Overview:
Since labeling data is very expensive, various machine learning algorithms are designed to use labeled and unlabeled data efficiently. An active area of research over the last few years has focused on attempts to exploit data from multiple tasks or domains in order to improve learning performance on a target task. Algorithms for ``multi-task learning'', ``transfer learning'' and ``domain adaptation'' belong to this category. These algorithms are especially important in areas that are naturally divided into many domains. For example, a successful Part-of-Speech (POS) tagger trained on news articles can perform very badly on medical articles. Moreover, it is not possible to annotate a lot of labeled data for all domains. Therefore, the use of labeled data from different tasks is crucial.
In this project, we analyze a special form of learning algorithms which learns multiple tasks together. This type of algorithms assumes that the weight vector of every task has a ``shared'' linear component. This special form of algorithms have been used in the tasks of 'multi-task learning' and 'domain adaptation'. Our goal is to provide theoretical understanding of such algorithms and propose new algorithms with better performance.
