Imagine you have a very smart, but mysterious, "black box" computer program (a deep neural network) that looks at a picture of a breast tissue sample and decides if it's benign or malignant. You know what it decided, but you have no idea why. It's like a doctor giving you a diagnosis but refusing to show you the X-ray or explain their reasoning.

To solve this, scientists have invented "Explainable AI" (XAI) tools. Think of these tools as different translators trying to explain the black box's logic. However, until now, these translators spoke completely different languages:

GradCAM points to the "hot spots" on the image using gradients.
SHAP plays a game of "what if we remove this feature?"
LIME builds a simple, local map around the specific image.
Integrated Gradients traces a path from a blank image to the real one.

The problem? You couldn't compare their answers. It was like trying to compare a map drawn in miles to one drawn in kilometers without a conversion formula.

Enter GRALIS: The Universal Translator

This paper introduces GRALIS (Gradient-Riesz Averaged Locally-Integrated Shapley). Think of GRALIS not just as a new tool, but as a master framework that proves all these different translators are actually speaking the same underlying language, just with different accents.

Here is the core idea, broken down with simple analogies:

1. The "Universal Recipe" (The Canonical Form)

The authors discovered that if you strip away the specific tricks of GradCAM, SHAP, LIME, and Integrated Gradients, they all follow the exact same mathematical recipe. They are all just calculating a weighted average of contributions.

Imagine you are making a smoothie to explain the AI's decision.

The Ingredients ( $\Delta$ ): These are the "marginal contributions." How much did adding a specific feature (like a pixel or a group of pixels) change the AI's mind?
The Recipe Book ( $w$ ): This is the "weight function." It decides how much importance to give to each ingredient.
The Blender ( $Q$ ): This is the "index space." It's the container where you mix everything together.

GRALIS proves that any fair, linear, and continuous way of explaining the AI's decision must look like this smoothie recipe. This is based on a famous math theorem called the Riesz Representation Theorem, which essentially says, "If you want to measure something fairly and continuously, you have to do it this way."

2. Fixing the "Broken Tools"

The paper points out that the old tools had specific flaws, like a car with a flat tire or a broken engine:

GradCAM had a "ReLU" filter (a filter that cuts off negative values). The authors say this filter breaks the math, making it impossible to compare with other tools. They propose a "linearized" version (GradCAM-lin) that removes this filter, making it fit the universal recipe.
LIME often failed to add up to the total prediction (like a budget that doesn't balance). GRALIS fixes this by ensuring the "completeness" axiom is met.
SHAP ignored the "curvature" (how features interact smoothly). GRALIS fills this gap by looking at the path between features, not just the start and end points.

3. The "Game of Coalitions"

One of the paper's coolest insights is how it handles interactions.
Imagine a team project where the success depends on how people work together.

Old methods usually just asked, "How much did Person A contribute?"
GRALIS asks, "How much did Person A contribute when working with Person B? What about when A, B, and C work together?"

It does this by turning the image into a cooperative game. It groups pixels into "coalitions" (like superpixels) and calculates exactly how much each group adds to the final score. The paper proves mathematically that GRALIS calculates these "interaction values" exactly, not as an approximation.

4. The "Multi-Scale" View

Sometimes you need to look at a picture from far away (the big picture) and sometimes up close (the details).

Old methods usually picked one scale.
GRALIS has a feature called MS-GRALIS (Multi-Scale GRALIS). It looks at the image at different levels of detail (like zooming in and out) and combines them using "optimal weights." It's like a photographer who takes a wide shot, a medium shot, and a close-up, then blends them perfectly so you don't miss any important details.

5. The "Proof" (Theorems)

The paper doesn't just say "this works"; it provides seven formal theorems (mathematical proofs) that guarantee:

Completeness: The explanations add up to 100% of the decision.
Convergence: If you run the calculation many times, the answer gets closer and closer to the truth (with a known error bound).
Uniqueness: There is only one correct way to write this formula.
Interaction: It correctly calculates how features influence each other.

