Research Projects
My current interest mainly focuses on deep learning for manifold-structured data, control-based algorithms and building connections between computational mean field games and machine learning theories.
In-Context Learning on the Space of Probability Measures
We formulate optimal transport as an in-context operator-learning problem. We train a single solution operator that maps a pair of probability distributions to their optimal transport map. Given only a small number of samples from a previously unseen source--target pair as context, the model can infer the corresponding transport map without gradient updates.
Training Dynamics and Control-based Training Algorithm for Transformer
We propose an auxiliary-control algorithm based on the Pontryagin’s maximum principle for training a Transformer architecture.
</div>