CAREER: Towards the Design of Models and Systems for Efficient Prediction

职业：设计高效预测的模型和系统

• 批准号: 1846431
• 负责人: Joseph Gonzalez
• 金额: $ 45.06万
• 依托单位: University of California-Berkeley
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2019
• 资助国家: 美国
• 起止时间: 2019-06-15 至 2024-05-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1846431&HistoricalAwards=false
• 关键词: CAREER; Towards; Design; Models; Systems

Advances in artificial intelligence (AI) are enabling new services and applications including voice-enabled digital assistants, machine translation, personalized medicine, fraud detection, and autonomous vehicles. This new class of intelligent services and applications are critical to the economy and transforming century-old industries ranging from commerce and banking to transportation and healthcare. In this research, the problem of deploying advanced machine learning models and prediction pipelines is studied. A holistic approach is taken for the design of the computer systems needed to render predictions and the models and algorithms that they run. Models are developed that can scale their complexity to avoid overthinking and thereby reduce the energy required to make predictions along with algorithms and systems to automatically configure complex prediction pipelines to simplify model deployment while also improving efficiency and reliability. This research will also produce and promote open-source software systems that can be used across industries. Building these new AI-powered applications will also require new skills and education. A key part of this proposal is open-source course development. A new Data Science major is being designed and this grant will fund new open modules in the data science courses. These courses are taken by students in engineering as well as across the entire campus and provide the fundamental skills needed to understand how machine learning can be used to solve real-world problems.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.

人工智能（AI）的进步正在催生新的服务和应用，包括语音数字助理、机器翻译、个性化医疗、欺诈检测和自动驾驶汽车。这种新型智能服务和应用对经济和改变从商业和银行到运输和医疗保健等百年历史的行业至关重要。在本研究中，研究了部署先进的机器学习模型和预测管道的问题。计算机系统的设计需要采用一种整体的方法来呈现预测以及它们运行的模型和算法。开发的模型可以扩展其复杂性，以避免过度思考，从而减少进行预测所需的能量，以及自动配置复杂预测管道的算法和系统，以简化模型部署，同时提高效率和可靠性。这项研究还将产生和推广可跨行业使用的开源软件系统。构建这些新的人工智能应用程序还需要新的技能和教育。这个提议的一个关键部分是开源课程开发。一个新的数据科学专业正在设计中，这笔拨款将资助数据科学课程中的新开放模块。这些课程由工程专业的学生以及整个校园的学生学习，提供了理解如何使用机器学习来解决现实世界问题所需的基本技能。该奖项反映了美国国家科学基金会的法定使命，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

Train Big, Then Compress: Rethinking Model Size for Efficient Training and Inference of Transformers

• 发表时间: 2020-07
• 作者: Zhuohan Li;Eric Wallace;Sheng Shen;Kevin Lin;K. Keutzer;Dan Klein;Joey Gonzalez

D3: a dynamic deadline-driven approach for building autonomous vehicles

• DOI: 10.1145/3492321.3519576
• 发表时间: 2022-03
• 期刊: Proceedings of the Seventeenth European Conference on Computer Systems
• 作者: Ionel Gog;Sukrit Kalra;Peter Schafhalter;Joseph Gonzalez;I. Stoica

Deep Mixture of Experts via Shallow Embedding

• 发表时间: 2018-06
• 作者: Xin Wang;F. Yu;Lisa Dunlap;Yian Ma;Yi-An Ma;Ruth Wang;Azalia Mirhoseini;Trevor Darrell;Joseph Gonzalez

FBNetV2: Differentiable Neural Architecture Search for Spatial and Channel Dimensions

• DOI: 10.1109/cvpr42600.2020.01298
• 发表时间: 2020-04
• 期刊: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)
• 作者: Alvin Wan;Xiaoliang Dai;Peizhao Zhang;Zijian He;Yuandong Tian;Saining Xie;Bichen Wu;Matthew Yu

Accel: A Corrective Fusion Network for Efficient Semantic Segmentation on Video

• DOI: 10.1109/cvpr.2019.00907
• 发表时间: 2018-07
• 期刊: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)
• 作者: Samvit Jain;Xin Wang;Joseph E. Gonzalez