I am interested in unsupervised scene representation learning and I am focusing on re-usable and re-composable object-centric representations. I also explore possible benefits and applications of the above findings in reinforcement learning.
I spent the summer of 2021 as an Applied Scientist Intern at Amazon. Before starting a Ph.D. in 2018, I worked at IBM Research India. In 2015, I received a Bachelor of Technology degree from Indian Institute of Technology (IIT) Guwahati. I also spent a summer at Korea Advanced Institute of Science and Technology (KAIST) as a research intern in 2014.
- May 2022: Preprint of our new work, STEVE, for unsupervised decomposition of complex and naturalistic videos is out on Arxiv! See our project page here.
- Jan 2022: Our recent work, SLATE, has been accepted at ICLR 2022 and was also presented at the NeurIPS 2021 Workshop on "Controllable Generative Modeling in Language and Vision".
- Oct 2021: Preprint of our new work, SLATE, for building a text-free DALL-E via object-centric learning is out on Arxiv! Our code is released here and the project page here.
- Simple Unsupervised Object-Centric Learning for Complex and Naturalistic Videos
- Illiterate DALL-E Learns to Compose
- Structured World Belief for Reinforcement Learning in POMDP
- Robustifying Sequential Neural Processes
- Jaesik Yoon, Gautam Singh, Sungjin Ahn
- ICML 2020 [pdf]
- SPACE: Unsupervised Object-Oriented Scene Representation via Spatial Attention and Decomposition
- Sequential Neural Processes
- Oct 2021: Rutgers University’s AI Researchers Propose A Slot-Based Autoencoder Architecture, Called SLot Attention TransformEr (SLATE)
- Reviewer: ICML22
- Workshop Reviewer: OSC@ICLR22