Visual Representations and Learning - 5MMACRV

Goal(s)

Images, videos and more generally visual data in various forms have established themselves as an important means for acquiring and exchanging information, especially on social networks. This type of data however requires the need for dedicated representations since visual content, such as a photograph, has rich as well as ambiguous information. Thus, it is important to develop visual representations adapted to not only the input (image, video, etc.) but also the task to which these representations will be applied. Furthermore, one representation can be used for several complementary tasks, sequentially or simultaneously. The challenge now is to learn representations which take into account all these constraints, and which can evolve over time, if these constraints themselves evolve.

This course, at the intersection of computer vision and machine learning, offers an introduction to different visual representation techniques and their associated models, as well as methods for continuously updating these models. The course will cover in particular: supervised, semi-supervised, and self-supervised learning of visual representations, domain adaptation, approaches to counter the so-called catastrophic forgetting problem. It will mainly focus on natural images and video sequences, but the techniques covered remain valid for other types of visual data.

The course is taught by two researchers who work in computer vision and machine learning at Inria (Karteek Alahari) and NAVER LABS (Diane Larlus). The two instructors will contribute their expertise in academic but also industrial research during the presentation of this course. They will present the latest techniques in the field, drawing on recent research articles.

Content(s)

In particular, the course will cover the following topics:

• Visual data representation learning
• Domain adaptation
• Continual learning
• Action recognition
• Object detection, segmentation

Basics in statistics, computer science (data structures and algorithms), image processing.