HUMAN TEACHABLE CONCEPT HIGHLIGHTER FOR POST-HOC VISUAL EXPLANATIONS

Abstract

In recent years, deep learning (DL) models have surpassed human experts in a variety of tasks. However, when it comes to educational contexts, human experts possess something these high-performing models currently lack—the ability to provide clear and approachable explanations tailored to human learners. Efforts to develop explainable methods for these models have typically been designed for DL-experts rather than learners. In this study, we propose a novel approach that combines high-performance models with the teachability of human expert explanations. Specifically, we focus on generating post-hoc, human-understandable explanations for key expert-defined concepts on a novel task: training citizen scientists to identify pollinators from images. Our method, HuTCH, transforms the representational space and highlights relevant segments of an image for learners based on essential concepts — for example, automatically highlighting hair in an image as a key concept for bee identification. We compare HuTCH’s performance by comparing against traditional saliency maps and expert annotations, and show that HuTCH concepts better align with expert annotations. The proposed framework bridges the gap between models’ accuracy and human-teachable features, contributing to the advancement of explainable AI for use in pedagogy.