MAPS: Masked Attribution-based Probing of Strategies- A computational framework to align human and model explanations

Sabine Muzellec, Yousif Kashef Alghetaa, Simon Kornblith, Kohitij KararXiv 2025 · DOI · arXiv

Introduction: The Interpretability Gap

In the field of Explainable AI (XAI), researchers frequently rely on "attribution maps"—vibrant heatmaps designed to highlight the pixels a model deems most important for a given classification. However, these visualizations often fall into the "eye candy" trap: a heatmap that looks intuitive to a human observer does not necessarily reflect the model's actual internal logic. Furthermore, there has historically been a significant measurement gap; validating whether an AI's strategy aligns with biological vision (human or non-human primate) has required labor-intensive psychophysical methods like "Bubbles," which can demand tens of thousands of trials for even a small set of images.

To address these challenges, researchers have introduced MAPS (Masked Attribution-based Probing of Strategies), a computational framework designed to adjudicate between competing XAI methods by grounding them in behavioral data.

The MAPS Framework: Causal Testing of AI Strategies

The core innovation of MAPS is the Explanation-Masked Image (EMI). Rather than simply looking at a heatmap, the framework uses the attribution map from a "Reference" model (e.g., an Artificial Neural Network or ANN) to create a mask. By applying a percentile cutoff—keeping only the top 5% or 10% of pixels identified as "important"—researchers generate a minimal image that is then shown to a "Target" system (a human, monkey, or another AI).

The MAPS pipeline (A), model-human consistency (B), XAI heatmaps (C), and similarity analysis (D).

This approach shifts XAI from qualitative visualization to causal testing. If the Target system maintains high accuracy and exhibits the same image-by-image error patterns (consistency) as the Reference model when viewing these sparse EMIs, the explanation is considered "aligned." This allows for a principled way to determine which interpretability tools truly capture the "how" of visual recognition.

Key Insights: Method Choice and Biological Alignment

The study, which utilized MS COCO images for a 10-way classification task, yielded several critical findings:

  • Method Choice Trumps Training: The specific XAI algorithm used (e.g., Saliency vs. Integrated Gradients) causes more variation in explanations than the random initialization of the network itself. This underscores the necessity of choosing the right interpretability tool for the task.
  • Noise Tunnel Saliency as the Gold Standard: Among 12 tested methods, Noise Tunnel Saliency consistently produced the most human- and monkey-aligned explanations. In contrast, older methods like Deconvolution failed to capture biological visual strategies.
  • Neural Grounding: The model-method combinations that showed the best behavioral alignment also best predicted the firing patterns of neurons in the primate inferior temporal (IT) cortex, suggesting that MAPS-derived strategies reflect the representational logic of the ventral visual stream.

Validating Against the "Gold Standard"

MAPS was benchmarked against the traditional "Bubbles" method, which randomly masks images to identify diagnostic regions. While Bubbles is highly accurate, it is inefficient, requiring roughly 36,000 trials to map 40 images. MAPS achieved the same behavioral validity using significantly fewer trials (~4,000), making it a scalable alternative for large-scale strategy probing.

Comparison of visual strategies (Object only vs. Bubble-based) and their impact on accuracy and behavioral consistency.

Conclusion and Future Directions

MAPS provides a robust, behaviorally grounded framework for aligning AI explanations with biological vision. While currently focused on static images, future work aims to extend MAPS to dynamic video and temporally evolving decisions. By bridging the gap between colorful heatmaps and actual visual strategies, MAPS moves the field closer to AI systems that not only perform like humans but also "see" like them.

Summary updated

Abstract

Human core object recognition depends on the selective use of visual information, but the strategies guiding these choices are difficult to measure directly. We present MAPS (Masked Attribution-based Probing of Strategies), a behaviorally validated computational tool that tests whether explanations derived from artificial neural networks (ANNs) can also explain human vision. MAPS converts attribution maps into explanation-masked images (EMIs) and compares image-by-image human accuracies on these minimal images with limited pixel budgets with accuracies on the full stimuli. MAPS provides a principled way to evaluate and choose among competing ANN interpretability methods. In silico, EMI-based behavioral similarity between models reliably recovers the ground-truth similarity computed from their attribution maps, establishing which explanation methods best capture the model's strategy. When applied to humans and macaques, MAPS identifies ANN-explanation combinations whose explanations align most closely with biological vision, achieving the behavioral validity of Bubble masks while requiring far fewer behavioral trials. Because it needs only access to model attributions and a modest set of behavioral data on the original images, MAPS avoids exhaustive psychophysics while offering a scalable tool for adjudicating explanations and linking human behavior, neural activity, and model decisions under a common standard.
BibTeX
@misc{muzellec2025maps,
  author = {Sabine Muzellec and Yousif Kashef Alghetaa and Simon Kornblith and Kohitij Kar},
  title = {{MAPS: Masked Attribution-based Probing of Strategies- A computational framework to align human and model explanations}},
  year = {2025},
  eprint = {2510.12141},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.12141},
}
Figure