Collaborative Research: PPoSS: Planning: Scalable Systems for Probabilistic Programming

协作研究：PPoSS：规划：概率编程的可扩展系统

• 批准号: 2029016
• 负责人: Tamara Broderick
• 金额: $ 12.5万
• 依托单位: Massachusetts Institute of Technology
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2020
• 资助国家: 美国
• 起止时间: 2020-10-01 至 2022-09-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2029016&HistoricalAwards=false
• 关键词: Collaborative; Research; PPoSS; Planning; Scalable

Statistical methods have had great successes for exploring data, making predictions, and solving problems in a wide range of problems. But in the world of big data, methods need to be scalable, so as to handle larger problems while modeling the real-world problems of messy and nonrepresentative data. The project’s novelties are developments in software and hardware facilitating full-stack integration of Bayesian inference to allow complex and realistic models to be fit to large datasets. The project's impacts are in many areas of pure and applied science, including fields as diverse as epidemiology, genetics, and political science, which are challenging because they are dense in parameters rather than in data. Examples include models for disease progression and drug development, decision making under uncertainty, and trends in public opinion.The project is exploring probabilistic programming, including hardware, high-performance computing, programming languages and compilers, and algorithms. The ultimate goal is to develop the tools necessary for an efficient, and scalable Bayesian workflow, building on the existing success of the open-source probabilistic programming language Stan. The team of researchers on this project are working on explorations of algorithms (model validation for approximate inference), programming languages and compilers (automating of approximate algorithms and advanced performance profiling), systems (probabilistic programming for streaming data), high-performance computing (parallel processing and GPUs), and hardware (exploring domain-specific hardware for Bayesian computation).This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

统计方法在探索数据、预测和解决广泛问题中的问题方面取得了巨大成功。但在大数据的世界里，方法需要具有可扩展性，以便在对杂乱和不具代表性的数据的现实问题进行建模的同时处理更大的问题。该项目的创新之处在于软件和硬件方面的发展，促进了贝叶斯推理的全栈集成，使复杂和现实的模型能够适应大型数据集。该项目的影响涉及纯科学和应用科学的许多领域，包括流行病学、遗传学和政治学等不同领域，这些领域具有挑战性，因为它们的参数密集，而不是数据密集。该项目正在探索概率编程，包括硬件、高性能计算、编程语言和编译器以及算法。最终目标是在开源概率编程语言Stan现有成功的基础上，开发高效、可伸缩的贝叶斯工作流所需的工具。这个项目的研究团队正在探索算法(近似推理的模型验证)、编程语言和编译器(近似算法的自动化和高级性能分析)、系统(流数据的概率编程)、高性能计算(并行处理和GPU)和硬件(探索贝叶斯计算的特定领域硬件)。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

项目成果

Statically bounded-memory delayed sampling for probabilistic streams

概率流的静态有界内存延迟采样

• DOI: 10.1145/3485492
• 发表时间: 2021
• 期刊: Proceedings of the ACM on Programming Languages
• 作者: Atkinson, Eric;Baudart, Guillaume;Mandel, Louis;Yuan, Charles;Carbin, Michael
• 通讯作者: Carbin, Michael

Sparseloop: An Analytical Approach To Sparse Tensor Accelerator Modeling

• DOI: 10.1109/micro56248.2022.00096
• 发表时间: 2022-05
• 期刊: 2022 55th IEEE/ACM International Symposium on Microarchitecture (MICRO)
• 作者: Yannan Nellie Wu;Po-An Tsai;A. Parashar;V. Sze;J. Emer

Semi-symbolic inference for efficient streaming probabilistic programming

• DOI: 10.1145/3563347
• 发表时间: 2022-09
• 期刊: Proceedings of the ACM on Programming Languages
• 作者: Eric Hamilton Atkinson;Charles Yuan;Guillaume Baudart;Louis Mandel;Michael Carbin

Simplifying dependent reductions in the polyhedral model

简化多面体模型中的相关约简

• DOI: 10.1145/3434301
• 发表时间: 2021
• 期刊: Proceedings of the ACM on Programming Languages
• 作者: Yang, Cambridge;Atkinson, Eric;Carbin, Michael
• 通讯作者: Carbin, Michael