OperatorSHAP: Fast and Accurate Shapley Value Estimation for Neural Operators
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: I'm Tom.
Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.
Tom: We're looking at the paper: "OperatorSHAP: Fast and Accurate Shapley Value Estimation for Neural Operators".
Jane: Here's the gist of the paper. Understanding model predictions is essential for physical applications, where outputs often inform safety-critical decisions such as structural load assessment, weather warnings, and clinical diagnosis OperatorSHAP introduces a grid-agnostic attribution method and training procedure that allows us to train FastSHAP-like explainers for neural operators We establish a theoretical framework for attributions in function space, connecting to Aumann–Shapley values We further show that OperatorSHAP’s explanations are consistent with state-of-the-art discrete Shapley values across resolutions and transfer across grid sizes without retraining Theoretical Foundations of Attributions The paper establishes a theoretical framework for attributions in function space, connecting to Aumann–Shapley values We investigate these theoretical foundations and show how to design grid-agnostic explainers for neural operators We connect this to the function-valued integrated gradient formula and show under which conditions the game and its values are well-defined for current neural operator architectures The smooth operator game is pNA, and admits a unique Aumann–Shapley value given by the diagonal formula This means that we can use the discrete Shapley values to train an approximator of the Aumann–Shapley values Amortized Grid-Agnostic Explainer We provide a FastSHAP-like training procedure for Aumann–Shapley values, allowing us to train amortized explainers even when the samples are provided on irregular grids We show that OperatorSHAP’s explanations are consistent with state-of-the-art discrete Shapley values across resolutions and transfer across grid sizes without retraining The training loop for OperatorSHAP involves sampling a input u ∼ p(u), coalition S ∼ pw, evaluating the backbone on the masked inputs uS to obtain the targets v(S), and updating the explainer against the FastSHAP weighted MSE Empirical Validation and Performance We validate our approach on a variety of partial differential equations (PDEs) with varying grids per sample and compare the results to classical Shapley approximators with regard to correlation, error, and faithfulness OperatorSHAP achieves comparable and often superior quality to RegressionMSR and IG across resolutions We show that for higher player numbers, classical methods need drastically higher budgets to outperform OperatorSHAP Connection to Integrated Gradients We show how the derived formula connects to the Integrated Gradient method The Aumann–Shapley allocation to a coalition S is the µ(S)-weighted sum of the masked pointwise attributions of Def.
Paper summary: Tom: We're starting with this paper, OperatorSHAP: Fast and Accurate Shapley Value Estimation for Neural Operators. The central thesis is that understanding what a neural operator predicts is crucial for safety-critical applications, like structural loads or weather warnings.
Jane: And the paper addresses the problem that standard Shapley values are too slow when the data comes from irregular grids, which is a major issue in physical applications. OperatorSHAP introduces a grid-agnostic attribution method and a training procedure for neural operators to fix this.
Lu: They establish a theoretical framework for attributions in function space, linking them to Aumann–Shapley values, which gives them the mathematical backbone for building these explainers.
Meng: The paper claims that OperatorSHAP’s explanations are consistent with state-of-the-art discrete Shapley values across different resolutions and grid sizes without needing any retraining. That consistency is a big claim because it suggests robustness across varying data setups.
Lalam: Essentially, they've shown how to train fast, FastSHAP-like explainers for neural operators by connecting them to the Aumann–Shapley values in function space.
Jane: So, what does this mean for us right now? It means we have a new way to get reliable explanations for complex physical models without being constrained by the specific grid setup of our input data.
Tom: Right. The paper is focused on establishing this theoretical link and showing how to train an approximator that captures those Aumann–Shapley values efficiently.
Lu: They also show how this connects to the function-valued integrated gradient formula, which gives them a rigorous way to define these attributions in function space under certain conditions for neural operator architectures.
Meng: But I still wonder about the practical side. How does this translate into something that runs fast enough for real-time use on irregular inputs? Is it just theoretical consistency, or is there a concrete performance guarantee we can rely on when deploying this?
Tom: The paper shows they train these explainers using a FastSHAP-like procedure, and they compare their results to classical Shapley approximators like KernelSHAP and FastSHAP, showing comparable quality across resolutions.
