🤖 AI

H-Sets: Hessian-Guided Discovery of Set-Level Feature Interactions in Image Classifiers

H-Sets is a novel two-stage framework that discovers and attributes higher-order feature interactions in image classifiers by using input Hessians to identify interacting pixel sets and IDG-Vis to provide more faithful, set-level saliency explanations.

Original authors: Ayushi Mehrotra, Dipkamal Bhusal, Michael Clifford, Nidhi Rastogi

Published 2026-04-27

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Ayushi Mehrotra, Dipkamal Bhusal, Michael Clifford, Nidhi Rastogi

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

Imagine you are trying to understand why a professional chef decided a soup was "perfect."

If you used a standard explanation method, it would be like a critic pointing at individual ingredients: "The salt is important. The pepper is important. The water is important." This is what most AI researchers do today—they look at "marginal effects," or how much each single pixel contributes to a decision.

But anyone who has cooked knows that a great soup isn't just a list of ingredients; it’s the interaction between them. Salt doesn't just add "saltiness"; it unlocks the flavor of the onion. The onion and the salt together create a third, new flavor that neither could achieve alone.

H-Sets is a new way to explain AI by looking for these "flavor combinations" (feature interactions) rather than just the individual "ingredients" (pixels).

The Two-Step Recipe

The researchers created a two-stage process to figure out these combinations:

1. The "Flavor Detective" (Interaction Detection)

Instead of just looking at how much a pixel changes the result, H-Sets uses a mathematical tool called a Hessian.

Think of the Hessian as a high-powered microscope that looks for "chemical reactions" between pixels. It asks: "If I change pixel A and pixel B at the same time, does the AI's decision change in a way that is much bigger than just adding A and B together?"

If the answer is yes, the detective has found a "set"—a group of pixels that work as a team. To make sure these teams actually make sense to humans, the researchers use a tool called SAM (Segment Anything Model). This acts like a "visual organizer," ensuring the AI doesn't group a random pixel from the sky with a pixel from a dog's ear just because they have a weird mathematical connection. It keeps the "teams" spatially coherent, like grouping all the pixels that make up a bird's wing.

2. The "Fair Judge" (Interaction Attribution)

Once the detective has found the teams (e.g., "The Wing Team," "The Beak Team"), we need to know how much credit each team deserves for the final decision.

The researchers use a concept from Game Theory called "Harsanyi dividends." Imagine a basketball team wins a game. You don't just want to know how many points each player scored; you want to know how much the chemistry between the Point Guard and the Center contributed to the win.

H-Sets calculates the "dividend" for each group, essentially asking: "How much extra value did this specific group bring to the table that they couldn't have provided if they were just playing as individuals?"

Why does this matter?

The paper proves that H-Sets is better than previous methods in two main ways:

It’s "Sparser" (Less Clutter): Most AI explanations look like "salt and pepper" noise—a messy spray of dots all over the image. H-Sets produces clean, concentrated maps that highlight specific, meaningful parts of the object (like just the beak or just the eye).
It’s "Faithful" (More Truthful): Because it actually understands how pixels work together, it is more "honest" about why the AI made a choice. If you remove the "team" of pixels that H-Sets identified, the AI's confidence crashes much faster than if you removed pixels identified by older methods.

The Bottom Line

In short, while older methods explain AI by pointing at a crowd of individuals, H-Sets explains AI by identifying the specialized teams that actually get the job done.

Technical Summary: H-Sets: Hessian-Guided Discovery of Set-Level Feature Interactions in Image Classifiers

1. Problem Statement

Standard feature attribution methods (e.g., Integrated Gradients, LRP) primarily focus on marginal effects—the contribution of individual features in isolation. However, in image classification, semantic meaning is often derived from feature interactions, where groups of pixels jointly influence a model's output in a non-additive way.

Existing attempts to capture these interactions face several challenges:

Coarseness: Some methods are restricted to superpixels or segments, losing pixel-level granularity.
Complexity: Game-theoretic approaches (like Shapley-Taylor indices) are computationally intractable for high-dimensional images.
Axiomatic Failures: Many methods fail to satisfy core interpretability axioms (e.g., monotonicity, superadditivity, or completeness).
Lack of Coherence: Methods often produce noisy or spatially disconnected saliency maps.

