United States
Research Scientist designing and training generative AI models for 3D. 2nd hire at venture-backed deep tech startup. Training cutting-edge large-scale generative models (diffusion, NeRFs, transformers). Designing, implementing, and iterating on novel data augmentation pipelines for training models on novel tasks using large (10s of TBs) datasets. Part of core team shaping development of R&D and iteration for product-market fit.
Research lead on a project to design and implement novel AI approaches to better capture human learning and decision-making. Extended Theory-Based RL (novel RL framework I introduced in my PhD) to richer domains that more closely capture real-world dynamics; approach outperforms SOTA deep RL models on learning efficiency. Designed novel abstract programming language that captures richness and dynamics of human modeling of artificial worlds (objects, agents, physics, goals). Key contributor to multiple projects building on Theory-based RL: • The Neural Architecture of Theory-based Reinforcement Learning • What is the Model in Model-Based Planning? • Learning to solve complex tasks by growing knowledge culturally across generations (Best Paper award, Neurips 2021 Workshop on Cooperative AI) Managed team of international and domestic researchers in-office and remotely. Reviewer for multiple papers on human-like AI agents for: ICML workshop, ICLR workshop, Cognition, Psychological Review.
Invented Theory-Based RL, a novel reinforcement learning framework that captures human-level learning and planning on complex tasks and outperforms SOTA deep-RL models. • Identified essential limitations of existing RL frameworks and designed novel algorithm inspired by computational and developmental cognitive science. • Invented novel object-oriented exploration algorithm that models human information-seeking. • Designed agent that distills its explicit knowledge about world dynamics to facilitate efficient planning. • Implemented and iterated over dozens of existing RL and planning algorithms, ultimately designing a novel one. • Rigorously evaluated TBRL model against state-of-the-art deep RL models. Coordinated large-scale team across MIT and Harvard. Managed/mentored over 25 students (undergrad, Master’s, PhD) over the course of PhD work. Awarded Angus MacDonald Award for Excellence in Undergraduate Teaching; Marcus Fellowship; Jeffrey and Nancy Hallis Fellowship. Authored papers on human-level reinforcement learning and on computational foundations of human cognitive development • Human-Level Reinforcement Learning through Theory-Based Modeling, Exploration, and Planning • Human Learning in Atari • Hypothesis-Space Constraints in Causal Learning • Information Selection in Noisy Environments with Large Action Spaces