Learning with fewer labels in Computer Vision – Amir Bar – TAU

Learning with fewer labels in Computer Vision

Abstract:
In recent years, deep neural networks have transformed the field of Computer Vision. Current neural networks excel at tasks like object detection and segmentation, and in some domains, even exceed human performance. Despite this success, current approaches require massive amounts of labeled data used to supervise the training of deep networks, and such data is costly to obtain. In my work, I propose to alleviate this difficulty by modeling structures in photos and video and show how such inductive biases can lead to models that generalize better with less annotated data. Specifically, I will discuss how modeling objects and learning from unlabeled data can benefit object detection, and how modeling actions can lead to models that generalize better and are more sample efficient for video synthesis.

Bio:
Amir Bar is a second-year Ph.D. student at Tel Aviv University and a visiting student at Berkeley AI Research, advised by Amir Globerson and Trevor Darrell. His research interests span video understanding, self-supervised learning, and machine learning. Most recently, he has been focusing on self-supervised learning.