Title: Interpretable and Explainable Deep Learning for Video
Promoter: Prof. Nikolaos Deligiannis

Deep learning has achieved outstanding performance in many signal processing tasks, which also include video-related tasks. However, deep networks are often considered as black-box models, due to their large size and high complexity. This lack of interpretability may be an obstacle to the deployment of deep learning in critical applications requiring trust, algorithm transparency, or performance guarantees. The field of interpretable and explainable deep learning aims to tackle this problem; for instance, by finding novel ways to design interpretable-by-design deep networks, by developing techniques to visualize the features learned by the model, and by explaining a particular model decision through highlighting the decisive features of the input.

My thesis focuses on these interpretability/explainability aspects, applied to the case of video deep learning. More precisely, we will design novel deep networks to learn representations of video using the deep-unfolding framework, which is a technique for designing interpretable-by-design neural networks that are based on existing optimization algorithms. Understanding how these models extract meaningful information in the data will be carried out by explaining the model decisions both using the visual content, the temporal features, and the dynamics captured in the video.

By this research, we participate to the effort of integrating efficient and trustworthy deep-learning solutions in video-related tasks, which include classifying video or objects in video (e.g., anomaly detection, environment understanding for automated vehicles, …), and video reconstruction in compression systems.