PAC-Bayes learning and Kernel methods

PAC-贝叶斯学习和核方法

• 批准号: 122405-2011
• 负责人: Marchand, Mario
• 金额: $ 2.11万
• 依托单位: Université Laval
• 依托单位国家: 加拿大
• 项目类别: Discovery Grants Program - Individual
• 财政年份: 2013
• 资助国家: 加拿大
• 起止时间: 2013-01-01 至 2014-12-31
• 项目状态: 已结题
• 来源: https://www.nserc-crsng.gc.ca/ase-oro/Details-Detailles_eng.asp?id=522279
• 关键词: PAC; Bayes; learning; Kernel; methods

Machine learning is concerned with the development of intelligent computer systems that are able to learn and generalize from collected data. This ability is required in many application domains. In computer vision for example, we are interested in learning to recognize a scene (and some objects within the scene) from a collection of annotated images. In bioinformatics, we would like to be able to predict which molecule can bind to a given protein by learning from a database containing several examples of molecule-protein pairs that bind together. In machine translation, we would like to be able to translate text from one language to another by learning this task from a set of text samples that have been translated by humans. Many important applications like these strongly rely on the performance of an underlying learning algorithm for constructing accurate predictors. The long-term objective of this research program is to propose learning algorithms that will meet the requirements of such large-scale applications that can improve our quality of life. To progress towards this goal, we plan to expand our previous work on learning algorithms that optimize a rigorous guarantee on the accuracy of predictors, called a risk bound. More specifically, we will focus on kernel methods and, by extending our previous work on PAC-Bayes sample-compression, we will develop PAC-Bayes bound-minimizing algorithms for learning the kernel. We also plan to propose novel learning algorithms for constructing structured output predictors and other predictors that are more general and potentially more powerful than the support vector machine.

机器学习与智能计算机系统的发展有关，这些系统能够从收集的数据中学习和推广。许多应用程序领域都需要这种能力。例如，在计算机视觉中，我们感兴趣的是从一组带注释的图像中学习识别场景（以及场景中的一些对象）。在生物信息学中，我们希望能够通过从包含几个结合在一起的分子-蛋白质对的数据库中学习来预测哪个分子可以与给定的蛋白质结合。在机器翻译中，我们希望能够通过从一组人类翻译过的文本样本中学习这项任务，将文本从一种语言翻译成另一种语言。像这样的许多重要应用都强烈依赖于底层学习算法的性能来构建准确的预测器。这项研究计划的长期目标是提出能够满足这种大规模应用需求的学习算法，从而提高我们的生活质量。为了实现这一目标，我们计划扩展我们之前在学习算法上的工作，优化对预测器准确性的严格保证，称为风险界限。更具体地说，我们将专注于核方法，通过扩展我们之前在PAC-Bayes样本压缩方面的工作，我们将开发用于学习核的PAC-Bayes边界最小化算法。我们还计划提出新的学习算法，用于构建结构化输出预测器和其他比支持向量机更通用、更强大的预测器。