I study how learning systems change during training and why additional data, updates, or model capacity sometimes fail to produce better capabilities. My interests include plasticity loss, optimization instability, gradient interference, and the conditions under which learning can scale.
My approach is to identify bottlenecks, understand how data, architecture, and training choices interact, and design targeted interventions. Deep reinforcement learning provides a foundation for this work; the broader goal is to understand and improve learning mechanisms across agents and learning systems.
My studies of plasticity examine how data, modules, and training stages affect an agent’s ability to keep learning. My work on network sparsity explores how connectivity changes the relationship between model capacity and effective learning. Both pursue a shared question: what enables a learning system to turn available resources into lasting improvements?