Seminar: Signal Processing and Systems

ECE Women Community

Domain Adaptation Using Multi-Kernel Matching

Date: November,30,2022 Start Time: 13:30 - 14:30 Add to:
Lecturer: Tamir Baruch Yampolsky
Affiliations: The Andrew and Erna Viterbi Faculty of Electrical & Computer Engineering

In this research, we address the problem of Domain Adaptation (DA), which is essential in many machine learning and data science tasks. The problem arises when the data sets do not reside in the same domain, which poses a significant challenge to analysis and processing. This can occur, for example, by acquiring data with different equipment, in different environmental conditions, or at different sites.

To make the problem concrete, consider two diffeomorphic spaces, where the diffeomorphism is unknown. Suppose we are given two data sets from these two spaces, termed source set and target set, which are not necessarily of equal size. In addition, suppose we are given a small number of corresponding sourcetarget pairs from the two sets. Our goal is to find an approximation of the diffeomorphism that maps the source set to the target set.

For this purpose, we developed a method based on multi-kernel matching. More specifically, our method is based on the representation of the source and target sets with several local kernels centered in the known corresponding samples. In turn, we match the local kernels and find a mapping between the source and target sets. We measure the performance of our method using downstream classification; that is, attempting to find the labels of the target set using the labels of the source set.

We test our method on simulations and real-world data such as EEG recordings for mental arithmetic identification and single-cell multiomics data, where we show superior or on-par results with the state of the art.

 

* M.Sc. student under the supervision of Professor Ronen Talmon.

 

 

All Seminars
Skip to content