Masked attribution-based probing of strategies as a computational framework to align human, non-human primate, and model explanations

Engineering notes

Key topics: autonomous driving. See the paper for implementation details and experimental results.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

What visual information do primate brains use to recognize objects, and can explanations from artificial neural networks (ANNs) help reveal these biological recognition strategies? Answering this question is important because humans and macaques both perform rapid, robust object recognition, yet the diagnostic image features guiding their behavior are difficult to measure at scale. Behavioral methods such as Bubbles can estimate these features but require extensive psychophysical data, whereas ANN explanation methods, including saliency and guided backpropagation, are efficient but often disagree with one another and lack direct biological validation. Here, we introduce MAPS, Masked Attribution-based Probing of Strategies, a framework that makes ANN-derived explanations testable in biological systems by linking them to neurobehavioral consequences. MAPS converts explanation maps into minimal explanation-masked images and asks whether these images preserve the original image-by-image recognition behavior. In silico, EMI-based behavioral similarity reliably recovers ground-truth similarity between model strategies. Applied to humans (n = 56) and macaques (n = 2), MAPS identifies explanation methods that best align with biological vision, achieving validity comparable to Bubbles without exhaustive psychophysics. MAPS provides a scalable, behaviorally grounded approach to evaluate and compare ANN explanations across brains and machines.

Links and sources

• PDF from original source

Comments