Jathushan Rajasegaran

I am a Ph.D. student at BAIR advised by Prof. Jitendra Malik. I am also visiting researcher at Meta AI working with with Dr. Christoph Feichtenhofer. I am broadly interested in Computer Vision and Deep Learning, with a focus on developing models for visual understanding in images and videos. Before coming to Berkeley, I was working with Prof. Salman Khan at Inception Institute. I completed my undergraduate study at University of Moratuwa, with a major in Electronic and Telecommunication Engineering. My Bachelor's Thesis was advised by Dr. Ranga Rodigo.

Email  /  Bio  /  Google Scholar  /  Github /  Photos /  Art

Cars don't run like cheetahs, Planes don't fly like birds and Machines won't think in a way same as humans. They will do better. --Richard Feynman

We want AI agents that can discover like we can, not which contain what we have discovered. --Richard Sutton

Machines should be able to understand the world outside our window. That world may change and evolve, but the machines should perceive trees, cars and spaceships not pixels. --Max Wertheimer, Stan Lee

My research interests lie in the general area of computer vision and deep learning, particularly in learning from videos, deep neural architectures and meta/continual learning.

Blog 1: Research Statement (as of June 2023)

We trained LLaMA models up to 1 billion parameters on 1 trillion visual tokens. The resulting model can do diverse tasks from image, video recognition, video tracking, action prediction, and robotics. We also study the scaling properties of these family of models.

We trained a masked autoencoder with 3D Gaussians as intermindiate representaions. This allows the models to do zero-shot figure-ground segmentation, image layering, edge detection, while performing same as supervised finetuning tasks.

Real-world humanoid control as a next token prediction problem, akin to predicting the next word in language.

A fully "transformerized" deisgn for Human Mesh Recovery achieves improved precision and remarkable robustness for 3D human reconstruction and tracking!.

Using 3D human reconstruction and tracking to recognize atomic actions in video.

Performing monocular tracking of people by predicting their appearance, pose and location and in 3D.

Performing monocular tracking of people by lifting them to 3D and then using 3D representations of their appearance, pose and location.

By learning generic represenatations from past tasks, we can easily adapt to new tasks as well as remember old tasks.

We increase the width of a ResNet like model by adding extra skip connections when new tasks are introduced.

Capsule Networks are cool, but they are shallow. We can increase the depth by 3D convolutions and skip connections.

Capsule Networks can capture actual variations that are present in human hand writing, so we generate more data and retrain the capsule networks.

We use content and style representations detect counterfeit apps in playstore.

Website source from Jon Barron here