PhD candidate specializing in making machine learning models more efficient without requiring larger architectures or more data.
Hugging Face
Research Award 2023
Developing techniques that improve model performance without increasing computational requirements or data needs.
Creating novel approaches like GRPO that fundamentally change how models learn from data.
Applying research to domains like molecular science and computer vision with measurable improvements.
A novel reinforcement learning framework that achieves superior performance with reduced computational overhead compared to traditional PPO methods. GRPO introduces group-based relative rewards that stabilize training while maintaining sample efficiency.
Compared to standard PPO in Atari benchmarks
More stable training dynamics across environments
Developed specialized architectures for molecular property prediction that outperform traditional methods while using significantly fewer parameters. Our approach combines geometric deep learning with efficient attention mechanisms.
On QM9 benchmark dataset
Compared to baseline models
Introducing GRPO, a novel RL algorithm that achieves superior performance with reduced computational overhead through group-based relative rewards.
A parameter-efficient architecture combining geometric deep learning with attention mechanisms for molecular science applications.