanisms. However, evaluating the performance of these methods on unseen data is challenging due to the lack of prior information on OOD data. Contemporary evaluation protocols often rely on open set simulations, which involve splitting a dataset into in- and out-of-distribution samples and averaging the performance over several synthetic random splits. In this work, the authors studied the effects of randomness in open set simulation frameworks and found that they include several sources of randomness that can significantly impact results. They conducted a large-scale study involving three orders of magnitude more simulations than recent publications to create a pool of experimental outcomes. The authors found that too few simulations may fail to provide reliable estimates of expected performance, as outcomes largely depend on chance. To address this issue, the authors proposed treating OOD method evaluation in open set simulations as a fundamentally probabilistic process and estimating expected performance using a Monte Carlo approach to draw more reliable conclusions. The authors conducted hypothesis tests that demonstrated even a considerable number of simulations was insufficient to infer statistically significant performance differences between compared methods. The authors suggest future work should investigate other OOD evaluation protocols not based on open set simulations and include different types of data such as natural language, sound, or video. Studying individual sources of randomness might help quantify their contribution to fluctuations and enable better experiment design. It's important to note that this work aimed to demonstrate the brittleness of current best practice evaluation protocols rather than evaluate specific methods. Different experimental setups or hyperparameters could lead to different conclusions regarding method performance; however, these experiments might also be subject to inherent randomness and should address it accordingly. Overall, this study highlights the importance of addressing randomness in evaluating OOD detection methods and proposes a new approach to estimate expected performance more reliably by treating OOD method evaluation in open set simulations as a probabilistic process through Monte Carlo approaches for drawing more reliable conclusions about method performances while accounting for different sources of randomness when designing experiments.