Uncovering Memorization Effect in the Presence of Spurious Correlations

TL;DR

This paper investigates how neural networks memorize spurious correlations, especially in minority groups, and demonstrates that removing such memorization can improve model fairness and robustness.

Contribution

It provides the first evidence linking memorization of spurious features in specific neurons to imbalanced group performance and proposes a framework to mitigate this during training.

Findings

Spurious features are stored in a small subset of neurons.

Memorization of minority group information correlates with imbalanced performance.

Removing spurious memorization improves minority group accuracy.

Abstract

Machine learning models often rely on simple spurious features -- patterns in training data that correlate with targets but are not causally related to them, like image backgrounds in foreground classification. This reliance typically leads to imbalanced test performance across minority and majority groups. In this work, we take a closer look at the fundamental cause of such imbalanced performance through the lens of memorization, which refers to the ability to predict accurately on atypical examples (minority groups) in the training set but failing in achieving the same accuracy in the testing set. This paper systematically shows the ubiquitous existence of spurious features in a small set of neurons within the network, providing the first-ever evidence that memorization may contribute to imbalanced group performance. Through three experimental sources of converging empirical evidence, we find the property of a small subset of neurons or channels in memorizing minority group information. Inspired by these findings, we hypothesize that spurious memorization, concentrated within a small subset of neurons, plays a key role in driving imbalanced group performance. To further substantiate this hypothesis, we show that eliminating these unnecessary spurious memorization patterns via a novel framework during training can significantly affect the model performance on minority groups. Our experimental results across various architectures and benchmarks offer new insights on how neural networks encode core and spurious knowledge, laying the groundwork for future research in demystifying robustness to spurious correlation.