Shortcut Learning in Vision Models

A learning algorithm rewards features that predict labels in its training data. It does not know which feature a human intended it to use. If all images of one category were taken with one camera or against one backdrop, that source cue may be easier to learn than object structure. The Nature Machine Intelligence perspective by Geirhos and colleagues calls this shortcut learning: a model solves the benchmark through cues that fail when conditions change. The cue can be subtle, and a good score on a random split from the source may not reveal it. Shortcuts differ from ordinary useful context by whether the correlation is expected to hold in the target task. Snow can legitimately help a classifier in some environmental study, but it is unreliable evidence of an animal species. A hospital-specific marker may predict a diagnosis in a dataset because of referral patterns without being a clinical sign. The same concern can arise from compression artifacts, borders, text overlays or systematic annotation practices. Training on more of the same source may reinforce rather than remove the dependence. Investigation begins with a hypothesis about the suspect cue. Keep the object while changing the background, or keep the background while changing the object. Test different acquisition sources, sites and time periods. Inspect failures, not only averages. Saliency maps can suggest where a model looks, but they do not alone prove which feature caused the prediction; controlled changes provide stronger evidence. Group related images so one scene does not leak across train and test. Mitigation may require collecting counterexamples, balancing contexts, removing leakage, changing the objective or building a model that uses a more stable feature. None guarantees immunity to new shortcuts. Document the intended concept and the operating conditions; then validate after a source change. In consequential uses, a high benchmark result should trigger scrutiny of what was learned rather than an assumption that the model understands the scene.