Reproduce “Shortcut Learning in Deep Neural Networks” on OpenResearch

About this paper

Robert Geirhos, Jörn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, Felix A. Wichmann

Deep neural networks often learn superficial decision rules, termed "shortcut learning," which perform well on standard benchmarks but fail in more challenging, real-world scenarios. This work unifies various observed limitations, from adversarial examples to poor out-of-distribution generalization, by framing them as manifestations of models exploiting unintended statistical regularities in training data, analogous to phenomena in biological learning.

Key insights

• Shortcut learning commonly arises from "dataset shortcut opportunities" (spurious correlations in data) and the inherent tendency of discriminative learning to find the "easiest" sufficient features.
• The entire inductive bias of a model, encompassing its architecture, training data, loss function, and optimization method, collectively determines which shortcut solutions are learned.
• A critical shift is needed in model evaluation from solely relying on independent and identically distributed (i.i.d.) test sets to including out-of-distribution (OOD) generalization tests as standard practice.

Summary by alphaXiv.