Gaussian perturbation interpretability for regression techniques to understand retrieval black-boxes

arXiv:2601.21685 · astro-ph.EP, astro-ph.IM · Submitted 2026-01-29 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Astrophysics Radio. Generated commentary on the latest astrophysics papers.

Vera: We're looking at the paper: "Gaussian perturbation interpretability for regression techniques to understand retrieval black-boxes".

Jocelyn: Here's the gist of the paper. PERTURB-c introduces PERTURB-c, a correlation-aware framework for interpreting black box regression models with one-dimensional structured inputs This framework addresses the issue where traditional interpretability methods like SHAP struggle to account for interactions between inputs in highly correlated data The authors demonstrate this framework on a simulated case study involving machine learning based transit spectroscopy retrievals of exoplanet WASP-107b.

Vera: The paper highlights that conventional feature importance methods suffer when observations are highly correlated, as modifying individual wavelength bin values can easily lead to non-physical spectra.

Jocelyn: Atmospheric retrieval problems, such as those involving exoplanet atmospheres, exemplify this challenge due to the specific ways different wavelength bins relate to each other through spectral line broadening and molecular transmission.

Vera: PERTURB-c is designed specifically to mitigate these issues by generating modified spectra governed by physically motivated relations rather than naive modifications.

Title and authors: Vera: So we're diving into this paper called "Gaussian perturbation interpretability for regression techniques to understand retrieval black-boxes." It’s about how we can look inside these complex models that handle things like exoplanet atmospheres.

Jocelyn: The authors, Jools D Clarke, Gordon Yip, and Nikolaos Nikolaou, they are the ones who developed PERTURB-c. They focus on making sure the AI isn't just guessing randomly but is actually using physically plausible features when it makes its predictions.

Subrahmanyan: That focus on physical plausibility is key because in atmospheric retrieval problems, changing one wavelength bin can easily create a spectrum that looks totally non-physical, which we don't want happening.

Vera: The core idea they are pushing is to use this PERTURB-c method to inspect the model and verify if it’s basing its results on known physical relationships in the data or if it has accidentally learned some weird patterns.

Jocelyn: They are essentially trying to help us check for biases in both our actual sky data and within the model itself, which is really important since we’re often working with simulated data.

Subrahmanyan: If we can trace a model's decision back to physical features in the data, it gives us some lasting scientific value and helps us improve how we use these machine learning techniques for things like understanding planetary populations.

Vera: So what does this whole PERTURB-c thing actually do in practice, beyond just having a fancy name? It’s a way to get inside the black box without breaking the physical reality of the problem.

Jocelyn: Right, and they achieve this by generating modified spectra that follow physically motivated relations instead of just making random changes to individual wavelength bins.

The paper's summary: Vera: So let’s talk about what the paper actually summarizes as the method works. PERTURB-c uses a specific technique to generate these modifications, and it bases this on probing the response of a cluster of adjacent bins using something called a Gaussian mask.

Jocelyn: They formalize this by defining a neighborhood, which they call N with parameters epsilon and m, where neighboring points are more positively correlated than points that are further away.

Subrahmanyan: This is implemented using a weighting function, wgauss with epsilon and m, which approximates the actual data correlation and lets them create way more samples that are actually in-distribution compared to just using standard methods like SHAP.

Vera: That sounds like they're smart about how they sample the space—they aren't just throwing random points at it; they’re intelligently choosing where to probe based on how correlated the inputs are.

Jocelyn: And because of this, the efficiency gains are huge; for example, in their simulated case study with WASP-107b, PERTURB-c was about two hundred twenty seven times faster than SHAP for that specific analysis <ref:2601.21685#pg1>.

Subrahmanyan: Two hundred twenty seven times is a big number when you’re talking about computational time for complex atmospheric retrievals. It shows that this approach is way more efficient at extracting the response gradient information they need.

The paper's improvements: Vera: So the paper lays out some specific improvements over existing methods, and one big thing they highlight is how PERTURB-c handles efficiency compared to what we usually use.

Jocelyn: They show that PERTURB-c generates fewer out of distribution samples than standard SHAP, with only about one in two hundred SHAP samples being out of distribution, whereas the PERTURB-c samples they generate are all supposed to be in-distribution <ref:2601.21685#pg1>.

Subrahmanyan: That's a nice contrast; it means the perturbation method is much more controlled and less likely to give us misleading results based on data we haven't seen before.

Vera: And they also talk about how this framework can be applied directly to retrieval constraints, which is really practical for real science work in astronomy.

