An Investigation of Why Overparameterization Exacerbates Spurious Correlations

Abstract

We study why overparameterization -- increasing model size well beyond the point of zero training error -- can hurt test error on minority groups despite improving average test error when there are spurious correlations in the data. Through simulations and experiments on two image datasets, we identify two key properties of the training data that drive this behavior: the proportions of majority versus minority groups, and the signal-to-noise ratio of the spurious correlations. We then analyze a linear setting and theoretically show how the inductive bias of models towards "memorizing" fewer examples can cause overparameterization to hurt. Our analysis leads to a counterintuitive approach of subsampling the majority group, which empirically achieves low minority error in the overparameterized regime, even though the standard approach of upweighting the minority fails. Overall, our results suggest a tension between using overparameterized models versus using all the training data for achieving low worst-group error.

References in corpus (10)

• Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift
• A systematic study of the class imbalance problem in convolutional neural networks
• Equality of Opportunity in Supervised Learning
• Reconciling modern machine learning practice and the bias-variance trade-off
• Benign Overfitting in Linear Regression
• Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalization
• The generalization error of random features regression: Precise asymptotics and double descent curve
• Natural Adversarial Examples
• Does Distributionally Robust Supervised Learning Give Robust Classifiers?
• Rethinking Bias-Variance Trade-off for Generalization of Neural Networks

Cited by in corpus (31)

• On the Opportunities and Risks of Foundation Models
• WILDS: A Benchmark of in-the-Wild Distribution Shifts
• The Risks of Invariant Risk Minimization
• Understanding the Failure Modes of Out-of-Distribution Generalization
• Towards out of distribution generalization for problems in mechanics
• Do We Actually Need Dense Over-Parameterization? In-Time Over-Parameterization in Sparse Training
• Simple data balancing achieves competitive worst-group-accuracy
• Label-Imbalanced and Group-Sensitive Classification under Overparameterization
• False Sense of Security: Leveraging XAI to Analyze the Reasoning and True Performance of Context-less DGA Classifiers
• Causally motivated Shortcut Removal Using Auxiliary Labels
• Identifying Spurious Correlations for Robust Text Classification
• DORO: Distributional and Outlier Robust Optimization
• Implicit Regularization via Neural Feature Alignment
• Self-training Avoids Using Spurious Features Under Domain Shift
• Robustness in Compressed Neural Networks for Object Detection
• Are Larger Pretrained Language Models Uniformly Better? Comparing Performance at the Instance Level
• Focus on the Common Good: Group Distributional Robustness Follows
• Distinguishing rule- and exemplar-based generalization in learning systems
• Learning Stable Classifiers by Transferring Unstable Features
• Adapting Language Models for Zero-shot Learning by Meta-tuning on Dataset and Prompt Collections
• Towards Robust and Adaptive Motion Forecasting: A Causal Representation Perspective
• Optimal Protocols for Continual Learning via Statistical Physics and Control Theory
• Robust Representation Learning via Perceptual Similarity Metrics
• Out-of-Distribution Generalization in Kernel Regression
• Enhancing Model Robustness and Fairness with Causality: A Regularization Approach
• Unsupervised Learning of Debiased Representations with Pseudo-Attributes
• On complementing end-to-end human behavior predictors with planning
• Human-Imitating Metrics for Training and Evaluating Privacy Preserving Emotion Recognition Models Using Sociolinguistic Knowledge
• Distributionally Robust Learning with Stable Adversarial Training
• Combating Unknown Bias with Effective Bias-Conflicting Scoring and Gradient Alignment
• Boosted CVaR Classification