Explaining deep learning of galaxy morphology with saliency mapping

Explainable AI outperformed standard deep learning for this astronomy task: The XAI method using saliency mapping achieved a correlation coefficient of 0.76 compared to 0.59 for direct prediction, suggesting that understanding how an AI makes decisions can lead to better results than simply optimizing for accuracy alone.

Transparency enabled feature reuse across different scientific problems: By using saliency mapping to visualize which parts of galaxy images the model focused on, researchers could extract measurements (like bar length) and potentially apply the same approach to measure other galaxy features (bulge-to-disk ratio) without retraining—demonstrating how explainability creates efficiency.

XAI techniques can help identify and remove dataset bias: The authors suggest that saliency mapping and similar explainability methods can reveal problematic patterns in training data, offering a practical tool for responsible AI practitioners to audit and improve datasets before they perpetuate biases in predictions.

Larger training datasets made the explainable approach more effective: The XAI method's superior performance was partly attributed to access to more training data from Galaxy Zoo, highlighting that responsible AI implementation may require investment in quality datasets—not just better algorithms.

Visual explanations bridge human expertise and AI predictions: By making the AI's decision-making process visible through saliency maps, the method allows human astronomers to verify whether the model is measuring galaxies based on scientifically meaningful features, enabling meaningful human-AI collaboration.