Lune

ACL Findings2023

Measuring and Mitigating Local Instability in Deep Neural Networks

Arghya Datta, Subhrangshu Nandi, Jingcheng Xu, Greg Ver Steeg, He Xie, Anoop Kumar, Aram Galstyan

Abstract

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 dataagnostic 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 dataagnostic methods, and comes within 90% of the gold standard, achieved by ensembling, at a fraction of the computational cost.

Ask Lune

Know this paper in minutes, not an afternoon.

Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.

Questions to start from

Lune can draw on

the full paper · related work

Get the answer

Free to start. No credit card required.

Lineage

Builds onCited byLeft to right by year; the more cited, the higher.

Builds on5

  1. What Are Bayesian Neural Network Posteriors Really Like?Pavel Izmailov, Sharad Vikram, Matthew D. Hoffman, Andrew Gordon WilsonICML 2021 · 458 citations
  2. How Good is the Bayes Posterior in Deep Neural Networks Really?Florian Wenzel, Kevin Roth, Bastiaan S. Veeling, Jakub Swiatkowski et al.ICML 2020 · 409 citations
  3. MASSIVE: A 1M-Example Multilingual Natural Language Understanding Dataset with 51 Typologically-Diverse LanguagesJack FitzGerald, Christopher Hench, Charith Peris, Scott Mackie et al.ACL 2023 · 88 citations
  4. Churn Reduction via DistillationHeinrich Jiang, Harikrishna Narasimhan, Dara Bahri, Andrew Cotter et al.ICLR 2022 · 18 citations
  5. Dataset Cartography: Mapping and Diagnosing Datasets with Training DynamicsSwabha Swayamdipta, Roy Schwartz, Nicholas Lourie, Yizhong Wang et al.EMNLP 2020 · 12 citations

Related papers

  1. Lightweight Approaches to DNN Regression Error Reduction: An Uncertainty Alignment PerspectiveZenan Li, Maorun Zhang, Jingwei Xu, Yuan Yao et al.ICSE 2023 · 3 citations
  2. Regression Bugs Are In Your Model! Measuring, Reducing and Analyzing Regressions In NLP Model UpdatesYuqing Xie, Yi-An Lai, Yuanjun Xiong, Yi Zhang et al.ACL 2021
  3. Inconsistency, Instability, and Generalization Gap of Deep Neural Network TrainingRie Johnson, Tong ZhangNeurIPS 2023 · 11 citations
  4. Noise Stability of Transformer ModelsThemistoklis Haris, Zihan Zhang, Yuichi YoshidaICLR 2026
  5. Probing Linguistic SystematicityEmily Goodwin, Koustuv Sinha, Timothy J. O'DonnellACL 2020 · 4 citations
Dusk over the sea between two cliffs drawn in fine vertical lines