Explainability via Interactivity: Supporting Nonexperts' Sensemaking of Pretrained CNN

The research introduces a mobile application designed to help non-expert users, specifically design students, understand the decision-making processes of pretrained Convolutional Neural Networks (CNNs). By utilizing Class Activation Mapping (CAM), the tool visualizes the specific image regions that influence a model’s classification, allowing users to interact with their immediate physical surroundings to test the capabilities and limitations of AI models. The primary thesis is that interactive, playful engagement with AI in real-world contexts effectively scaffolds sensemaking for users who lack formal machine learning expertise.

The study was conducted with 30 industrial design students who had no prior experience with machine learning. Methodology included a deployment of the mobile tool within a university course, followed by in-depth interviews with three participants and an affinity diagram analysis of their recorded observations and screenshots. The findings categorize student sensemaking into three distinct levels of cognitive depth: the collection of interesting or humorous classification results to identify general model behavior; the inference of how environmental factors like lighting, perspective, and adjacent objects affect model accuracy; and active, on-the-spot experimentation to test specific hypotheses about model logic.

The results suggest that mobile-based, interactive XAI tools provide a more accessible and intuitive learning experience than traditional desktop-based interfaces. By enabling users to manipulate variables in their own environment, the tool successfully bridges the gap between complex AI architecture and practical, real-world application. This approach demonstrates that providing non-experts with visual feedback on model decision-making fosters a deeper, more nuanced understanding of AI, which is essential for those looking to integrate machine learning into their own professional design practices.