📊 statistics

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

This paper proposes an operational framework for the ESA's Ariel mission that utilizes reformulated influence functions on Extreme Learning Machines to enable label-free data attribution and derive a conservative error proxy for scientific machine learning in the absence of ground truth.

Original authors: Nikki Grens, Luís F. Simões, Kai Hou Yip, Theresa Lueftinger

Published 2026-08-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nikki Grens, Luís F. Simões, Kai Hou Yip, Theresa Lueftinger

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the vast silence between the stars, astronomers are listening for the whispers of distant worlds. When a planet passes in front of its home star, a tiny fraction of that starlight filters through the planet's atmosphere, carrying with it a chemical fingerprint of the air surrounding that alien world. This signal is incredibly faint and buried within a sea of noise, requiring sophisticated tools to decode. For decades, scientists have relied on classical physics-based methods to extract these secrets, but the sheer volume of data expected from future space missions is pushing these traditional tools to their limits. To keep pace, researchers are turning to machine learning, teaching computers to recognize patterns in the light and infer the composition of these distant atmospheres. However, a new problem has emerged alongside this technological leap. In the high-stakes environment of space exploration, where there is no way to send a probe back to check the answer, scientists must be able to trust the computer's reasoning. If a machine learning model makes a prediction, the operators need to know not just what the answer is, but why the model chose it, and which pieces of data from its training history were most responsible for that decision. Without this transparency, the models remain black boxes, and in the unforgiving vacuum of space, a black box is a risk no mission can afford.

This is the challenge that a team of researchers set out to solve for the European Space Agency's upcoming Ariel mission, a project designed to survey the atmospheres of about one thousand exoplanets. The scientists developed a new way to trace the decisions of a machine learning model back to the specific examples it learned from during its training. Instead of simply asking which features of the input data mattered most, they asked a different question: which specific training examples shaped the final prediction? To do this, they used a mathematical technique known as influence functions, which estimates how much a single piece of training data would change the model's output if that piece were removed. In most machine learning applications, this calculation is incredibly slow and computationally expensive, often requiring the model to be retrained thousands of times. However, the researchers found a clever shortcut by using a specific type of neural network called an Extreme Learning Machine. This architecture allows them to calculate the influence of every single training sample almost instantly, without needing to retrain the model or even know the correct answer for the test data.

The team tested their method using simulated data that mimics what the Ariel mission will eventually observe. They trained their model to predict the transmission spectra of exoplanets based on light curve data, a complex task where the computer must translate the dimming of starlight into a detailed map of atmospheric gases. Once the model was trained, the researchers used their new influence-tracking system to look at a specific prediction and identify the handful of training examples that had the most significant impact on that result. They found that the most influential samples were not necessarily the ones that looked most similar to the test case in a simple visual sense. Instead, the model relied on a broader set of examples that drove the mathematical structure of its predictions. Crucially, the system could also identify "harmful" samples—training data points that the model struggled to learn correctly. By combining the influence of a sample with the error the model made on it during training, the system flagged data points that were likely to introduce bias or errors into future predictions.

Perhaps the most practical result of this work is a new way to estimate how wrong a prediction might be, even when the true answer is unknown. The researchers derived a "conservative error proxy," a score that estimates the potential size of an error by looking at how much the model's training data was imperfect and how sensitive the prediction was to those imperfections. When they compared this estimated error score against the actual errors in their simulated data, they found a strong correlation. The score was particularly good at predicting errors related to the shape of the spectral curve, which is vital for identifying specific molecules. This means that in the future, when the Ariel mission is flying and collecting real data, the system will be able to tell operators not only what the atmosphere likely contains, but also how much confidence to place in that answer and which parts of the training history might be causing uncertainty.

The efficiency of this approach is what makes it viable for space operations. Calculating the influence of every training point for a single prediction took the team just over a minute using their method. In contrast, the traditional way of checking this by removing one data point at a time and retraining the model would have taken nearly a full day to complete the same task. This speed difference transforms the technique from a theoretical curiosity into a practical tool for real-time mission support. The researchers also demonstrated that their method works well with more complex, ensemble-based models, suggesting it can scale to the sophisticated algorithms needed for future missions. By providing a transparent window into the model's decision-making process, this work offers a framework for scientific machine learning where trust is not assumed but verified. It ensures that when we finally decode the atmospheres of distant worlds, we will understand exactly how we got there, grounding our discoveries in the data that made them possible.

Technical Summary: Traceable Spectral Inference via Influence Functions

Problem Statement
Machine learning (ML) models are increasingly deployed in scientific space missions, such as the European Space Agency's (ESA) Ariel mission, to analyze exoplanet atmospheric data. However, the growing complexity of these models often compromises interpretability. In operational contexts like Ariel, where ground truth for in-flight observations is unavailable, the inability to validate model behavior or understand the physical plausibility of outputs restricts trust and complicates error assessment. While existing Explainable AI (XAI) methods focus on feature attribution (identifying which input variables drive a prediction), there is a critical gap in Training Data Attribution (TDA). Specifically, there is a need for efficient mechanisms to identify which specific training samples shaped a model's prediction, to detect harmful data artifacts, and to estimate prediction errors without access to ground-truth labels.

