Visual examination of relations between known classes for deep neural network classifiers

Lukasz Sobczak, Katarzyna Filus, Marzena Halama, Joanna Domanska

Abstract: Classification is the key task of deep learning. Among others, it is used in computer vision for object recognition, and in natural language processing for the masked language modeling. All of these tasks are crucial for modern automatic and semi-automatic production facilities powered by Automated Guided Vehicles (AGVs), because they enable intelligent inspection, maintenance log analysis and operation control. Especially in such safety-critical domains, in which humans and machines often coexist, it is crucial to provide new methods that could help us understand the operation of these black-boxes. That is why we present a novel visual-analytic methods that can be used to examine how networks perceive relations/similarities between the known classes. Our methods operate solely on the trained models and do not require any data samples. They can also reveal some quality issues in the training datasets and indicate low model accuracy. Our methods are suitable for generic vision and language models, but can also be used in transfer learning scenarios. To empirically validate our approach in such a scenario, we conduct experiments on a state-of-the-art mobile vision model - MobileNetV2 - fine-tuned for vehicle classification. We release a new dataset - UtilityVehicles - featuring images of various vehicles that can occur in industry. The presented use case is a vision-based application for Augmented Reality applications for smartphones and embedded devices for AGVs in automatic and semi-automatic production facilities. We release a GitHub repository with data and code: https://github.com/iitis/UtilityVehicles.