Research map / Optimization / Implicit bias and generalization
The Pitfalls of Simplicity Bias in Neural Networks
Harshay Shah, Kaustav Tamuly, Aditi Raghunathan, Prateek Jain, Praneeth Netrapalli
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train an FCN on ^LMS-7, where the simplest feature is noisy and less predictive than the 7-slab features | SGD-trained FCNs · ^LMS-7 test | accuracy: 0.940 ± 0.002 to 0.949 ± 0.003 across listed FCNs; text reports 95% test accuracy | Hurt |
| Randomize the simple linear feature | FCN on LMS-5 · LMS-5 | randomized AUC: 1.0 to 0.5 | Hurt |
| Randomize all complex features | FCN on LMS-5 · LMS-5 | randomized AUC: 1.00 | No clear effect |
| Train on MNIST-CIFAR and randomize the MNIST block | MobileNetV2 · MNIST-CIFAR:A | AUC: 1.00 ± 0.00 to 0.53 ± 0.01 | Hurt |
| Randomize the CIFAR10 block | MobileNetV2 · MNIST-CIFAR:A | AUC: 1.00 ± 0.00 to 1.00 ± 0.00 | No clear effect |
| Replace SGD with Adam or RMSProp, or change the activation among ReLU, Leaky ReLU, PReLU, and Tanh (ablation) | (100,2)-FCN · LMS-7 and MS-(5,7) | simple-feature-randomized AUC: approximately 0.50 across tested combinations | No clear effect |
| Add dropout and vary ℓ2 regularization (ablation) | FCNs on LMS-7 · LMS-7 | simple-feature-randomized AUC: approximately 0.5 for all models | No clear effect |
| Use a vanilla ensemble of independently trained models (ablation) | Models on proposed datasets · LMS-5 | reliance on complex features | No clear effect |
| Use adversarial training (ablation) | FCNs on proposed datasets · Synthetic datasets | reliance on complex features | No clear effect |
| Remove the simplest feature from the training data (ablation) | SGD-trained models · LMS-5 and related synthetic data | accuracy: 100% | Improved |
Closest papers by meaning
- Simplicity Bias in Overparameterized Machine Learning (AAAI 2024 · similarity 0.93)
- Simplicity Bias of Two-Layer Networks beyond Linearly Separable Data (ICML 2024 · similarity 0.92)
- Do We Always Need the Simplicity Bias? Looking for Optimal Inductive Biases in the Wild (CVPR 2025 · similarity 0.91)
- The Rich and the Simple: On the Implicit Bias of Adam and SGD (NeurIPS 2025 · similarity 0.89)
- Simplicity Bias via Global Convergence of Sharpness Minimization (ICML 2024 · similarity 0.89)
- Bias of Stochastic Gradient Descent or the Architecture: Disentangling the Effects of Overparameterization of Neural Networks (ICML 2024 · similarity 0.89)
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.
