Explainable Deep Classification Models for Domain Generalization

Explainability and accuracy are not inherently at odds: The paper demonstrates that models can be made more interpretable through saliency-based training without sacrificing classification accuracy, challenging the common assumption that there is a necessary trade-off between these two important properties.

Visual explanations improve real-world performance: By training models to focus on ground-truth object regions (represented as saliency maps showing which pixels influenced decisions), the approach improved generalization when models were applied to new domains, suggesting that interpretable models may be more robust in practice.

Explainability can be measured and validated: The researchers quantified explainability using both automated metrics and human judgment, establishing that model explanations are not purely subjective and can be systematically evaluated—an important step toward standardizing transparency in AI systems.

Interpretability helps separate relevant features from domain bias: By using model explanations to distinguish between features that genuinely identify objects versus features that are specific to particular visual domains, the approach reduced unwanted domain-specific biases, demonstrating a practical benefit of explainability for fairness and reliability.

This approach is applicable to real-world deployment: The method works on standard deep learning architectures for object classification without performance penalties, making it potentially feasible for organizations seeking to deploy more transparent AI systems in production environments.