Jocelyn: They use PERTURB-c in a T2 analysis mode to impose specific uncertainty limits on the abundance of a target molecule on a per wavelength bin basis. For instance, they use it to constrain the minimum abundance of silicon while also checking the uncertainty for sulfur dioxide at the same time.

Subrahmanyan: That capability is useful because it helps coordinate different instruments and shows how much observational time might actually be needed to meet specific hypotheses about planetary atmospheres.

Conclusion: Vera: So, wrapping up this paper on "Gaussian perturbation interpretability for regression techniques to understand retrieval black-boxes," the main point is that PERTURB-c offers a visualization that's much more intuitive than the standard SHAP feature importance plots when you have a huge number of input features.

Jocelyn: They successfully show how this method attributes high importance to regions where both silicon and sulfur dioxide spectral features are present, and they even infer how the model has learned to separate those two species, breaking the degeneracy between them.

Subrahmanyan: That inference about how the model breaks that degeneracy is significant because it tells us something about what the underlying physics might be in a way that's hard to see directly.

Vera: Overall, this work provides a novel way to look at these models and gives researchers a better tool for understanding how they work under the hood when dealing with complex data structures.

Jocelyn: It sets the groundwork for more confident and cohesive retrieval strategies in future studies of planetary populations by letting us see what the model is actually relying on.

Subrahmanyan: I think this is important because it moves us toward better confidence in our models, which is essential as we try to make population level studies with exoplanets more robust.

Vera: That's where we leave it for now. We'll be watching for the next paper on the show.

University College London · King’s College London

astro-ph.EP, astro-ph.IM

Submitted: 2026-01-29

Updated: 2026-10-30

Comments: Heavily revised new version! 18 pages, 13 figures, in review

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 78/100

The gist: PERTURB-c introduces PERTURB-c, a correlation-aware framework for interpreting black box regression models with one-dimensional structured inputs This framework addresses the issue where traditional…

Key concepts

Correlation Awareness
This is the core idea that the framework uses to understand data. Instead of treating every input feature independently, it recognizes that certain inputs are related (correlated), especially in spectral data. This awareness ensures that modifications made to one part of the spectrum are physically consistent with how neighboring parts should behave, preventing non-physical results.
Gaussian Mask Perturbation
The method generates modified observations by probing a cluster of adjacent wavelength bins using a Gaussian mask. This mask is mathematically defined based on the known physical correlations between neighboring points in the spectral data. This allows the framework to create new samples that are statistically more likely to be 'in-distribution' (physically plausible) than those generated by naive random modifications.
T1 Mode Analysis
This is a sample efficiency technique where PERTURB-c infers a first-order qualitative assessment of the perturbation response for a given wavelength bin. By fitting a linear approximation, it achieves sufficient accuracy using only about 50 samples, significantly reducing the computational cost compared to methods requiring hundreds of samples for full uncertainty intervals.
Retrieval Constraints
The framework allows researchers to impose specific uncertainty limits on molecular abundances at individual wavelengths. This capability is crucial for coordinating multi-instrument studies, helping determine the minimum observational time needed to test specific hypotheses about exoplanet atmospheres by constraining the model's predictions.

Terminology

Summary

PERTURB-c introduces PERTURB-c, a correlation-aware framework for interpreting black box regression models with one-dimensional structured inputs This framework addresses the issue where traditional interpretability methods like SHAP struggle to account for interactions between inputs in highly correlated data The authors demonstrate this framework on a simulated case study involving machine learning based transit spectroscopy retrievals of exoplanet WASP-107b. p. 0

Where the paper says it

Received YYY; in original form ZZZ ABSTRACT In this paper we introduce PERTURB-c, a correlation-aware framework for interpreting black box regression models with one-dimensional structured inputs. We demonstrate this framework on a simulated case study with machine learning based transit spectroscopy retrievals of exoplanet WASP-107b.

Motivation for Correlation Awareness

The paper highlights that conventional feature importance methods suffer when observations are highly correlated, as modifying individual wavelength bin values can easily lead to non-physical spectra. Atmospheric retrieval problems, such as those involving exoplanet atmospheres, exemplify this challenge due to the specific ways different wavelength bins relate to each other through spectral line broadening and molecular transmission. p. 1

Where the paper says it

Atmospheric retrieval represents one such example of these correlated problems, as there are very specific ways in which each wavelength bin in an observation is subsequently related to the others. At small scales this is due to the effects of spectral line broadening (Guest et al.

