Collective Machine Learning for Semantic Data Interpretation

本研究将研究集体机器学习算法，该算法将来自异构来源的信息结合起来，自动生成复杂数据的语义解释器。两种新兴趋势推动了这项研究。首先，在大数据时代，有越来越多的免费数据源，这些数据源与任何特定的解释问题相关。尽管这些数据源的大小和注释覆盖范围各不相同，但是可以利用它们的联合来降低在目标解释任务中获得能力所需的注释成本。其次，机器学习的日益成功增加了从学习简单分类模型到适应复杂输出类别的相关语义预测器的雄心。******跨复杂输出空间执行集体学习的主要挑战在于数据源的异质性，即记录的不同输入特征、捕获的不同输出注释和考虑的不同预测任务。为了解决这一基本挑战，本研究将开发新的表征学习算法，以揭示不同数据集和不同输出目标背后的共享结构。******如果成功，该研究计划将克服传统机器学习和数据分析系统的界限，并为异构数据分析提供新的工具，以解决现代大数据背景下的重大需求。此外，本研究还将显著提高语义数据分析系统的自主性和鲁棒性，同时减少甚至在某些情况下消除对人类指导的依赖。******拟议的研究计划既有基础方面的，也有应用方面的，并有望在这两个方面都取得进展。特别是，这项研究不仅将在机器学习和数据分析研究中贡献新的数学和算法发展，还将扩大自动化数据分析系统在更广泛的自然语言处理、计算机视觉、生物信息学、社会和商业数据分析问题中的适用性。由此产生的方法将适用于政府、工业、组织或个人收集的广泛类别的异构数据，并将大大减少在开发有用的数据解释系统时对领域专业知识的依赖。

This research will investigate collective machine learning algorithms that combine information from heterogeneous sources to automatically induce semantic interpreters of complex data. Two emerging trends motivate this research. First, in the era of big data there are an increasing number of freely available data sources that are relevant to any particular interpretation problem. Although such data sources vary in size and annotation coverage, their union can be leveraged to reduce the annotation cost required to achieve competence in a target interpretation task. Second, the growing success of machine learning has increased the ambition to move beyond learning simple classification models to adapting related semantic predictors across complex output categories.******The primary challenge of performing collective learning across complex output spaces lies in the heterogeneity of data sources, in terms of the different input features recorded, different output annotations captured, and different prediction tasks considered. To address this fundamental challenge, this research will develop novel representation learning algorithms that can uncover the shared structure underlying different data sets and different output targets.******If successful, this research program will overcome the boundaries of traditional machine learning and data analysis systems, and provide new tools for heterogeneous data analysis that address a significant need in the modern context of big data. Moreover, this research will also dramatically increase the autonomy and robustness of semantic data analysis systems while reducing, and in some cases eliminating, their dependence on human guidance.******The proposed research program has both fundamental and applied aspects and is expected to contribute progress in both respects. In particular, this research will not only contribute new mathematical and algorithmic developments in machine learning and data analysis research, it will also broaden the applicability of automated data analysis systems to a wider range of natural language processing, computer vision, bioinformatics, social and commercial data analysis problems. The resulting methods will be applicable to broad classes of heterogeneous data collected by governments, industry, organizations or individuals, and will significantly reduce the dependence on domain expertise in developing useful data interpretation systems.