Shortcut Learning in Deep Neural Networks.¶
Geirhos, R., Jacobsen, J., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., & Wichmann, F. A. (2020). Shortcut Learning in Deep Neural Networks. Nature Machine Intelligence, 2, 665-673.
Cited by¶
10 citations across 10 artifacts.
Each citation links to the sentence it supports in the citing article.
Primes¶
- Cue Outcome Decoupling
- In machine learning under distribution shift, a classifier latches onto a scanner watermark correlated with disease at the training site but not the new site — shortcut learning and spurious correlation.
This sourceDefines shortcut learning: models latch onto spurious correlations that fail under distribution shift, with better-fit shortcuts generalizing worse.
SupportedVerified against the work's full text
Geirhos et al. describe a pneumonia classifier that keyed on a hospital-specific metal token and failed on scans from new hospitals, as a case of shortcut learning.
“a machine classifier successfully detected pneumonia from X-ray scans of a number of hospitals, but its performance was surprisingly low for scans from novel hospitals: The model had unexpectedly learned to identify particular hospital systems with near-perfect accuracy (e.g. by detecting a hospital-specific metal token on the scan”
- In machine learning under distribution shift, a classifier latches onto a scanner watermark correlated with disease at the training site but not the new site — shortcut learning and spurious correlation.
- Feature Engineering
- Running the other way, education psychometrics' long experience with construct validity — does this feature actually measure the latent thing it claims to? — ports into machine learning, where its absence now appears under names like spurious correlation and shortcut learning.
This sourceNames spurious-correlation / shortcut learning — the ML manifestation of features lacking construct validity — supporting the construct-validity transfer into ML.
Supported in partVerified against the work's full text
Geirhos et al. define shortcut learning in ML and draw a parallel to surface learning in Education, but never mention construct validity or psychometrics.
“Shortcuts are decision rules that perform well on standard benchmarks but fail to transfer to more challenging testing conditions, such as real-world scenarios.”
- Running the other way, education psychometrics' long experience with construct validity — does this feature actually measure the latent thing it claims to? — ports into machine learning, where its absence now appears under names like spurious correlation and shortcut learning.
- Proxy-Target Divergence
- The same pattern recurs in a hedge whose historical spread relationships decouple from default risk during a crisis (stress decoupling), a sea-turtle hatchery where a brightest-horizon cue decouples from ocean direction under light pollution (environmental decoupling), and a pneumonia classifier whose scanner artefact decouples from pneumonia signal under deployment shift (distribution-shift decoupling).
This sourceShows classifiers latching onto spurious features (e.g., scanner artifacts) that decouple from the intended signal under distribution shift.
Supported in partVerified against the work's full text
Supports only the pneumonia-classifier example: a hospital-specific cue on the scan stood in for pneumonia and failed at new hospitals; the hedge and sea-turtle examples are not in this paper.
“a machine classifier successfully detected pneumonia from X-ray scans of a number of hospitals, but its performance was surprisingly low for scans from novel hospitals”
- The same pattern recurs in a hedge whose historical spread relationships decouple from default risk during a crisis (stress decoupling), a sea-turtle hatchery where a brightest-horizon cue decouples from ocean direction under light pollution (environmental decoupling), and a pneumonia classifier whose scanner artefact decouples from pneumonia signal under deployment shift (distribution-shift decoupling).
- Relevance Substitution
- In machine learning, shortcut learning substitutes spurious features (image background, scanner artefact, watermark) for the task-defining features, the model's inference running through the irrelevant channel.
This sourceNames and surveys shortcut learning, where models rely on spurious features rather than task-defining ones.
SupportedVerified against the work's full text
Geirhos et al. define shortcut learning with grass-background and hospital scan-token examples, where models use unintended cues instead of intended features.
“The model had unexpectedly learned to identify particular hospital systems with near-perfect accuracy (e.g. by detecting a hospital-specific metal token on the scan, see Figure 1).”
- In machine learning, shortcut learning substitutes spurious features (image background, scanner artefact, watermark) for the task-defining features, the model's inference running through the irrelevant channel.
- Shortcut Learning
- The general claim is compact: optimization under a sufficient statistic finds the cheapest sufficient statistic, not the target.
This sourceCanonical statement of shortcut learning: optimizers seize cheap predictive features (the 'cheapest sufficient statistic') that succeed in-distribution and fail under distribution shift; surveys ML cases and stress-set/OOD diagnostics.
Supported in partVerified against the work's full text
Geirhos et al. support the informal gist (networks settle on easy rules that pass in-distribution tests but are not the intended solution) but not the 'sufficient statistic' formulation or its generality.
“DNNs always learn the easiest possible solution to a problem”
- The general claim is compact: optimization under a sufficient statistic finds the cheapest sufficient statistic, not the target.
Mechanisms¶
- Artifact Red-Team Review
- Behavioral Feature Refresh Cycle
- … visible tricks — or the wrong "invariant" — is shortcut learning: the detector latches onto a signal that merely correlated with abuse in the training window (a particular hosting provider, a time-of-day) and collapses the moment that correlate shifts or, worse, penalizes a legitimate population that shares it.
This sourceDefines shortcut learning as reliance on decision rules that perform in familiar settings but fail when spurious training correlations shift.
- … visible tricks — or the wrong "invariant" — is shortcut learning: the detector latches onto a signal that merely correlated with abuse in the training window (a particular hosting provider, a time-of-day) and collapses the moment that correlate shifts or, worse, penalizes a legitimate population that shares it.
- Convolutional Feature Extractor
- Its failure mode is that a learned kernel will happily latch onto whatever correlates with the label in the training data rather than what actually causes it — shortcut learning.
This sourceDefines shortcut learning as a model's reliance on unintended decision rules that perform on standard data but fail under changed conditions.
- Its failure mode is that a learned kernel will happily latch onto whatever correlates with the label in the training data rather than what actually causes it — shortcut learning.
- Shortcut Probe Holdout Set
- The failure has a name: shortcut learning, the modern face of the Clever Hans effect.
This sourceDocuments shortcut learning; the modern face of the Clever Hans effect.
- The failure has a name: shortcut learning, the modern face of the Clever Hans effect.
- Supervised Representation Learning
- Its failure mode is shortcut learning: the model latches onto a nuisance feature that happens to correlate with the label in the training set, so its output looks clean and then collapses under distribution shift.
This sourceDefines shortcut learning as reliance on unintended training-set correlations that can fail under distribution shift.
- Its failure mode is shortcut learning: the model latches onto a nuisance feature that happens to correlate with the label in the training set, so its output looks clean and then collapses under distribution shift.
Verification¶
Does it exist? Confirmed. This work's DOI resolves to a registered record, which fixes its identity. That is all it fixes.
Does it back the claim? Read against the text for 5 of 10 citations: 2 supported, 3 supported in part. Each verdict is shown under its citation below, with what in the work backs the sentence.
Support is checked per citation rather than per work — the same source can be cited soundly in one article and wrongly in another. Per-citation recording began recently, so a citation with no recorded check is a gap in the record rather than evidence it went unchecked.
See how references were verified.
Registry ID ref:ce3d33590cf8 · see in the full table