Introducing students to out-of-distribution detection with deep neural networks

In the past ten years, deep learning has transformed data science, engineering, and even art in countless ways. The key challenge for training engineering students in deep learning is to provide hands-on topics, in which students can apply their problem-solving skills. Out-of-distribution detection is a subtle problem vexing real-world applications, when a deep neural network is faced with an object of a previously unseen class. It provides a challenging playground for students to explore different approaches, applying mathematics and statistics, while ultimately gaining a deeper understanding of deep learning models. In this paper, we review several starting points to explore this application area, including experiments on CIFAR-IO datasets.