Data augmentation and explainability for bias discovery and mitigation in deep learning

Bias in AI systems comes from multiple sources: The paper identifies that bias can originate from both data collection and model design stages in machine learning pipelines, requiring systematic categorization to understand where problems arise.

Explainable AI is essential for discovering hidden biases: By making AI decision-making transparent through explainability methods, researchers can identify problematic patterns that models learn—such as the paper's example of skin lesion classifiers relying on texture rather than medically relevant features.

Manual inspection alone is insufficient at scale: The dissertation demonstrates that while careful human review of datasets (like examining skin lesion images) can find biases, this process is labor-intensive; semi-automatic approaches using explainability methods offer more practical alternatives for larger datasets.

Data augmentation can actively reduce bias: Rather than passively accepting biased training data, the paper proposes techniques like Style Transfer Data Augmentation that deliberately create diverse training examples to prevent models from learning spurious correlations (e.g., confusing texture with disease type).

Multiple mitigation strategies are needed: The research presents three complementary approaches—Style Transfer, Targeted Augmentations, and Attribution Feedback—suggesting that addressing bias effectively requires combining different technical methods rather than relying on a single solution.