Explainable Image Classification

The field studies how to make image classifiers’ decisions understandable, so people can inspect evidence, detect shortcuts or bias, and judge when a prediction is reliable. Common approaches include saliency and class activation maps, perturbation and causal tests, and concept- or prototype-based models that link decisions to named features or image parts. Some methods explain existing models after training, while others build explanations into the architecture through concept bottlenecks, sparse representations, or prototypes; evaluation uses faithfulness and localization tests alongside human studies and robustness checks. A central challenge is that explanations that look clear may not faithfully reflect the model’s reasoning, and results remain difficult to compare across varied metrics, architectures, and real-world settings.