DISCO: Mitigating Bias in Deep Learning with Conditional Distance Correlation

TL;DR

This paper introduces DISCO methods that leverage causal theory to effectively mitigate dataset bias in deep learning models, improving robustness and scalability across diverse datasets.

Contribution

It presents the SAM causal framework and scalable estimators DISCO$_m$ and sDISCO for bias mitigation, bridging causal theory and deep learning practice.

Findings

DISCO methods outperform existing bias mitigation techniques.

The methods require fewer hyperparameters.

They scale effectively to multi-bias scenarios.

Abstract

Dataset bias often leads deep learning models to exploit spurious correlations instead of task-relevant signals. We introduce the Standard Anti-Causal Model (SAM), a unifying causal framework that characterizes bias mechanisms and yields a conditional independence criterion for causal stability. Building on this theory, we propose DISCO$_m$ and sDISCO, efficient and scalable estimators of conditional distance correlation that enable independence regularization in black-box models. Across five diverse datasets, our methods consistently outperform or are competitive in existing bias mitigation approaches, while requiring fewer hyperparameters and scaling seamlessly to multi-bias scenarios. This work bridges causal theory and practical deep learning, providing both a principled foundation and effective tools for robust prediction. Source Code: https://github.com/***.