| Human beings are able to transfer knowledge learned in one or more tasks in order to tackle new tasks. For example, the abilities achieved while learning to walk presumably apply when one learns to run, and knowledge acquired while learning to recognize cars could apply when recognizing trucks. In Machine Learning (ML), this process is cast into a ``learning from example'' paradigm known as Multi-task learning (MTL). MTL concerns the study and development of systems that generalize and transfer knowledge from a collection of tasks. The recent availability of more and more real-life data sets from possibly related tasks has recently boosted this research area and its applications. A central issue in MTL is task-relatedness. Current MTL systems are still far from human beings' capability to relate and exploit information acquired from various tasks, mainly due to the fact that tasks can be related in various ways. Very recently, a number of parametric learning approaches for MTL based on different types of task-relatedness have been proposed. However, at present there is no general, non-parametric framework for describing and analyzing task-relatedness in MTL. The goal of this project is then to develop such a framework, for analyzing task-relatedness based on an approach that combines graphs and non-parametric machine learning techniques. The framework will be used for improving generalization performance of multi-task learning algorithms as well as the interpretability of models generated by these algorithms. |
