Email: yilundu [at] mit [dot] edu
I am a third-year PhD student at MIT EECS, advised by Prof. Leslie Kaelbling, Prof. Tomas Lozano-Perez and Prof. Joshua B. Tenenbaum. Previously, I obtained my bachelor's degree from MIT, worked at OpenAI and FAIR, and got a gold medal at the International Biology Olympiad.
I am interested in constructing machine learning tools that enable the development of autonomous embodied agents. As the world is always changing, models must be adapt to out-of-distribution examples at test time and incrementally learn from new experiences. Towards these challenges, my recent research uses the tools of iterative energy-based optimization as a mean to adapt to out-of-distribution samples and a way to construct composable systems which can combinatorially generalize and incrementally learn. Second, models should be able to infer and discover structure across a different modalities such as vision, text, sound and touch, such as the underlying three-dimensional geometry of the world. I am interested in leveraging neural fields as a generic way to discover such rich structure in the world. Finally I'm interested in broader applications of these tools to other domains such as computational biology.
- Check out a list of our work on energy-based models!
- If you are looking for research experience, feel free to reach out -- we have several projects actively looking for additional collaborators.
- Check out our work at NeurIPS on discovering composable concepts from images, representing visual relations in scenes, and modality independent structural discovery!
- Check out our work at ICCV on learning embodied visual representations, learning 3D generative models, learning representations of dynamic scenes and on finding interactions between objects and humans!
- Compositionality: constructing modular / composable models which enable combinatorical generalization, incremental learning, and controllability.
- Perception and Scene Understanding: inferring structured representations of the world for downstream embodied tasks.
- Generative Modeling: constructing models of the world's structure.
- Interactive Learning: building intelligent agents which may interact in the surrounding world.