PERTURB-c is designed specifically to mitigate these issues by generating modified spectra governed by physically motivated relations rather than naive modifications. p. 1
Where the paper says it

We set out to detail the shortcomings of two main techniques, ceteris paribus and also SHAP in this application, and in lieu of this propose PERTURB-c3 , a new interpretability framework specifically tailored for highly correlated spectral data. PERTURB-c is designed to mitigate the issues with out-of-thebox feature importance methods when used for interpreting machine learning based retrievals, handling highly correlated spectral data by generating modified spectra which are instead governed by physically motivated relations.

Methodology of Perturbation Explainability

The core of the method involves generating modified observations that maintain physical correlations, which PERTURB-c achieves by probing the response of a cluster of adjacent bins using a Gaussian mask. The authors formalize this by defining a neighborhood N (ε, m) where neighboring points are more positively correlated than non-neighboring points. p. 9

Where the paper says it

Clarke et al. Using this notation allows us to define the relationship where neighbouring points are more positively correlated than non-neighbouring points, 0 ⪅ 1 𝑀−ℓ 𝑀Í−ℓ 𝑚=1 𝜌 𝑛 x:,𝑚1(𝑀−ℓ ) ⊺ , X:,m\N ( 𝜀,𝑚) < 1 ℓ Í ℓ 𝑚=1 𝜌 𝑛 x:,𝑚1ℓ ⊺ , X:,N ( 𝜀,𝑚) .

This is implemented using a weighting function wgauss. p. 9
Where the paper says it

(16) In practice, we can define this as a windowing function 𝑤gauss. (𝜀, 𝑚) with a Gaussian weighting, 𝑤gauss.

(ε, m) which approximates the real data correlation and allows for the generation of significantly more in-distribution samples than standard SHAP methods. p. 9
Where the paper says it

(17) With the data correlation constrained in this way, it becomes far easier to generate samples which are consistent with the prior distribution. By applying the windowing function which has been found to approximate the real data correlation, we can ‘diffuse’ the way in which perturbation is applied, allowing for generation of significantly more in distribution samples [Figure 9].

Performance and Efficiency Gains

The PERTURB-c framework shows superior efficiency compared to existing techniques, achieving a speedup of approximately 227 times over SHAP for the WASP-107b case study. The authors show that PERTURB-c generates fewer out-of-distribution samples than SHAP, with only about 1 in 200 SHAP samples being out of distribution, whereas PERTURB-c samples are generated such that they should all be in-distribution. p. 8

Where the paper says it

On average in this highly correlated environment only 1 in 200 SHAP samples are indistribution, whereas PERTURB-c samples are generated such that they should all be in-distribution. As such, PERTURB-c needs to generate a reduced number of samples (approximately 200× fewer than SHAP) in order to extract the same information content (same number of in-distribution samples) [Table 2].

This efficiency allows for the extraction of response gradient information at a minimum number of samples, requiring only 50 samples for the default T1 analysis to achieve a 2. p. 9
Where the paper says it

negative contribution to the output) assessment of the perturbation response for a given wavelength bin. We sample the minimum viable number of points from the response curve and perform linear regression on the resulting perturbation response to acquire the gradient, or the mean posterior response to injected perturbation.

0412σ confidence interval. p. 2
Where the paper says it

The most important aspect that we want to untangle is the feature importance, which is the relative impact of each input wavelength bin in our observed data to the results of our retrieval. This allows for statements to be made such as “we detect phosphine at a 3𝜎 confidence level given that the uncertainty on the 0.

Application in Retrieval Constraints

The framework is applied to retrieval constraints, allowing researchers to impose specific uncertainty limits on molecular abundances per wavelength bin. For instance, the authors use PERTURB-c at T2 analysis to constrain the minimum abundance of SiO while simultaneously assessing the uncertainty for SO2. p. 3

Where the paper says it

60 SO2 log abundance -6. 72 SiO log abundance -9.

This capability helps in coordinating multi-instrument synergies and informs the required observational time needed to meet specific hypotheses about planetary atmospheres.

Conclusion on Interpretability

The study concludes that PERTURB-c provides a novel visualization that is more intuitive than standard SHAP feature importance visualizations for conveying information given a large number of input features. The framework successfully attributes high feature importance to regions where SiO and SO2 spectral features are present, even inferring how the model has learned to break degeneracy between these two molecular species. This work lays the groundwork for further exploration into improving confidence and cohesion in retrieval strategies for future population level studies. p. 2

Where the paper says it

The most important aspect that we want to untangle is the feature importance, which is the relative impact of each input wavelength bin in our observed data to the results of our retrieval. This allows for statements to be made such as “we detect phosphine at a 3𝜎 confidence level given that the uncertainty on the 0.

