CAREER: Scaling Approximate Inference and Approximation-Aware Learning

职业：扩展近似推理和近似感知学习

• 批准号: 1553547
• 负责人: Wolfgang Gatterbauer
• 金额: $ 55万
• 依托单位: Carnegie-Mellon University
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2016
• 资助国家: 美国
• 起止时间: 2016-04-01 至 2017-11-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1553547&HistoricalAwards=false
• 关键词: CAREER; Scaling; Approximate; Inference; Approximation

The last decade has seen an enormous increase in our ability to gather and manage large amounts of data; business, healthcare, education, economy, science, and almost every aspect of society are accumulating data at unprecedented levels. The basic premise is that by having more data, even if uncertain and of lower quality, we are also able to make better-informed decisions. To make any decisions, we need to perform "inference" over the data, i.e. to either draw new conclusions, or to find support for existing hypotheses, thus allowing us to favor one course of action over another. However, general reasoning under uncertainty is highly intractable, and many state-of-the-art systems today perform approximate inference by reverting to sampling. Thus for many modern applications (such as information extraction, knowledge aggregation, question-answering systems, computer vision, and machine intelligence), inference is a key bottleneck, and new methods for tractable approximate inference are needed.This project addresses the challenge of scaling inference by generalizing two highly scalable approximate inference methods and complementing them with scalable methods for parameter learning that are "approximation-aware." Thus, instead of treating the (i) learning and the (ii) inference steps separately, this project uses the approximation methods developed for inference also for learning the model. The research hypothesis is that this approach increases the overall end-to-end prediction accuracy while simultaneously increasing scalability. Concretely, the project develops the theory and a set of scalable algorithms and optimization methods for at least the following four sub-problems: (1) approximating general probabilistic conjunctive queries with standard relational databases; (2) learning the probabilities in uncertain databases based on feedback on rankings of output tuples from general queries; (3) approximating the exact probabilistic inference in undirected graphical models with linearized update equations; and (4) complementing the latter with a robust framework for learning linearized potentials from partially labeled data.

在过去的十年里，我们收集和管理大量数据的能力有了巨大的提高。商业、医疗保健、教育、经济、科学以及社会的几乎各个方面都在以前所未有的水平积累数据。基本前提是，通过拥有更多数据，即使数据不确定且质量较低，我们也能够做出更明智的决策。为了做出任何决定，我们需要对数据进行“推理”，即得出新的结论，或者找到对现有假设的支持，从而使我们能够支持一种行动方案而不是另一种行动方案。然而，不确定性下的一般推理非常棘手，当今许多最先进的系统通过恢复采样来执行近似推理。因此，对于许多现代应用（例如信息提取、知识聚合、问答系统、计算机视觉和机器智能）来说，推理是一个关键瓶颈，需要新的易于处理的近似推理方法。该项目通过推广两种高度可扩展的近似推理方法并用“近似感知”的可扩展参数学习方法对其进行补充来解决扩展推理的挑战。因此，该项目不是单独处理 (i) 学习和 (ii) 推理步骤，而是使用为推理开发的近似方法，也用于学习模型。研究假设是，这种方法提高了整体端到端预测准确性，同时提高了可扩展性。具体来说，该项目至少针对以下四个子问题开发了理论和一套可扩展的算法和优化方法：（1）用标准关系数据库逼近一般概率联合查询； （2）根据一般查询输出元组排名的反馈来学习不确定数据库中的概率； (3) 用线性化更新方程逼近无向图模型中的精确概率推理； （4）用一个强大的框架来补充后者，用于从部分标记的数据中学习线性化势。