Large scare semi-supervised pattern recognition and data mining from images and text

大规模模式识别和数据挖掘技术已经对应用产生了广泛而切实的影响，从改进通用互联网搜索和半自动互联网门户建设到加速生物信息学的发现。模式识别和数据挖掘中的许多问题出现在这样的环境中：相对于大量可用的未标记数据，容易获得少量的标记训练数据。半监督模式识别方法允许标记数据与未标记数据相结合，以提高识别系统的结果。信息提取是一种数据挖掘，旨在从非结构化数据中提取结构化数据库记录。一旦提取了信息，就可以将底层结构化表示用于各种任务，例如增强搜索或后续数据分析。本文提出的研究将发展新的通用半监督信息提取技术。人们对解决模式识别和数据挖掘中的问题也越来越感兴趣，数据挖掘可以从计算机视觉和文本处理技术的紧密耦合中受益。具体的例子包括：互联网规模的图像搜索，具有数千个潜在类别的对象识别以及从生物实验的图像和文本描述中提取信息；然而，许多其他新出现的问题也具有类似的性质。因此，本提案的研究也旨在开发半监督模式识别，数据挖掘和信息提取的原则和有效方法，重点是涉及图像和文本同时处理的问题。虽然这项工作的重点将放在通用算法和技术上，但为了开发和评估技术，我们建议使用具体的场景，通过挖掘网络创建一个超大规模的视觉百科全书，并创建适合生物图像和文本分析的更专业设置的算法。研究将培养出高素质的人才，以开发适合互联网规模和基因组规模处理的技术——这些技能需求量很大。

Large scale pattern recognition and data mining techniques have resulted in broad and tangible impacts on applications ranging from improved general purpose internet search and semi-automated internet portal construction to accelerated discovery in bioinformatics. Many problems in pattern recognition and data mining arise in a setting where it is easy to obtain a small amount of labeled training data relative to the amount of available unlabeled data. Semi-supervised methods for pattern recognition allow labeled data to be combined with unlabeled data to improve the results of a recognition system. Information extraction is a type of data mining that seeks to extract structured database records from unstructured data. Once information is extracted it is possible to use the underlying structured representation for a variety of tasks such as enhancing search or subsequent data analysis. The research proposed here will develop new general purpose semi-supervised information extraction techniques. There is also a growing interest in addressing problems in pattern recognition and data mining can benefit from a tighter coupling of computer vision and text processing techniques. Concrete examples include: internet scale image search, object recognition with thousands of potential categories and the extraction of information from both images and text descriptions of biological experiments; however, many other emerging problems have similar properties. The research in this proposal thus also aims to develop principled and efficient methods for semi-supervised pattern recognition, data mining and information extraction with an emphasis on problems that involve the simultaneous processing of images and text. While the emphasis of this work will be on general purpose algorithms and techniques, to develop and evaluate techniques we propose to use the concrete scenarios of creating an extremely large scale visual encyclopedia by mining the web and creating algorithms also tailored to the more specialized settings of biological image and text analysis. Research will result in training highly qualified personnel in the development of techniques appropriate for internet scale and genome scale processing - skills in high demand.