--- Page 1 ---

PERTURB-c introduces PERTURB-c, a correlation-aware framework for interpreting black box regression models with one-dimensional structured inputs The authors demonstrate this framework on a simulated case study with machine learning based transit spectroscopy retrievals of exoplanet WASP-107b.

--- Page 2 ---

PERTURB-c introduces PERTURB-c, a correlation-aware framework for interpreting black box regression models with one-dimensional structured inputs The authors demonstrate this framework on a simulated case study with machine learning based transit spectroscopy retrievals of exoplanet WASP-107b.

--- Page 3 ---

PERTURB-c introduces PERTURB-c, a correlation-aware framework for interpreting black box regression models with one-dimensional structured inputs The authors demonstrate this framework on a simulated case study with machine learning based transit spectroscopy retrievals of exoplanet WASP-107b.

--- Page 4 ---

PERTURB-c introduces PERTURB-c, a correlation-aware framework for interpreting black box regression models with one-dimensional structured inputs The authors demonstrate this framework on a simulated case study with machine learning based transit spectroscopy retrievals of exoplanet WASP-107b.

--- Page 5 ---

PERTURB-c introduces PERTURB-c, a correlation-aware framework for interpreting black box regression models with one-dimensional structured inputs The authors demonstrate this framework on a simulated case study with machine learning based transit spectroscopy retrievals of exoplanet WASP-107b.

--- Page 6 ---

PERTURB-c introduces PERTURB-c, a correlation-aware framework for interpreting black box regression models with one-dimensional structured inputs The authors demonstrate this framework on a simulated case study with machine learning based transit spectroscopy retrievals of exoplanet WASP-107b.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Develop a correlation aware perturbation explainability framework called PERTURB-c to interpret black box regression models for one-dimensional structured inputs by leveraging physical knowledge of known spectral correlation. This allows AI systems to inspect for biases and verify predictions are made on the basis of physically plausible features.

  2. Implement a method for generating modified spectra governed by physically motivated relations, which makes the method much more stable in situations where out-of-the-box feature importance methods fail due to highly correlated spectral data. This mitigates the issue where naïvely modifying individual wavelength bin values in an observation it is very easy to arrive at a spectrum which is highly non-physical and therefore not within the learned parameter space of the model.

  3. Utilize PERTURB-c in a T2 analysis mode to visualize response curves of multiple input wavelengths concurrently as heatmaps, avoiding the limitations of standard SHAP waterfall plots which can become difficult and unintuitive to read and the positioning of the bars becomes overall less informative for large numbers of features. This provides a more intuitive visualization than standard SHAP feature importance.

  4. Apply PERTURB-c to constrain retrieval uncertainties by imposing limits on the response deviation of a target molecule on a per-wavelength-bin basis, which can inform the number of visits which may be required to constrain the target for this specific hypothesis. This allows for the coordination of multi-instrument synergies without requesting telescope time in excess of what is required.

  5. Improve sample efficiency during interpretability analysis by using Tier 1 (T1) mode where we want to infer a firstorder qualitative (e.g. positive vs. negative contribution to the output) assessment of the perturbation response for a given wavelength bin by fitting a linear approximation, which requires only 50 samples for sufficient accuracy.

Abstract

Characterizing the atmospheres of large numbers of exoplanets from their observed spectra can answer key questions about planetary populations. Traditional retrievals are resource-intensive; machine-learning methods offer a fast alternative, but their complexity renders them black-boxes. Understanding how they reach their predictions allows us to inspect them for biases and verify predictions are based on physically plausible spectral features. Commonly used perturbation-based methods for explaining machine learning struggle with the strong correlations between inputs in spectra. We introduce Gaussian perturbation methodology for explaining black-box regression models with one-dimensional structured inputs, which addresses these issues by leveraging physical knowledge of spectral correlation. We compare it with naive perturbation of one wavelength bin at a time. Although traditional retrievals are built on physical forward models, their mapping from spectrum to retrieved parameters is a black-box. We therefore first demonstrate this methodology using TauREx-III as a proxy black-box retrieval on HST observations of HD 209458 b and JWST observations of HAT-P-26 b, showing the efficacy of Gaussian perturbation-based interpretability against a well-established, physically motivated retrieval framework, before applying it to a convolutional neural network retrieval of HAT-P-26 b. For the gases the retrievals constrain, the importance from Gaussian perturbation agrees with the expected importance, including at the core of the CO 2 band of HAT-P-26 b, where a degeneracy inverts the response. The method shows which parts of a spectrum a retrieval relies on, which the physics of its forward model alone does not, and shows the neural network has learned the same degeneracy.

Related papers