Honda Research Institute Europe

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).

• A mobile application using Class Activation Maps (CAM) allows non-expert users to visualize which image regions drive a Convolutional Neural Network's (CNN) classification decisions, helping them understand model capabilities and limitations.
• The tool utilizes TensorFlow.js and the MobileNet model, which is pretrained on the ImageNet database to recognize over 1,000 object categories, enabling immediate, installation-free use on mobile devices.
• In a study of 30 industrial design students, the tool facilitated three levels of sensemaking: collecting anecdotal examples of model performance, inferring how environmental factors like lighting and adjacent objects influence predictions, and conducting controlled experiments to test specific hypotheses.