Reproduce “On the Foundations of Shortcut 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

Katherine L. Hermann, Hossein Mobahi, Thomas Fel, Michael C. Mozer

Researchers at Google DeepMind and Google Research demonstrate that deep neural networks exhibit "shortcut bias" by prioritizing features based on their "availability" – how easily they can be extracted – in addition to their predictivity. Their work, using synthetic and naturalistic datasets alongside theoretical analysis, shows this bias is amplified by network depth and nonlinearity, leading to reliance on less robust features.

Key insights

  • "Availability"—how easily a feature can be extracted and leveraged by a model—is a distinct and critical factor influencing shortcut learning, often overriding a feature's statistical predictivity.
  • Deep nonlinear architectures, particularly with increased depth and nonlinear activations (like ReLU), inherently amplify shortcut bias towards more available features.
  • Linear networks are theoretically unbiased with respect to feature availability, while nonlinear networks inevitably exhibit this bias due to their architectural properties.

Summary by alphaXiv.

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