Reproduce “Underspecification Presents Challenges for Credibility in Modern Machine Learning” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Alexander D'Amour, Katherine Heller, Dan Moldovan, Ben Adlam, Babak Alipanahi, Alex Beutel, Christina Chen, Jonathan Deaton and 32 more

The authors define underspecification as a training pipeline’s ability to produce predictors with similar held-out performance but different behavior under application-specific stress tests. They compare 50 ResNet-50 models trained with different random seeds: mean accuracy was 0.759 (standard deviation 0.001) on clean ImageNet and 0.197 (0.024) on level-5 pixelation. This evidence is specific to this model, task, metric, and condition.

Key insights

  • Underspecification is a prevalent issue in modern, production-grade deep learning systems, affecting domains from computer vision to clinical risk prediction.
  • Models demonstrating indistinguishable in-distribution performance can possess significantly different inductive biases, leading to highly variable and often undesirable behaviors under deployment conditions.
  • Arbitrary choices within the ML pipeline, such as random seeds, can determine the specific inductive biases learned by a model, making its real-world generalization properties unpredictable if only i.i.d. performance is considered.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.