Traceable Spectral Inference via Influence Functions: Efficient Data Attribution and Error Proxies for the Ariel Mission

Data attribution can work without labels: The paper reformulates influence functions to operate on predictions rather than loss values, enabling the method to identify which training data points most influenced a model's decisions even when ground truth labels aren't available—a critical capability for space missions like Ariel where validation data is scarce during operations.

Computational efficiency matters for space deployment: By using the mathematical closed-form solution of an Extreme Learning Machine, the researchers achieved efficient computation of influence scores, making the approach practical for real-time operational use rather than just post-hoc analysis.

Error prediction improves scientific reliability: The paper demonstrates that influence-based error proxies correlate strongly with actual spectral errors in simulated data, suggesting the method can predict when model outputs may be unreliable—essential for assessing "physical plausibility" in scientific missions where incorrect conclusions could waste resources or mislead research.

Identifying problematic training data enhances model trustworthiness: The approach enables scientists to pinpoint both the most helpful and most harmful training samples, allowing them to understand and potentially improve their datasets before deployment, rather than treating the model as a "black box."

Bridges the gap between AI interpretability and scientific requirements: Unlike typical explainable AI methods focused on feature importance, this work directly addresses the operational needs of scientific missions by combining data attribution with error estimation and physical validation concerns.