Understanding Multiple Kernel Learning Methods

Multiple kernel learning has been the subject of nearly a decade of research. Designing and integrating kernels has proven to be an appealing approach to address several, challenging real world applications, involving multiple, heterogeneous data sources in computer vision, bioinformatics, audio processing problems, etc. The goal of this workshop is to step back and evaluate the achievements of multiple kernel learning in the past decade, covering a variety of applications.

In short, this workshop seeks to understand where and how kernel learning is relevant (with respect to accuracy, interpretability, feature selection, etc.), rather than exploring the latest optimization techniques and extension formulations. More specifically, the workshop envisions to discuss the following two questions:

1. Kernel learning vs. kernel design: Does kernel learning offer a practical advantage over the manual design of kernels?

2. Given a set of kernels, what is the optimal way, if any, to combine them (sums, products, learned or non learned, with or without cross-validation)?

We are seeking participants interested in presenting their work and relating their experience in the workshop, providing insight on the above two questions. This includes evidence of MKL improving accuracy beyond any existing method based on single kernels (provided with insights as to why there is such improvement), as well as evidence of the opposite (with insights as to why). We welcome presentation of recent results, as well as presentations based on previously published work that shed light on the above questions.

If you are interested in participating and contributing a presentation, please send the organizers an email with an abstract or a summary. If the presentation is based on previously published work, please include details of such publications.