Methodology
The authors propose a framework for traceable spectral inference based on influence functions, adapted for operational spectroscopy pipelines. The methodology relies on three core technical pillars:

Reformulation for Prediction (Label-Free):
Traditional influence functions estimate the effect of a training point on test loss. Since ground truth is unavailable during mission operations, this work reformulates influence to quantify the change in the predicted value ( $f(x_{test})$ ) induced by a specific training instance. This shift allows for label-free deployment, focusing on the model's output response surface rather than an unobservable error metric.
Efficient Computation via Extreme Learning Machines (ELM):
To avoid the computational prohibitions of Leave-One-Out (LOO) retraining, the study employs an Extreme Learning Machine (ELM). In an ELM, input weights are fixed and random, while output weights are learned via Ridge Regression. This structure yields a convex objective with a closed-form solution. Crucially, the Hessian matrix ( $H$ ) of the squared-loss objective is constant and depends only on the hidden layer activations and the regularization term. This allows for the rapid, exact calculation of the inverse Hessian ( $H^{-1}$ ), enabling efficient influence computation without iterative approximations.
Influence-Based Error Proxy and Harmfulness:
The framework derives a conservative error proxy by propagating training residuals through influence sensitivities.
- Prediction Influence: Calculated as $I(z_i, x_{test}) = \phi_{test}^T H^{-1} (\phi_i \cdot r_i)$ , where $\phi$ represents hidden layer activations and $r_i$ is the training residual.
- Harmfulness Metric: Adapted from Cook's Distance, "harmfulness" is defined as the product of the absolute influence and the training residual ( $|I \cdot r|$ ). This identifies samples that exert high leverage on the prediction while simultaneously exhibiting poor model fit (high error).
- Error Proxy: An infinitesimal Jackknife-inspired score aggregates squared, residual-weighted influences across wavelength channels, normalized by the predicted amplitude to create a scale-invariant observation-level risk score.

Key Contributions
The paper introduces three specific contributions for operational spectroscopy:

Label-Free Influence: A reformulation of influence functions from loss-based to prediction-based, enabling attribution in scenarios where ground truth is unknown.
Closed-Form Efficiency: The leveraging of the ELM's Ridge Regression solution to compute infinitesimal prediction influence efficiently, avoiding the need for Hessian-vector product approximations required by deep neural networks.
Conservative Error Proxy: A derived metric that propagates training residuals through influence sensitivities to estimate prediction uncertainty and identify "harmful" training samples that may introduce bias or artifacts.

Results
Evaluated against simulated spectra from the Ariel Data Challenge 2021 (mapping transit light curves to transmission spectra), the proposed framework demonstrated:

Strong Correlation with Errors: The influence-based error proxy showed a strong Spearman rank correlation with realized spectral errors. Specifically, it correlated at 0.81 with Spectral Angle Mapper (SAM) (shape-based error) and 0.69 with Root Mean Squared Percentage Error (RMSPE) (scale-based error).
Identification of Influential vs. Harmful Samples: t-SNE visualizations revealed that the most influential training samples are not necessarily the geometric nearest neighbors in feature space. While influential samples matched the query's spectral shape well, they diverged significantly in absolute scale.
Validation against LOO: Comparisons with actual LOO retraining confirmed that influence functions capture the direction and relative magnitude of the retraining effect. While LOO retraining showed an amplification effect due to full sample removal, the influence estimates provided a highly accurate first-order approximation.
Computational Efficiency: Calculating influence matrices across all cross-validation folds (approx. 6,810 training points) took 1 minute and 15 seconds. In contrast, performing a full LOO retraining for the same dataset would have taken approximately 24 hours.
Generalization: The framework was successfully applied to a Gradient-Boosted ELM (GB-ELM), maintaining strong correlations with error metrics (0.83 for SAM), demonstrating adaptability beyond single-model architectures.

Significance and Claims
The authors claim that this approach provides a viable operational framework for scientific machine learning in space missions. By enabling the identification of the most influential and potentially harmful training samples, the method offers a transparent account of the data driving specific predictions. This facilitates targeted data inspection by astrophysicists, allowing for the flagging of high-risk spectra for removal or reweighting. Furthermore, the derived error proxy serves as a robust, label-free mechanism for operational risk assessment, ensuring that models learn correct physical structures even when ground truth is inaccessible. The work suggests that influence functions can bridge the gap between complex ML models and the rigorous validation requirements of scientific space exploration.

Technical Summary: Traceable Spectral Inference via Influence Functions

More like this