Reactive Model Correction: Mitigating Harm to Task-Relevant Features via Conditional Bias Suppression

Neural networks can learn harmful shortcuts: Deep learning models often rely on spurious correlations in training data rather than legitimate patterns, which can have serious consequences in high-risk applications like medical diagnosis.

Previous bias-correction methods create new problems: While existing post-hoc methods (applied after training without retraining) can efficiently suppress reliance on harmful features, they tend to damage overall model performance by indiscriminately altering how the model processes information.

A "reactive" approach preserves performance while removing bias: The paper's proposed R-ClArC method uses explainable AI insights to selectively correct only the specific cases where the model relies on spurious features, rather than making blanket changes that harm accuracy across the board.

The method was validated on both controlled and real-world tests: Experiments using the FunnyBirds dataset (controlled setting) and ISIC2019 skin cancer images (real-world application) demonstrated that the reactive approach successfully minimized reliance on spurious features while maintaining model performance.

This offers a practical path forward for high-risk AI applications: By combining post-hoc correction with explainability tools, the approach provides a way to deploy safer AI systems in critical domains without sacrificing the accuracy that users depend on.