A New Generation of Trainable Machines for Multi-Task Learning

点击翻译按钮获取中文摘要

The field of Machine Learning plays an increasingly important role in Computer Science and related disciplines. Over the past decade, the availability of powerful desktop computers has opened the door to synergistic interactions between empirical and theoretical studies of Machine Learning, showing thevalue of the ``learning from example'' paradigm in a wide variety of applications. Much effort has been devoted by Machine Learning researchers to the standard single task learning problem and exciting results have been derived. However, Machine Learning capabilities are still extremely limited when compared to those of humans. The human ability to generalise knowledge learned in one task in order to solve a new task is not available in current Machine Learning systems. Multi-task learning research has not yet received sufficient attention in the field. The standard single task learning approach builds on assumptions that are too restrictive to be easily extended to the novel learning scenarios which are envisaged in this proposal. Although interesting insights on multi-task learning have been provided, at present there is no comprehensive framework for multi-task learning and no cornerstone has yet been placed in the field. Thus, the main purpose of this proposal is to develop this area of Machine Learning research. The proposal focuses on Statistical Machine Learning methods for learning multiple related (classification or regression) tasks and integrating information across them. We shall design formal models of relationships between the tasks and develop (learning algorithms) for learning these relationships from data. We shall also develop the mathematical foundations (generalisation bounds, approximation results, convergence results) for multi-task learning, extending some key theoretical results for single tasklearning. Furthermore, the learning algorithms will be applied to two key applications, namely user preference modelling and multiple microarray gene expression data analysis. A central role in our approach is played by certain graph structures which allow us to model task relationships. This approach is very general and can be adapted to increasingly complex learning scenarios. The computational methods are based on the minimisation of certain penalty functionals via a large number of hyper-parameters associated with the tasks. The proposed research will lead to a new generation of trainable machines for multi-task learning, which will be more powerful and flexible models of learning, closer to human learning than previously developed Machine Learning frameworks.