Let Samples Speak: Mitigating Spurious Correlation by Exploiting the Clusterness of Samples

TL;DR

Spurious correlations limit deep models, especially when bias attributes are unknown. The paper proposes NSF, a four-step, bias-label-free debiasing pipeline that identifies bias presence from dispersed samples in feature space, estimates a bias-invariant mean, learns a linear transform $t(\vec{x})$ to neutralize spurious features, and finetunes the classifier on the transformed features. The authors provide theoretical justification under a strong spurious assumption and demonstrate substantial improvements in Worst Group Accuracy (WGA) across image and NLP benchmarks, including medical imaging, with fast training times. NSF is architecture-agnostic, non-intrusive, and practical for real-world deployments where bias labels are unavailable.

Abstract

Deep learning models are known to often learn features that spuriously correlate with the class label during training but are irrelevant to the prediction task. Existing methods typically address this issue by annotating potential spurious attributes, or filtering spurious features based on some empirical assumptions (e.g., simplicity of bias). However, these methods may yield unsatisfactory performance due to the intricate and elusive nature of spurious correlations in real-world data. In this paper, we propose a data-oriented approach to mitigate the spurious correlation in deep learning models. We observe that samples that are influenced by spurious features tend to exhibit a dispersed distribution in the learned feature space. This allows us to identify the presence of spurious features. Subsequently, we obtain a bias-invariant representation by neutralizing the spurious features based on a simple grouping strategy. Then, we learn a feature transformation to eliminate the spurious features by aligning with this bias-invariant representation. Finally, we update the classifier by incorporating the learned feature transformation and obtain an unbiased model. By integrating the aforementioned identifying, neutralizing, eliminating and updating procedures, we build an effective pipeline for mitigating spurious correlation. Experiments on image and NLP debiasing benchmarks show an improvement in worst group accuracy of more than 20% compared to standard empirical risk minimization (ERM). Codes and checkpoints are available at https://github.com/davelee-uestc/nsf_debiasing .

Figures (7)

• Figure 1: 1) Ideally, bias attributes (e.g., color) should be evenly distributed and non-predictive of the class; 2) Sampling bias can introduce unintended patterns, like most circles being red and most stars being purple, causing some features to mistakenly correlate with class labels. Since ERM training minimizes the mean loss, an ERM-trained model is highly likely to fit these spurious correlations due to their large population in the data; 3) Intuitively, a transformation producing invariant representation for different values of bias attributes reduces the possible of learning bias.
• Figure 2: NSF leverages a task-independent strong spurious assumption, enabling us to (1) use the presence of minority sample (purple circles) as the indicator of spurious features; (2) distinguish majority (purple stars) and minority groups (red stars) and estimate a bias-invariant representation using the groups found. NSF mitigates biases by (3) first eliminating the bias attributes by transforming features to align training samples with estimated unbiased mean values of the true data distribution, then (4) debiasing classifiers through fine-tuning.
• Figure 3: A linear transformation t(x) is learned to eliminate the spurious features by shifting them to their unbiased mean and reducing variance so that the correlation between the spurious feature and label is removed.
• Figure 4: (a) Features and the decision boundary using models trained on the waterbirds dataset, using group-balanced sampling for better visualization. 1) Original features $\vec{x}$ allows more biased solutions for the ERM training; 2) Elimination of spurious features using $t(\vec{x})$ leaves a smaller room for biased solutions; 3) Finetuning $h'(\vec{x})$ using $t(\vec{x})$ results in a unbiased classifier. (b) Combinations of classifiers of ERM $h(\vec{x})$ and debiased $h'(\vec{x})$, and the features of raw $\vec{x}$ and transformed $t(\vec{x})$. The debiased classifier $h'(\vec{x})$ performs well using the original features $\vec{x}$ indicating $h'(\vec{x})$ relies on core rather than spurious features.
• Figure 5: The WGA and mAcc of discarding channels by the lowest of the coefficient $w$ in the transformation $t(x)$ are higher than choosing randomly, validating lower $w$ highly correlate with spurious features so that they are eliminated, as in \ref{['fig:tx']}.

Theorems & Definitions (2)

• Theorem 1
• Theorem 2