- This event has passed.
Webinar on Meta Learning : Search for a unified method for remembering the past and predicting the future
Abstract: With the rise of the deep learning era, data became the new oil. However, data is expensive to store and collect. Every day there are millions of images, videos, and texts added to the internet. Our systems/models should be able to adapt to the changes in the new data and able to remember the knowledge from the past. Continual learning addresses one of these problems, by learning new tasks/knowledge while remembering the past. This allows us to keep only the very recent data in our systems. On the other hand, some of the data is very expensive and is not readily available. In these scenarios, few-shot learning helps to quickly adapt to new data distributions with a very limited amount of data. Recently, meta-learning became very popular due to its ability to learn more general representations. In this talk, we will discuss continual learning, few-shot learning, and how both of these methods can be unified via meta-learning.
Bio: Jathushan Rajasegaran received his BSc in Electronics and Telecommunication Engineering from the University of Moratuwa, Sri Lanka in 2018. Following his undergraduate, he spent 2 years at the Inception Institute of Artificial Intelligence, UAE, working on continual learning, meta-learning, and few-shot learning. After that, he joined the University of California Berkeley, and he is currently a first-year Ph.D. student at BAIR. His research interests are mainly in learning methods for deep neural networks including continual learning, meta-learning, and few-shot learning, and 3D vision.