Research map / Label-efficient and robust learning / Recognition and long-tail
On the Foundations of Shortcut Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds a generative framework with latent predictive and available features to show that nonlinear (ReLU/Tanh) networks exhibit shortcut bias toward more available features, supported by NTK theory and naturalistic image experiments.
Problem
Models' feature reliance depends on both predictivity and availability of features, but prior work mostly studied predictivity alone; it is unclear which input properties make features more available and how the two interact to produce shortcut learning.
Approach
Introduces availability (amplification and nesting) as a factor separate from predictivity, uses a controllable generative framework with latent features to quantify shortcut bias relative to an LDA optimal classifier, and gives an NTK-based account showing ReLU networks are biased toward more available features while linear networks are not.
Key results
- Linear models are relatively unbiased while adding a single hidden layer with ReLU or Tanh induces shortcut bias; e.g., a model reaches shortcut bias 0.53 with alpha_s/alpha_c=64 and rho_s=0.85.
- Theory: for a ReLU network the first non-zero sensitivity difference appears at order m=9 and favors the more-available feature when predictivities are equal; linear networks show zero bias.
- In Waterbirds, ResNet18 is more sensitive to the background than a Bayes-optimal classifier; reducing background availability (e.g., smaller background extent, grayscale) increases bird accuracy on incongruent probes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase shortcut-feature availability relative to the core feature | MLP on synthetic vector classification · Synthetic vector classification | shortcut bias | Helped |
| Increase shortcut-feature availability relative to the core feature | ResNet-18 on synthetic images · Synthetic image classification | shortcut bias | Helped |
| Increase MLP depth | MLP on synthetic vector classification · Synthetic vector classification | shortcut bias | Helped |
| Use nonlinear hidden activations instead of linear activations (ablation) | MLP on synthetic vector classification · Synthetic vector classification | shortcut bias | Helped |
| Use a linear network rather than a nonlinear network (ablation) | Synthetic vector classification · Synthetic vector classification | shortcut bias | No clear effect |
| Use a ReLU network rather than a linear network (ablation) | Synthetic vector classification · Synthetic vector classification | shortcut bias | Helped |
| Make the shortcut feature more predictive | ResNet-18 on Waterbirds · Waterbirds congruent probes | Bird accuracy | Helped |
| Make the non-core feature more predictive | ResNet-18 on Waterbirds · Waterbirds incongruent probes | Bird accuracy | Hurt |
| Make the non-core feature more predictive | ResNet-18 on CelebA · CelebA probes | Attractive accuracy | Mixed |
| Reduce background availability by increasing bird size, removing background patches, or changing background color (ablation) | ResNet-18 on Waterbirds · Waterbirds incongruent probes | Bird accuracy | Helped |
| Make the shortcut feature shallower than the core feature (ablation) | MLP on synthetic vector classification · Synthetic vector classification | shortcut bias | Helped |
| Increase shortcut-feature pixel footprint | ResNet-18 on synthetic images · Synthetic image classification | shortcut bias | Helped |
Surprising findings
- Availability can override predictivity: models can prefer a less predictive but more available shortcut feature over a more predictive core feature.
- Pretrained ImageNet ResNet50 models are also sensitive to background availability manipulations.
Limitations
- Theory relies on a quadratic approximation of the ReLU kernel and small-covariance asymptotics.
- Synthetic framework is minimal and only two latent features; findings on naturalistic data are correlational with manipulations.
- Availability is defined through hypothesized factors (amplification, nesting) rather than measured directly.
Experimental setup
Datasets: Waterbirds CelebA CUB-200 Places ImageNet
Benchmarks: Waterbirds CelebA
Metrics: shortcut bias reliance accuracy on congruent and incongruent probes
Baselines: LDA (Bayes-optimal classifier) linear MLP
Builds on: Neural Tangent Kernel Waterbirds CelebA
Closest papers by meaning
- Improved Training Technique for Shortcut Models (NeurIPS 2025 · similarity 0.88)
- Do ImageNet-trained Models Learn Shortcuts? The Impact of Frequency Shortcuts on Generalization (CVPR 2025 · similarity 0.88)
- Navigate Beyond Shortcuts: Debiased Learning Through the Lens of Neural Collapse (CVPR 2024 · similarity 0.86)
- Understanding Visual Feature Reliance through the Lens of Complexity (NeurIPS 2024 · similarity 0.86)
- Shortcut Features as Top Eigenfunctions of NTK: A Linear Neural Network Case and More (NeurIPS 2025 · similarity 0.85)
- Intriguing Properties of Generative Classifiers (ICLR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.