Jane: So the key takeaway here is that they've managed to make Shapley-like explanations practical for neural operators by solving the grid dependency problem through a new training framework.
Lu: It’s about moving from homogeneous inputs to handling irregular grids in a way that respects the underlying theory of games in function space.
Meng: I need more concrete data on how much faster this is compared to what we are currently using for deployment, especially when the input data itself is highly heterogeneous.
Lalam: I’m quite sure they are showing a path forward by providing a method that works across different grid sizes without requiring you to retrain your model.
Conclusion: Tom: So, wrapping up OperatorSHAP: Fast and Accurate Shapley Value Estimation for Neural Operators. The title itself really captures the essence of what they’ve done—fast and accurate estimation of Shapley values specifically for neural operators.
Jane: It’s about taking a computationally expensive attribution method and making it work efficiently, especially when the data doesn't live on a neat, regular grid. They are providing a practical tool for physical applications where explanations matter.
Lu: The implication is that we can start to trust these explanations more in complex physical simulations because the theoretical foundation they built connects function space theory directly to usable training procedures for neural operators.
Meng: For production, it means we have a new candidate for explainers that should be robust against grid heterogeneity during inference, provided the training procedure holds up under real-world data variations.
Lalam: I think the big picture is that they’ve successfully connected the abstract mathematical concepts of Aumann–Shapley values to a fast, trainable method that works across different resolutions and input structures.
Tom: It seems like they’ve built something that bridges the gap between high-level game theory in function space and deployable, practical model explanations for neural operators.
Jane: That’s right. It’s not just a new algorithm; it's a new way to think about how we attribute importance in these kinds of complex AI models when those models are used for real-world physics simulations.
Joshua Stiller, Santo M. A. R. Thies, Felix Czaja, Eyke Hüllermeier
Munich Center of Machine Learning · German Research Foundation
cs.LG, cs.AI
Submitted: 2026-06-26
Updated: 2026-11-06
Comments: Accepted at NeurIPS 2026
Code: https://github.com/facebookresearch/hydra
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 91/100
The gist: Understanding model predictions is essential for physical applications, where outputs often inform safety-critical decisions such as structural load assessment, weather warnings…
Key concepts
- Aumann–Shapley Values
- These are a theoretical way to fairly distribute the contribution of each input feature (or 'player') when predicting an output in a complex system. The paper uses these values as the target for training, aiming to approximate them accurately for neural operators.
- Grid-Agnostic Explainer
- This refers to an explanation method that works regardless of how the input data is structured or sampled on a grid. OperatorSHAP allows this explainer to be trained efficiently even when samples are provided on irregular grids, unlike methods that require retraining for different resolutions.
- FastSHAP-like Training
- This describes the training procedure used by OperatorSHAP, which is modeled after the FastSHAP technique but adapted for neural operators. It allows the system to learn how to calculate Shapley values efficiently by using a weighted Mean Squared Error update loop.
Terminology
Summary
Understanding model predictions is essential for physical applications, where outputs often inform safety-critical decisions such as structural load assessment, weather warnings, and clinical diagnosis OperatorSHAP introduces a grid-agnostic attribution method and training procedure that allows us to train FastSHAP-like explainers for neural operators We establish a theoretical framework for attributions in function space, connecting to Aumann–Shapley values We further show that OperatorSHAP’s explanations are consistent with state-of-the-art discrete Shapley values across resolutions and transfer across grid sizes without retraining
Theoretical Foundations of Attributions
The paper establishes a theoretical framework for attributions in function space, connecting to Aumann–Shapley values We investigate these theoretical foundations and show how to design grid-agnostic explainers for neural operators We connect this to the function-valued integrated gradient formula and show under which conditions the game and its values are well-defined for current neural operator architectures The smooth operator game is pNA, and admits a unique Aumann–Shapley value given by the diagonal formula This means that we can use the discrete Shapley values to train an approximator of the Aumann–Shapley values
Amortized Grid-Agnostic Explainer
We provide a FastSHAP-like training procedure for Aumann–Shapley values, allowing us to train amortized explainers even when the samples are provided on irregular grids We show that OperatorSHAP’s explanations are consistent with state-of-the-art discrete Shapley values across resolutions and transfer across grid sizes without retraining The training loop for OperatorSHAP involves sampling a input u ∼ p(u), coalition S ∼ pw, evaluating the backbone on the masked inputs uS to obtain the targets v(S), and updating the explainer against the FastSHAP weighted MSE
Empirical Validation and Performance
We validate our approach on a variety of partial differential equations (PDEs) with varying grids per sample and compare the results to classical Shapley approximators with regard to correlation, error, and faithfulness OperatorSHAP achieves comparable and often superior quality to RegressionMSR and IG across resolutions We show that for higher player numbers, classical methods need drastically higher budgets to outperform OperatorSHAP
Connection to Integrated Gradients
We show how the derived formula connects to the Integrated Gradient method The Aumann–Shapley allocation to a coalition S is the µ(S)-weighted sum of the masked pointwise attributions of Def.
