# Operation is the hardest teacher: estimating DNN accuracy looking for mispredictions

arXiv: https://arxiv.org/abs/2102.04287 · HTML: https://mlsanity.com/paper/2102.04287/

## Key facts

- Cited by 8 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 8 Feb 2021.

## Built on by (most cited first, 8 of 8)

- [Reliability Assessment and Safety Arguments for Machine Learning Components in System Assurance](https://mlsanity.com/paper/2112.00646/) (2021)
- [Assessing the Reliability of Deep Learning Classifiers Through Robustness Evaluation and Operational Profiles](https://mlsanity.com/paper/2106.01258/) (2021)
- [Detecting Operational Adversarial Examples for Reliable Deep Learning](https://mlsanity.com/paper/2104.06015/) (2021)
- [Benchmarking Generative AI Models for Deep Learning Test Input Generation](https://mlsanity.com/paper/2412.17652/) (2024)
- [Reliability Assurance for Deep Neural Network Architectures Against Numerical Defects](https://mlsanity.com/paper/2302.06086/) (2023)
- [Iterative Assessment and Improvement of DNN Operational Accuracy](https://mlsanity.com/paper/2303.01295/) (2023)
- [DeepSample: DNN sampling-based testing for operational accuracy assessment](https://mlsanity.com/paper/2403.19271/) (2024)
- [Enhancing Deployment-Time Predictive Model Robustness for Code Analysis and Optimization](https://mlsanity.com/paper/2501.00298/) (2024)

## Abstract (arXiv)

Deep Neural Networks (DNN) are typically tested for accuracy relying on a set of unlabelled real world data (operational dataset), from which a subset is selected, manually labelled and used as test suite. This subset is required to be small (due to manual labelling cost) yet to faithfully represent the operational context, with the resulting test suite containing roughly the same proportion of examples causing misprediction (i.e., failing test cases) as the operational dataset. However, while testing to estimate accuracy, it is desirable to also learn as much as possible from the failing tests in the operational dataset, since they inform about possible bugs of the DNN. A smart sampling strategy may allow to intentionally include in the test suite many examples causing misprediction, thus providing this way more valuable inputs for DNN improvement while preserving the ability to get trustworthy unbiased estimates. This paper presents a test selection technique (DeepEST) that actively looks for failing test cases in the operational dataset of a DNN, with the goal of assessing the DNN expected accuracy by a small and ''informative'' test suite (namely with a high number of mispredictions) for subsequent DNN improvement. Experiments with five subjects, combining four DNN models and three datasets, are described. The results show that DeepEST provides DNN accuracy estimates with precision close to (and often better than) those of existing sampling-based DNN testing techniques, while detecting from 5 to 30 times more mispredictions, with the same test suite size.