Shortcut Learning in Deep Neural Networks
Robert Geirhos, Jörn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, Felix A. Wichmann
Nature Machine Intelligence (h
cs.CV, cs.AI, cs.LG, q-bio.NC
2020-04-17
Geirhos et al. unify adversarial examples, cross-dataset failures, and a pneumonia model that flags hospitals not disease as one problem: shortcut learning. The fix is OOD testing over i.i.d. tests.
Deep learning keeps winning, but strangely. Image classifiers beat humans, yet shifting a photo of a car by a few pixels or changing its background breaks the prediction. Captioning models produce plausible captions without looking at the image. A pneumonia detector performed well across hospitals but collapsed at a new one: it had learned to spot a hospital-specific metal token on the scan, not the disease.
This 2020 Nature Machine Intelligence perspective argues these unrelated failures share one cause. The cause is "shortcut learning": the model learns a feature that happens to work rather than the solution the task intends. Shortcuts score well on standard in-distribution (i.i.d.) tests and collapse under out-of-distribution (OOD) conditions.
The paper offers a taxonomy of decision rules, ranked from loose to tight:
The key gap sits between the third and fourth tiers: tested only on i.i.d. data, you cannot tell whether a model uses shape, location, or white-pixel count, since all three score equally on the training set. A toy star-vs-moon classifier shows it: a fully connected network hits 100% on training but learned a location rule, and shuffling object positions drops test accuracy to 51% (chance). A convolutional network, whose architecture suppresses the location shortcut, scores 100% on both.
The paper traces shortcuts to four interlocking components (architecture, training data, loss function, optimizer), jointly the model's "inductive bias." Cross-entropy encourages the model to stop once a simple predictor works, and gradient descent favors simple functions. That is why shortcuts are "easier to learn" than the intended solution.
Shortcuts span domains, not just vision:
| Domain | Shortcut |
| Vision | texture over shape; background as object cue (cow = grass) |
| NLP | BERT infers arguments from the word "not," ignoring content |
| RL | a Tetris agent pauses to avoid losing; Mario wins via time-travel exploits |
| Fairness | Amazon's hiring tool used gender as a predictor, inferring it from women's-college names after names were removed |
Shortcut learning appears in biology too: maze rats apparently navigating by color were sniffing paint odor; rote-learning students ace multiple choice but fail essays. The paper proposes a "Morgan's Canon for machine learning": never attribute to high-level abilities what shortcut learning can explain.
The value is not a new algorithm but a shared mirror. Adversarial robustness, domain adaptation, causal representation, and fairness, previously siloed, turn out to fight the same enemy.
The practical warning: a high i.i.d. score is not proof the model learned what you intended. To check, you need OOD tests such as ImageNet-C (15 corruptions), ObjectNet, or cue-conflict stimuli, closer to the real exam.
As a perspective it synthesizes rather than proposes methods; the argument rests on cited prior work with no systematic experimental comparison. The toy experiment is too simple to quantify how bad shortcuts get on real tasks.
Its OOD prescription is not a permanent fix: good OOD benchmarks age as models improve (the Winograd Schema is the example), and "the intended solution" is often ill-defined. The analogy to animal "unintended cue learning" is suggestive but rhetorical, not a proof.