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
-
Grid-agnostic attribution for neural operators allows for training FastSHAP-like explainers on datasets with irregular grids, addressing a limitation of current amortized explainers that are restricted to homogeneous inputs. This enables reliable explanation generation for physical applications involving heterogeneous data, such as those from sensor placements or irregular geometries.
-
Consistent attribution across resolutions and transfer between grid sizes without retraining means an operator explainer can provide explanations across different input discretizations, which is crucial for real-world deployment where model inputs might vary in resolution. This avoids the need to re-train explainers for every new grid size encountered during inference.
-
The theoretical framework connects operator games to Aumann–Shapley values, providing a rigorous mathematical foundation for attributions in function space and showing that the smooth operator game lies in the pNA class with a well-defined value. This allows for the discretization of continuous values to predict Shapley values over finite games without retraining.
-
The integration of pointwise attribution methods like Integrated Gradients with Aumann–Shapley theory provides a way to define a
Partition-of-unity attribution
that equals the Aumann–Shapley allocation, allowing for efficient computation via the formula:(φv)(S) = µ(S)-weighted sum of the masked pointwise attributions of Def. 5.
-
The system can perform inference in a single forward pass after training, as
OperatorSHAP returns full attributions in a single forward pass, with zero per-sample budget,
significantly improving the speed and efficiency compared to classical Shapley estimators like KernelSHAP or RegressionMSR.
Abstract
Understanding model predictions is essential for physical applications, where outputs often inform safety-critical decisions, such as structural load assessment, weather warnings, and clinical diagnosis. Shapley values satisfy many desirable properties as an attribution method, but their computational cost during inference hinders their practical use. Current amortized explainers, such as FastSHAP, are limited to homogeneous inputs, which is problematic for physical applications where data often comes from irregular grids and geometries. We introduce OperatorSHAP, a grid-agnostic attribution method that allows us to train FastSHAP-like explainers for neural operators. We establish a theoretical framework for attributions in function space, connecting to Aumann-Shapley values. We further show that OperatorSHAP's explanations are consistent with state-of-the-art discrete Shapley values across resolutions and transfer across grid sizes without retraining.
Sources
- Functional relevance based on the continuous Shapley value
- An Odd Estimator for Shapley Values
- PolySHAP: Extending KernelSHAP with Interaction-Informed Polynomial Regression
- FastSHAP: Real-Time Shapley Value Estimation
- A Library for Learning Neural Operators
- A New Baseline Assumption of Integated Gradients Based on Shaply value
- Amortized Linear-time Exact Shapley Value for Product-Kernel Methods
- Regression-adjusted Monte Carlo Estimators for Shapley Values and Probabilistic Values
- Linear TreeShap
Related papers
- FlexMoE: Learning One-for-All Multi-Budget Expert Pruning for Flexible MoE Deployment
- OptiMind: Improving Small Language Models for Optimization Modeling through Formulation Guidance and Multi-Turn Learning
- Staying in the Sweet Spot: Boost Reasoning Evolution via Responsive Rollout Difficulty Manipulation
- Examining the Impact of Data and Training Methods on Emergent and Subliminal Misalignment
- Rethinking Black-Box On-Policy Distillation: A Rubric-based Analysis
- (Mis)generalization of Helpful-Only Fine-tuning