Domain Adaptation is an area that deals with transfering knowledge from a trained model to a new related situation, which differs particularly in the input representation of the data. In the current medical setting, we are interested in predicting a specific heart signal (regression) for a new individual, based only on ground truth recordings of several other ones. This challenging problem is referd to as Multiple Source Domain Adaptation, and promising algorithms do exist (e.g., [1], often inspired by their single source domain adaptation variants [2]).

While the single source domain adaptation learning setup for pairs of individuals for the current problem has been investigated, the goal of this thesis/project would be to investigate potential benefits due to modelling the problem as a Multiple Source Domain Adaptation problem.

Literature

1. Zhao, Han, et al. “Adversarial multiple source domain adaptation.” https://proceedings.neurips.cc/paper_files/paper/2018/hash/717d8b3d60d9eea997b35b02b6a4e867-Abstract.html
2. Ganin, Yaroslav, et al. “Domain-adversarial training of neural networks.” https://jmlr.org/papers/v17/15-239.html