Because of the limited availability of labelled data and computing resources, it is often challenging to deploy machine learning algorithms to real-world scientific applications. Domain adaptation, which aims to transfer a well-trained model for a specific machine learning task to similar tasks within the same class, offers a viable route to solve scientific machine learning problems of this type. For example, how can the decision strategy for the COVID management plan of City M be adapted to City S? It is necessary to find the optimal plan without repeating the expensive experiments. By casting scientific machine learning tasks into a probabilistic framework, this project will investigate various avenues in applying (optimal) transport methods to address problems in domain adaptation.
This project has two aspects. First is to identify sufficient conditions on the underlying problem and the target learning task (e.g. classification, regression, and decision-making), which allows the application of the transport-based domain adaptation framework in a certified manner. Then, the next aim is to implement high-performance solvers that can address the domain adaptation problems at scale.
Thanh Dat Tran is currently a third-year undergraduate student majoring in Data Science at Monash University. His current interest is in the field of domain adaptation related to machine learning, in which the main focus is to deal with the situation where the source domain of the data used in training machine learning models is different but related to a target domain that the models can be used on. With his upcoming project, he aims to apply optimal transportation theory in many real-word applications to bring the distribution of a labelled source data to be similar to the unlabelled target domain.