2. Methodology

The authors propose H-Sets, a two-stage framework designed to discover and attribute higher-order feature interactions.

Stage 1: Interaction Set Detection (Hessian-Guided)
Instead of searching for masks, H-Sets uses the Hessian matrix (second-order derivatives) to identify statistical non-additivity.

Pairwise Detection: The method identifies pairs of features $(x_i, x_j)$ where the mixed partial derivative $\frac{\partial^2 f_c(x)}{\partial x_i \partial x_j}$ exceeds a threshold $\mu$ .
Recursive Expansion: To find higher-order interactions ( $|I| > 2$ ), the framework uses a breadth-first strategy. It starts with a "seed" feature (the highest-attributed pixel within a segmentation mask) and recursively merges strongly interacting neighbors into a set until a size limit $\nu$ is reached.
Spatial Prior: To ensure semantic coherence, the authors use the Segment Anything Model (SAM) as a spatial grouping prior. SAM guides the merging process to prefer features within the same semantic region, though it does not influence the gradient calculations themselves.
Smoothing: To handle the non-differentiability of ReLU, they employ a specific smooth approximation $h(z)$ that preserves ReLU-like behavior while ensuring stable Hessian computation.

Stage 2: Interaction Attribution (IDG-Vis)
Once sets are identified, the framework assigns them importance scores using IDG-Vis (Integrated Directional Gradients for Vision), a set-level extension of the IDG framework.

Directional Gradients: For a set $I$ , a direction vector $\hat{a}$ is constructed to represent the change from a baseline to the input within that set.
Path Integration: The method integrates the directional gradient along a linear path to mitigate gradient saturation and noise.
Game-Theoretic Aggregation: To isolate the unique contribution of the interaction (excluding individual effects), the method uses Harsanyi dividends from cooperative game theory. This allows the attribution to satisfy key axioms like superadditivity and monotonicity.

3. Key Contributions

A Novel Two-Stage Framework: Combines curvature-driven detection (Hessian) with path-integrated attribution (IDG-Vis).
Hessian-Based Discovery: Moves beyond heuristic mask-searching to a principled statistical detection of non-additive dependencies.
IDG-Vis: Introduces a vision-specific adaptation of Integrated Directional Gradients that works at the set level.
Axiomatic Rigor: Provides a method that satisfies a comprehensive set of interpretability axioms from both game theory and attribution literature.

4. Results

The authors evaluated H-Sets on ImageNet and CUB datasets across various architectures (VGG, ResNet, DenseNet, MobileNet).

Qualitative Performance: Compared to baselines like Integrated Gradients (noisy/diffuse), Archipelago (coarse), and MOXI (too localized), H-Sets produces saliency maps that are compact, sparse, and semantically coherent, accurately highlighting discriminative object parts.
Sparsity: H-Sets consistently achieved higher Gini index scores, meaning it produces more concise explanations by concentrating importance on fewer, more decisive features.
Faithfulness: Using the ROAD metric (Area Over the Perturbation Curve), H-Sets outperformed all baselines. This indicates that removing the features identified by H-Sets causes a more significant and accurate drop in model confidence, proving the explanations are more "truthful" to the model's actual logic.
Robustness: Ablation studies showed the method is robust to hyperparameter tuning ( $\mu$ and $\nu$ ) and successfully identifies non-semantic "decoy" features (as shown in the DecoyMNIST experiment), proving the spatial prior (SAM) does not bias the model toward semantic shortcuts.

5. Significance

H-Sets represents a significant step forward in Explainable AI (XAI) for computer vision. By shifting the focus from "which pixels are important" to "which groups of pixels interact to drive a decision," it provides a more human-like and mathematically grounded explanation of deep neural networks. Its ability to balance high-fidelity interaction detection with spatial coherence makes it a powerful tool for debugging and building trust in high-stakes AI applications.