SHF: Medium: Program Synthesis for Weak Supervision

SHF：中：弱监督的程序综合

• 批准号: 2106707
• 负责人: Aws Albarghouthi
• 金额: $ 90万
• 依托单位: University of Wisconsin-Madison
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2021
• 资助国家: 美国
• 起止时间: 2021-07-01 至 2025-06-30
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2106707&HistoricalAwards=false
• 关键词: SHF; Medium; Program; Synthesis; Weak

Artificial intelligence, in the form of machine learning, has been transformative in automating difficult tasks and extracting insights in numerous problem domains, including language, vision, and recommendations, and offers many further possibilities. For instance, machine learning holds the promise of automatically analyzing medical images, a task previously reserved for human specialists. However, machine-learning algorithms typically require learning from large amounts of hand-labeled data. Manually labeling data is an expensive and human-intensive process. This project seeks to radically minimize the amount of labeled data needed for machine learning, with the goal of enabling cheap and rapid application of machine learning to new and important domains.The project leverages program synthesis and weak supervision technology to minimize the amount of labeled data needed to build performant models. Weak supervision replaces hand labels with a number of imprecise sources providing a rough signal for supervised training. Such sources are expressed by labeling functions: small, rough programs that encode knowledge about the task at hand. The goal of this project is to have labeling functions be generated automatically using program synthesis, eliminating the manual writing of labeling functions and the need for programming expertise. The project develops a generic language of labeling functions, and explores efficient re-use of synthesized functions and richer means of user interaction to further reduce label requirements. The project is training a diverse group of students. Additionally, the PIs are designing a novel course on machine learning with less labeled data.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.

以机器学习为形式的人工智能在自动化困难任务和提取众多问题领域（包括语言、视觉和建议）的见解方面具有变革性，并提供了许多进一步的可能性。例如，机器学习有望自动分析医学图像，这是以前留给人类专家的任务。然而，机器学习算法通常需要从大量手工标记的数据中学习。手动标记数据是一个昂贵且人力密集的过程。该项目旨在从根本上减少机器学习所需的标记数据量，目标是使机器学习能够廉价快速地应用于新的重要领域。该项目利用程序合成和弱监督技术，最大限度地减少构建高性能模型所需的标记数据量。弱监督用许多不精确的源代替手工标签，为监督训练提供粗略的信号。这些信息源由标记函数表示：小而粗糙的程序，对手头任务的知识进行编码。该项目的目标是使用程序合成自动生成标记函数，消除标记函数的手动编写和编程专业知识的需要。该项目开发了一种通用的标签功能语言，并探索了有效地重用合成功能和更丰富的用户交互手段，以进一步减少标签的要求。该项目正在培训一批不同的学生。此外，PI正在设计一个关于机器学习的新课程，使用更少的标记数据。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值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

AutoWS-Bench-101: Benchmarking Automated Weak Supervision with 100 Labels

• DOI: 10.48550/arxiv.2208.14362
• 发表时间: 2022-08
• 期刊: ArXiv
• 作者: Nicholas Roberts;Xintong Li;Tzu-Heng Huang;Dyah Adila;Spencer Schoenberg;Chengao Liu;Lauren Pick;Haotian Ma;Aws Albarghouthi;Frederic Sala

NAS-Bench-360: Benchmarking Neural Architecture Search on Diverse Tasks

NAS-Bench-360：针对不同任务的神经架构搜索基准测试

• 发表时间: 2022
• 期刊: Advances in neural information processing systems
• 作者: Tu, Renbo;Roberts, Nicholas;Khodak, Mikhail;Shen, Junhong;Sala, Frederic;Talwalkar. Ameet
• 通讯作者: Talwalkar. Ameet

Universalizing Weak Supervision

• 发表时间: 2021-12
• 期刊: ArXiv
• 作者: Changho Shin;Winfred Li;Harit Vishwakarma;Nicholas Roberts;Frederic Sala

Lifting Weak Supervision To Structured Prediction

• DOI: 10.48550/arxiv.2211.13375
• 发表时间: 2022-11
• 期刊: ArXiv
• 作者: Harit Vishwakarma;Nicholas Roberts;Frederic Sala

Generative Modeling Helps Weak Supervision (and Vice Versa)

• DOI: 10.48550/arxiv.2203.12023
• 发表时间: 2022-03
• 期刊: ArXiv
• 作者: Benedikt Boecking;W. Neiswanger;Nicholas Roberts;Stefano Ermon;Frederic Sala;A. Dubrawski