Machine learning to understand images and text

点击翻译按钮获取中文摘要

Machine learning research aims to build computer systems that extract useful information from data. Considerable progress has been made in the two main areas of learning. Supervised methods, which rely on having target labels for each input, have proved to be very powerful and widely applicable, particularly on classification problems, in which each input example is assigned to one of a small set of classes. These methods generally scale poorly with the number of training examples and classes. Unsupervised methods utilize unlabeled data, where no targets are provided, and attempt to construct representations useful for many input-output mappings. These methods scale readily but are not as powerful, because without labels considerable structure must be imposed to make the algorithms work. This proposal focuses on the development of learning methods that utilize both labeled and unlabeled examples, addressing two fundamental problems. The first is extracting informative summaries of structured objects, which contain several inter-dependent components. Two very common structured objects are images and documents. A sample task in this domain is scene analysis, which entails picking out multiple items in visual input. For example, a video of a busy street recorded digitally as a stream of color pixel-maps may be re-represented in terms of the cars and pedestrians and their motions. This is an essential and difficult problem, involving a combination of low-level image properties with high-level, object-specific knowledge. The second problem is rapid learning. The human ability to acquire a visual concept from a few examples, and to recognize instances in a complex scene, poses a central challenge to the fields of computer vision, cognitive science, and machine learning. Improving machines' ability to acquire new concepts readily has applications not only to vision, enabling analysis of scenes containing relatively novel objects, but also to the domain of online interaction, as a rapid learning system can build an accurate model of a user with just a few questions, or limited knowledge of the person's web-surfing history or product preferences.