New machine learning methods for collaborative filtering, document retrieval and image retrieval

这项研究将使用多伦多大学最近开发的新型机器学习算法来改进基于网络的系统的几个方面。协作过滤系统根据用户与其他已知偏好的用户之间的相似性来向用户推荐书籍或DVD等产品。有许多不同的方法来使用这些相似性，新的非线性机器学习方法已经比目前使用的方法更准确。这些方法将得到进一步发展。在检索类似于查询文档的文档时，一种新发现的被称为“语义散列”的方法比现有方法要快得多。语义散列使用机器学习将地址与数据库中的每个文档相关联，从而使相似文档的地址仅相差几位。给定查询文档的地址，通过简单地查看相邻地址可以非常快速地找到类似的文档。这项研究将提高机器学习算法为相似文档分配相似地址的能力。它还将使用类似的方法进行图像检索。这将涉及到学习图像的特征，这些特征对于图像的内容的信息量与文字关于文件内容的信息一样多。将开发新的方法来从未标记的卫星图像的大型数据库中学习特征，以便可以用更少的标记训练样本来学习地面覆盖类型的分类器。通过学习使用附近区域的标签提供的上下文信息，分类器也将得到改进。

The research will use novel machine learning algorithms recently developed at the University of Toronto toimprove several aspects of web-based systems.Collaborative filtering systems recommend products such as books or DVD's to a user based on similaritiesbetween that user and other users whose preferences are known. There are many different ways of using thesesimilarities and new, non-linear machine learning methods are already more accurate than the methodscurrently in use. These methods will be developed further.A newly discovered method called "semantic hashing" is much faster than existing methods at retrievingdocuments that are similar to a query document. Semantic hashing uses machine learning to associate anaddress with each document in the database in such a way that similar documents have addresses that differ byonly a few bits. Given the address of the query document, similar documents can than be found very rapidly bysimply looking at neighboring addresses. The research will improve the ability of the machine learningalgorithms to assign similar addresses to similar documents. It will also apply similar methods to imageretrieval. This will involve learning features of images that are as informative about the content of an image aswords are about the content of a document.New methods will be developed for learning features from large databases of unlabeled satellite images, so thatclassifiers for types of ground cover can be learned with fewer labeled training examples. The classifiers willalso be improved by learning to use the contextual information provided by the labels of nearby regions.