[Paper Review] An Investigation of Why Overparameterization Exacerbates Spurious Correlations

Shiori Sagawa, Aditi Raghunathan+2 more|arXiv (Cornell University)|May 9, 2020

Reinforcement Learning in Robotics36 references83 citations

TL;DR

The paper shows that increasing model size beyond zero training error can worsen worst-group performance when spurious correlations are present, and that subsampling the majority group can mitigate this effect, sometimes better than upweighting minorities.

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.

Motivation & Objective

• Motivate and quantify how overparameterization affects worst-group error in the presence of spurious correlations.
• Identify data properties that modulate this effect and develop theoretical insights into the memorization bias of overparameterized models.
• Compare training strategies (ERM vs reweighting) in terms of average vs worst-group performance.

Proposed method

• Empirical evaluation on two image datasets (CelebA and Waterbirds) across a range of model sizes.
• Use reweighted empirical risk minimization to upweight minority groups.
• Also implement a synthetic synthetic-data setup with controlled core and spurious features and noisy components.
• Analyze a linear explicit-memorization setting to formalize memorization and inductive bias.
• Train with unregularized logistic regression over random features to study overparameterization effects.
• Vary model capacity by increasing width (CelebA) or number of projections (Waterbirds).

Experimental results

Research questions

• RQ1 When and why does overparameterization degrade worst-group performance in the presence of spurious correlations?
• RQ2 How do data distribution properties (majority/minority proportions and spurious-to-core informativeness) influence this effect?
• RQ3 Can alternative training strategies (e.g., subsampling or upweighting) mitigate worst-group errors in the overparameterized regime?
• RQ4 What theoretical mechanism explains memorization-driven failures, and under what conditions does minimizing memorization improve robustness?

Key findings

• Overparameterization improves average test error but worsens worst-group error on CelebA and Waterbirds under reweighting (worst-group error can exceed 60% with zero training error on overparameterized models).
• An underparameterized model can achieve substantially better worst-group test error (e.g., 25.6% on CelebA and 26.6% on Waterbirds) compared with overparameterized models.
• On synthetic data, the best worst-group error is achieved by underparameterized models (e.g., 28.5%), while highly overparameterized models plateau at around 55% worst-group error.
• Removing the spurious features eliminates the detrimental overparameterization effect, indicating reliance on spurious features drives failures.
• Two data properties amplify the effect: higher majority fraction p_maj and higher spurious-core information ratio r_s:c increase worst-group error in the overparameterized regime; balanced groups and lower r_s:c can reverse the effect.
• A theoretical linear analysis shows a minimum-norm inductive bias in overparameterized models favors memorizing minority points via noise features, leading to high worst-group error; underparameterization avoids this by relying on core features.