Interpretability Guarantees with Merlin-Arthur Classifiers

arXiv:2206.00759v3 Announce Type: replace
Abstract: We propose an interactive multi-agent classifier that provides provable interpretability guarantees even for complex agents such as neural networks. These guarantees consist of lower bounds on the mutual information between selected features and the classification decision. Our results are inspired by the Merlin-Arthur protocol from Interactive Proof Systems and express these bounds in terms of measurable metrics such as soundness and completeness. Compared to existing interactive setups, we rely neither on optimal agents nor on …

abstract agent agents arthur arxiv classification classifier classifiers cs.ai cs.lg decision express features information interactive interpretability merlin multi-agent networks neural networks protocol results systems type