# Measuring and Mitigating Local Instability in Deep Neural Networks

arXiv: https://arxiv.org/abs/2305.10625 · HTML: https://mlsanity.com/paper/2305.10625/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Builds on 22 of its 30 references that are in the map.
- First posted on arXiv 18 May 2023.

## Builds on (most cited first, 22 of 22)

- [Adam: A Method for Stochastic Optimization](https://mlsanity.com/paper/1412.6980/) (2014)
- [Rethinking the Inception Architecture for Computer Vision](https://mlsanity.com/paper/1512.00567/) (2015)
- [GLUE: A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding](https://mlsanity.com/paper/1804.07461/) (2018)
- [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://mlsanity.com/paper/1910.01108/) (2019)
- [Averaging Weights Leads to Wider Optima and Better Generalization](https://mlsanity.com/paper/1803.05407/) (2018)
- [When Does Label Smoothing Help?](https://mlsanity.com/paper/1906.02629/) (2019)
- [A Simple Baseline for Bayesian Uncertainty in Deep Learning](https://mlsanity.com/paper/1902.02476/) (2019)
- [Born Again Neural Networks](https://mlsanity.com/paper/1805.04770/) (2018)
- [Underspecification Presents Challenges for Credibility in Modern Machine Learning](https://mlsanity.com/paper/2011.03395/) (2020)
- [Deep Ensembles: A Loss Landscape Perspective](https://mlsanity.com/paper/1912.02757/) (2019)
- [Dataset Cartography: Mapping and Diagnosing Datasets with Training Dynamics](https://mlsanity.com/paper/2009.10795/) (2020)
- [An Evaluation Dataset for Intent Classification and Out-of-Scope Prediction](https://mlsanity.com/paper/1909.02027/) (2019)
- [Fine-Tuning Pretrained Language Models: Weight Initializations, Data Orders, and Early Stopping](https://mlsanity.com/paper/2002.06305/) (2020)
- [What Are Bayesian Neural Network Posteriors Really Like?](https://mlsanity.com/paper/2104.14421/) (2021)
- [Show Your Work: Improved Reporting of Experimental Results](https://mlsanity.com/paper/1909.03004/) (2019)
- [MASSIVE: A 1M-Example Multilingual Natural Language Understanding Dataset with 51 Typologically-Diverse Languages](https://mlsanity.com/paper/2204.08582/) (2022)
- [Optimization with Non-Differentiable Constraints with Applications to Fairness, Recall, Churn, and Other Goals](https://mlsanity.com/paper/1809.04198/) (2018)
- [DialoGLUE: A Natural Language Understanding Benchmark for Task-Oriented Dialogue](https://mlsanity.com/paper/2009.13570/) (2020)
- [Uncertainty Baselines: Benchmarks for Uncertainty & Robustness in Deep Learning](https://mlsanity.com/paper/2106.04015/) (2021)
- [On the Reproducibility of Neural Network Predictions](https://mlsanity.com/paper/2102.03349/) (2021)
- [Churn Reduction via Distillation](https://mlsanity.com/paper/2106.02654/) (2021)
- [Locally Adaptive Label Smoothing for Predictive Churn](https://mlsanity.com/paper/2102.05140/) (2021)

## Built on by (most cited first, 1 of 1)

- [The kernel-balanced equation for deep neural networks](https://mlsanity.com/paper/2309.07367/) (2023)

## Abstract (arXiv)

Deep Neural Networks (DNNs) are becoming integral components of real world services relied upon by millions of users. Unfortunately, architects of these systems can find it difficult to ensure reliable performance as irrelevant details like random initialization can unexpectedly change the outputs of a trained system with potentially disastrous consequences. We formulate the model stability problem by studying how the predictions of a model change, even when it is retrained on the same data, as a consequence of stochasticity in the training process. For Natural Language Understanding (NLU) tasks, we find instability in predictions for a significant fraction of queries. We formulate principled metrics, like per-sample ``label entropy'' across training runs or within a single training run, to quantify this phenomenon. Intriguingly, we find that unstable predictions do not appear at random, but rather appear to be clustered in data-specific ways. We study data-agnostic regularization methods to improve stability and propose new data-centric methods that exploit our local stability estimates. We find that our localized data-specific mitigation strategy dramatically outperforms data-agnostic methods, and comes within 90% of the gold standard, achieved by ensembling, at a fraction of the computational cost