MiSS: Explaining Point Cloud Classifiers via Minimal Sufficient Coalitions
A new framework named MiSS has been developed by researchers to elucidate 3D point cloud classifiers using perturbation-relative sufficiency reasoning. This system interprets superpoint partitions as understandable abstractions and utilizes a weighted MaxSAT approach to suggest minimal coalitions of geometric areas, which are subsequently validated by a statistical oracle. MiSS distinguishes between candidate proposals and their verification, incorporating strategies such as adaptive cardinality floor, certified exact-size fallback, and surrogate acquisition. The framework outputs statistically verified sufficient coalitions as binary attributes, allowing for explanations without needing white-box access to the classifier.
Key facts
- MiSS is a black-box, query-based framework for explaining 3D point cloud classifiers.
- It uses perturbation-relative sufficiency reasoning with superpoint partitions as interpretable abstractions.
- A weighted MaxSAT procedure proposes minimal coalitions of geometric regions.
- The system employs heuristics including adaptive cardinality floor, certified exact-size fallback, safely tightened upper bound, blocking clauses, and surrogate acquisition.
- A blackbox statistical oracle verifies sufficiency from prediction queries.
- The output is a statistically verified sufficient coalition as a binary attribute.
- MiSS does not require Boolean feature spaces or white-box logical encodings of the predictor.
- The framework is designed for 3D point cloud classifiers.
Entities
—