photo taken in Vancouver, Canada

I'm interested in solving fundamental AI problems through theoretical research that leads to real-world solutions.

Currently, my research focuses on learning in systems composed of dynamically and stochastically interacting components. Pease see my selected works for more details.

• PhD Student - Department of Computer Science (present)

• M.S. in Computer Science (2022)

• B.E. in Computer Science (2019)

• Backpropagation is the default learning rule for artificial neural networks and is often treated as the settled credit assignment approach whenever differentiability is available.
• In this work, we revisit this convention through a theoretical lens of sample efficiency. We characterize the conditions under which synthetic gradients can achieve a lower gradient-estimation mean squared error than backpropagation. We construct examples illustrating that this sample efficiency advantage can be arbitrarily large.

• Compound AI systems, comprising multiple interacting components such as LLMs and diffusion models, have demonstrated improvements compared to single models. However, aligning compound AI systems to human preferences is challenging.
• We propose a principled framework (SysDPO) for aligning all components in a compound AI system as a cohesive whole.

• It is challenging to describe complex systems composed of interacting and dynamic components. In particular, it becomes more challenging as these components in the system exhibit objective-driven behaviors.
• This paper develops a formal and elegant description of such systems using a field-theoretic language inspired by physics. The proposed theoretical framework is referred to as intelligent fields.