6. The "Test Drive"

The authors tested this on a real-world dataset of breast cancer images (BreaKHis). They didn't just say "it looks good"; they checked if removing the "important" parts the AI highlighted actually changed the AI's prediction.

Result: When they removed the top-highlighted areas, the AI's confidence in a "malignant" diagnosis dropped significantly (96% of the time). This proves the tool is actually finding the right spots, not just guessing.

Summary

GRALIS is a mathematical unification that says: "All these different ways of explaining AI are actually the same thing, just viewed through different lenses." It provides a single, rigorous framework that fixes the flaws of the old tools, allows them to be compared fairly, and guarantees that the explanations are mathematically sound, complete, and capable of detecting how features work together.

It's like finally realizing that all the different dialects of a language are actually the same language, and now we have a dictionary that translates them all perfectly.

Technical Summary: GRALIS – A Unified Canonical Framework for Linear Attribution Methods

1. Problem Statement

The field of Explainable AI (XAI) for deep neural networks is currently fragmented. Prominent attribution methods—such as GradCAM, SHAP, LIME, and Integrated Gradients (IG)—operate on distinct theoretical foundations, rendering them formally incomparable. This fragmentation leads to empirical rather than rigorous method selection, where attribution maps from different techniques cannot be systematically compared or combined.

Prior attempts to unify these methods have been partial:

Ancona et al. established that gradient-based methods (like GradCAM) can be expressed as a "gradient $\times$ input" linear form but did not prove this structure is necessary nor include SHAP or LIME.
Covert and Lee unified LIME, SHAP, and IG via Shapley games but excluded GradCAM because its post-aggregation ReLU violates the linearity required by their framework.

Consequently, six structural gaps remain in the literature:

Arbitrary Baselines: IG relies on a fixed baseline, drastically changing results based on that choice.
Ignored Curvature: SHAP compares coalitions but ignores the path (curvature) between them.
Lack of Completeness: LIME coefficients do not necessarily sum to the model's output difference.
Spatial Limitation: GradCAM is confined to CNN feature maps and does not apply to dense layers or Transformers.
Missing Interactions: Most methods produce marginal attributions, failing to capture integrated feature interactions.
No Multi-Scale Aggregation: No method aggregates attributions across abstraction levels with mathematically optimal weights.

2. Methodology: The GRALIS Framework

The paper proposes GRALIS (Gradient-Riesz Averaged Locally-Integrated Shapley), a mathematical framework that unifies linear additive attribution methods under a unique canonical structure derived from the Riesz Representation Theorem.

The Canonical Form

GRALIS posits that every additive, linear, and continuous attribution functional in $L^2(Q, \mu)$ admits a unique canonical representation:
$\phi_i(f, x, x') = \int_Q w(q) \cdot \Delta_i(f, x, x', q) \, d\mu(q)$
Where:

$Q$ is the integration index space (e.g., paths, coalitions, or feature maps).
$w(q)$ is a weight function.
$\Delta_i$ is the marginal contribution of feature $i$ .

This form subsumes existing methods as special cases:

GradCAM-lin: A linearized version of GradCAM (removing the post-aggregation ReLU) where $Q$ represents channels and positions.
SHAP: Where $Q$ represents coalitions.
LIME: Where $Q$ represents local perturbations.
Integrated Gradients: Where $Q$ represents integration paths.

Key Algorithmic Components

Conditioned Integration Paths: Unlike standard IG which integrates over a global path, GRALIS integrates over paths conditioned on specific coalitions $S$ . Features outside $S$ remain at the baseline during integration, capturing curvature specific to that coalition.
GRALIS-MC: To address the exponential complexity of exact Shapley values ( $O(2^n)$ ), the paper introduces a Monte Carlo approximation. This reduces complexity to $O(m \cdot n \cdot k)$ with an explicit error bound combining Monte Carlo sampling error ( $O(1/\sqrt{m})$ ) and Riemann integration error ( $O(1/k)$ ).
Interaction Values: GRALIS induces a cooperative game $v_G$ from the continuous space via a measurable projection $\rho$ . It computes Shapley Interaction Values (SIVs) exactly on this induced game using the Möbius transform, rather than approximating them.
Multi-Scale Extension (MS-GRALIS): For models with multiple layers, GRALIS aggregates attributions using weights $\lambda_\ell$ derived from inverse variance weighting, minimizing the total variance of the attribution.

