On the Value of Labeled Data and Symbolic Methods for Hidden Neuron Activation Analysis
摘要
We introduce a novel model-agnostic post-hoc Explainable AI method that provides meaningful interpretations for hidden neuron activations in a Convolutional Neural Network. Our approach uses a Wikipedia-derived concept hierarchy with approx. 2 million classes as background knowledge, and deductive reasoning based Concept Induction for explanation generation. Additionally, we explore and compare the capabilities of off-the-shelf pre-trained multimodal-based explainable methods. Our evaluation shows that our neurosymbolic method holds a competitive edge in both quantitative and qualitative aspects.