Stanford, California, United States
Graduating Stanford AI PhD student interested in evaluation and RL-based adaptation of large models. I use natural human data as evaluation, rubric, and reward signal for complex model behaviors. Interested in robotic foundation models that enable both dexterity and cross-task generalization to the point of interacting meaningfully with human needs. Check out: - BEHAVIOR: large-scale structured evaluation for embodied AI (https://behavior.stanford.edu/) - ROSETTA: RL reward generation from natural language (sanjanasrivastava.github.io/rosetta-project/) - Offline RL from binary preference, with DeepMind Interactive Agents (https://arxiv.org/pdf/2211.11602)
- Developed imitation and reinforcement learning-based agent for virtual agent in physics-simulated 3D Playhouse environment - Applied offline RL methods established in MuZero to Playhouse agent - Worked with Bradley-Terry binary preference-based reward models in offline RL pipeline
Investigated failure modes of convolutional neural networks (CNNs) in object recognition tasks; established similarities and differences between object recognition models and human visual recognition ability; added human model of visual attention to CNN to reduce required training data and improve performance.