Doctoral Thesis: Towards Robust and General-Purpose Vision via Multiview Contrastive Learning

Yonglong Tian

Abstract:

Representation learning plays a key role in building robust and general-purpose vision learners, and is a long-standing problem. It becomes increasingly interesting with the continuing explosion of data in our era. In recent years, contrastive learning between multiple views of the data has significantly reshaped the field.

In this talk, we start by describing the general framework of multiview contrastive learning. We demonstrate how this simple but generic framework can deal with various representation learning problems. Then we move forward by trying to understand the role of view selection in contrastive learning from an information-theoretic point of view, and come up with an “InfoMin principle”. Such principle is then demonstrated by supervised contrastive learning. In the last part, we will discuss other applications (such as knowledge distillation) and extensions of contrastive learning (e.g., how to apply contrastive learning to uncurated data). 

Details

• Date: Monday, November 14
• Time: 2:30 pm - 4:00 pm
• Category: Thesis Defense
• Location: 32-G449 (Patil/Kiva)

Additional Location Details:

Thesis Supervisor: Prof. Phillip Isola