Utilizing unlabeled data for machine learning tasks - theoretical analysis

点击翻译按钮获取中文摘要

Mainstream machine learning tools depend on the availability of human annotated training data. Nowaday, in what is called the ``big data" era, applications of machine learning have access to very large amounts of un-annotated (a.k.a. unlabeled) data. Consequently, there is vast interest in designing machine learning tools that can utilize such big pools of raw, unannotated data to reduce the need for human intervention in the learning process. Various recently arising machine learning paradigms address this issue. These include Clustering, Active Learning, Semi-Supervised Learning, Domain Adaptation and Transfer Learning, as well as Learning from ``Weak Teachers" (like supervision obtained via crowdsourcing). To cope with such scenarios, learning practitioners have developed heuristics that, while apparently working reasonably well in practice, are not supported by existing mathematical analyses. In the past couple of decades, machine learning provided a resounding demonstration of the impact of theoretical analysis on the development of practical applications. Algorithmic paradigms like Support-Vector-Machines, Decision-Trees and Boosting grew from theoretical models into popular and vastly applicable software packages. Can the success of theoretical analysis of machine learning be extended to the modern big-data-minimal-human-intervention scenarios? The proposed research aims to provide basis for such developments by building mathematical support for those emerging machine learning and data mining paradigms.   Some examples of recent initiatives taken by my team in that direction include a program that aims to provide tools for guiding users that wish to cluster big data sets on how to choose appropriate clustering algorithms and parameter settings. Such choices are critical to the success of clustering applications, and yet have so far been done in an ad hoc fashion. The PhD thesis of my student Margareta Ackerman took first steps in this direction but there are still big challenges to overcome, both in terms of developing such tools, and in terms of raising the awareness of the data mining community to the significance of the matching between clustering tasks and the algorithms employed to address them. Another project addresses the task of utilizing ``weak supervision".  The use of annotation by novice supervisors to help collect training data for classification prediction tasks has been drawing research attention recently due to the growing popularity of using crowdsourcing. In contrast with much of the current research in this direction, we consider the scenario in which the weak supervision is used in cooperation with human supervision, aiming to reduce  (rather than eliminate) calls to the expert. We are developing novel mathematical models to zoom in on the instances for which novice-generated labels cannot be trusted and need to be scrutinized by a human expert.