“Post-hoc prototype explanations for classification models”
Explainable AI traditionally forces a choice between flexible but vague post-hoc methods (like saliency maps) and intuitive but rigid prototype-based architectures. This talk introduces a framework that bridges this gap: post-hoc prototype explanations. We demonstrate how to extract intuitive, example-based explanations from pre-trained networks without architectural modifications or retraining. We will explore this approach across two modalities: 1. EPIC (Vision), Accepted as an Oral at AAAI 2026: Explaining pre-trained image classifiers across benchmarks from CUB-200 to ImageNet. 2.APEX (Audio), In Review at Interspeech 2026: Adapting prototypes for the acoustic domain by disentangling signals into precise time and frequency perspectives. Together, these methods provide a unified, plug-and-play toolkit for highly interpretable visual and audio classification.
Bio
Kornel Howil is a Research Engineer at the IDEAS Research Institute in the Computer Graphics team led by prof. Przemysław Spurek. He is currently pursuing an M.S. in Computer Science at Jagiellonian University and holds two B.S. degrees in Computer Science and Astronomy from the University of Warsaw. His technical expertise includes Neural Computer Graphics and Machine Translation.