Optimising information fusion in Unsupervised Multiple Kernel Learning Optimising information fusion in Unsupervised Multiple Kernel Learning

Unsupervised Multiple Kernel Learning has proven to be a very useful algorithm for the identification of relevant patterns from complex biomedical data. In short, it uses kernel functions to represent all types of data in a unified manner (kernel matrices) and merges their information in the kernel matrix space. Kernel functions can be interpreted as similarity quantification functions, and kernel matrices can be seen as similarity matrices. The combined similarity information is used to find a simplified similarity-based representation of the data that is useful for the identification of pathophysiological patterns.

Different kernel functions might be more appropriate depending on the type of data at hand. However, the optimal kernel functions for different types of data, and the optimal way to combine the resulting kernel matrices, are not yet well understood. On the other hand, there are alternative fusion schemes that can be explored (e.g. fusing at the simplified representation level).

The objective of this project is to explore different kernel functions and settings, different kernel-matrix combination schemes, as well as different fusion schemes, towards an optimised simplified representation of complex biomedical data. 

Preferred skills: comfortable with Matlab; notions of machine learning.

 

References:

[1] Sanchez-Martinez S, Duchateau N, Erdei T, Fraser AG, Bijnens BH, Piella G. Characterization of myocardial motion patterns by unsupervised multiple kernel learning. Med Image Anal. 2017 Jan;35:70-82. doi: 10.1016/j.media.2016.06.007. Epub 2016 Jun 11. PMID: 27322071.