Reveal to Revise: An Explainable AI Life Cycle for Iterative Bias Correction of Deep Models

AI models can learn misleading patterns from training data: The paper demonstrates that state-of-the-art deep learning models often pick up on spurious correlations in medical datasets (such as skin cancer detection), which can lead to unreliable decisions in high-stakes applications if left uncorrected.

Explainability tools can help identify hidden biases: The "Reveal to Revise" framework uses explainable AI techniques to make model weaknesses visible by examining which features the model relies on, allowing practitioners to spot problematic patterns that might otherwise remain hidden.

Bias correction requires an iterative process: Rather than a one-time fix, the framework shows that effectively removing bias involves multiple cycles of revealing problems, identifying their source in the data, correcting the model, and then re-evaluating—demonstrating that responsible AI deployment is an ongoing process.

Practical bias mitigation is achievable with minimal human effort: The paper shows that practitioners can iteratively identify and correct multiple different biases in medical imaging models (tested on melanoma detection and bone age estimation) without requiring extensive manual intervention at each step.

Responsible AI practices should be built into the model development lifecycle: Rather than treating bias correction as an afterthought, the framework emphasizes that explainability and bias mitigation should be integrated throughout the entire development process to ensure safer deployment of AI in medical and other high-stakes domains.