AI summaryⓘ
The authors focus on making machine learning models easier to understand for space missions where direct verification is impossible. They change the usual way of checking model decisions, looking at how training data affects predictions instead of errors, allowing the method to work without knowing correct answers. Using a special kind of model called Extreme Learning Machine, they efficiently measure how small changes in training data influence predictions. They also create a way to estimate errors based on these influences, which matches well with actual differences in test data. Overall, their work helps identify important training examples and estimate potential mistakes in scientific models used in space missions.
InterpretabilityExplainable AIInfluence functionsTraining data attributionExtreme Learning MachineRidge regressionSpectroscopyPrediction influenceError proxyScientific machine learning
Authors
Nikki Grens, Luís F. Simões, Kai Hou Yip, Theresa Lueftinger
Abstract
Interpretability is critical for machine learning models deployed in scientific space missions such as ESA's Ariel, where ground truth is unavailable during operations and physical plausibility must be assessed. While most explainable AI methods focus on feature attribution, this work investigates training data attribution through influence functions and introduces three key contributions for operational spectroscopy pipelines. First, influence is reformulated in terms of prediction rather than loss, enabling label-free deployment. Second, by leveraging the closed-form ridge solution of an Extreme Learning Machine, infinitesimal prediction influence is efficiently computed. Third, an influence-based conservative error proxy is derived by propagating training residuals through the influence sensitivities. Evaluated against simulated spectra, the proposed proxy correlates strongly with scale and shape-based spectral errors. Furthermore, influence functions enable the identification of the most influential samples and the approximation of the most harmful ones. Together, these results suggest that this approach can serve as an operational framework for scientific machine learning.