3. Key Contributions and Theoretical Guarantees

The paper establishes seven formal theorems that provide guarantees absent in individual methods:

T1 (Unified Canonical Form): Proves via the Riesz Theorem that the integral form $(Q, w, \Delta)$ is the necessary and unique representation for any additive, linear, and continuous attribution functional.
T2 (Exact Completeness): Guarantees that the sum of attributions equals the difference between the model output and the baseline ( $f(x) - f(x')$ ).
T3 (Convergence): Provides a convergence bound for GRALIS-MC, showing explicit error terms for both sampling and path discretization.
T4 (Exact SIVs): Demonstrates that GRALIS computes Shapley Interaction Values exactly on the induced cooperative game $v_G$ , avoiding the circularity or approximation often found in interaction estimation.
T5 (Hoeffding ANOVA): Shows that under feature independence, GRALIS terms coincide with the Hoeffding functional decomposition.
T6 (Sobol Indices): Establishes that Sobol sensitivity indices are a local limiting case of GRALIS.
T7 (Multi-Scale Optimization): Proves that inverse variance weighting provides the optimal weights for multi-scale aggregation.

Algebraic Justification: Appendix X utilizes the Möbius transform to rigorously justify the correspondence between the continuous GRALIS integral and discrete Shapley Interaction Values, proving that GRALIS constructs a valid cooperative game $v_G$ and computes SIVs exactly upon it.

4. Experimental Validation

The paper reports preliminary validation on a breast histology classification task using the BreaKHis dataset (1,187 images) and a DenseNet-121 model trained with knowledge distillation.

Implementation: Used SLIC superpixel segmentation ( $n_{seg} \approx 25$ ), 30 Monte Carlo permutations with antithetic sampling, and 10 integration steps.
Faithfulness: Evaluated via superpixel deletion. For malignant images, removing top-attribution superpixels reduced malignant confidence in 96% of cases (mean drop +0.025 to +0.027). For benign images, the effect was symmetric and theoretically coherent (removing benign evidence increased malignant confidence).
Metrics:
- SAL (Saliency): 0.762 (identifying semantically coherent regions).
- Compactness ( $\phi_{active}$ ): 0.39, a 19x improvement over feature-space variants.
- Deletion AUC: Preliminary estimates show positive AUC for malignant images and symmetric negative AUC for benign images, consistent with class-conditional structure.

Note: The authors explicitly state that a full comparative benchmark against baseline methods (GradCAM, KernelSHAP, LIME, IG) is planned for a companion paper.

5. Significance and Claims

The paper claims that GRALIS resolves the fragmentation of XAI by providing a unifying mathematical justification for linear attribution methods. Its significance lies in:

Formal Unification: It is the first framework to simultaneously encompass GradCAM (linearized), SHAP, LIME, and IG under a single necessary canonical form.
Structural Completeness: It satisfies a broader set of axiomatic properties (13.5/14 in the paper's structural comparison) than any existing method, including completeness, sensitivity, locality, and exact interactions.
Theoretical Rigor: It moves beyond empirical observation to prove that linearity is a structural necessity for additive attributions, resolving the "gap" between gradient-based and game-theoretic methods.
Optimality: It provides the first mathematically derived optimal weights for multi-scale aggregation.

The authors maintain a modest stance regarding experimental scope, acknowledging that the current validation is a proof-of-concept on a single dataset and architecture. They emphasize that the theoretical contributions (Theorems 1–7) hold unconditionally under the stated linearity and continuity conditions, independent of the empirical results. The framework does not cover nonlinear methods (e.g., standard GradCAM with ReLU, attention maps) as they fall outside the Riesz representation conditions, a limitation the authors explicitly note for future work.

GRALIS: A Unified Canonical Framework for Linear Attribution Methods via Riesz Representation