Abstract
Machine Learning (ML) models are typically trained on the available data for a single target task, for which they are meant to provide a solution. In contrast, multi-task learning (MTL) methods target multiple tasks simultaneously. In this case, the model is trained on the aggregated data from many target tasks and a single model representation is shared by all the tasks. The use of a shared representation enables the machine learning model to generalize better on each individual task as the inductive algorithm may learn shared features that are more robust to the noise present in the individual single-task datasets, thus reducing the risk of bias. MTL is contributing to the advancement of ML and is also generating impactful applications for realworld problems. This chapter presents MTL as a general optimization methodology. The MTL paradigm, the different variants, the internal MTL mechanisms that affect the quality of the ML models, and some specific examples of MTL approaches and applications are discussed.