时间

TIME

2022年10月13日（星期四）上午9:00-10:30

地点

VENUE

#腾讯会议：945-113-200

主讲人

SPEAKER

Chengcheng Wan is a Postdoctoral Scholar at the University of Chicago. She received her Ph.D. in Computer Science from UChicago, advised by Professor Shan Lu. Her research focuses on improving ML-enabled software systems, especially for correctness, performance, and energy efficiency. Her research has led to first-author papers published at top conferences (ICSE*2, USENIX ATC, ICML). She is a recipient of Siebel Scholarship, the Microsoft Research Dissertation Grant, and the ACM SIGSOFT Distinguished Paper Award.

Her Personal website: http://people.cs.uchicago.edu/~cwan/

主题

TITLE

Improving the Correctness, Performance, and Energy Efficiency of Machine Learning Software Systems

摘要

ABSTRACT

An increasing number of software applications adopt machine learning (ML) to solve real-world problems. The offering of ML cloud APIs further eases such adoption. However, to obtain a correct, fast, and energy-efficient ML application, developers still face challenges in the design of three crucial parts: the application context that encloses the ML component, the system environment that the application executes upon, and the ML algorithm.

In this talk, I will present my research on improving ML applications in all these three parts. On the application side, I will first discuss our empirical study that shows how widespread and severe ML API misuse is in open-source ML applications. I will then present our techniques in testing ML applications and detecting ML API misuses. On the ML algorithm side, I will present our nested AnyTime DNN architecture that offers flexible tradeoffs among accuracy, performance, and energy efficiency. On the system side, I will present our runtime scheduler, ALERT, that dynamically configures neural networks and system resources to meet the runtime requirements in accuracy, performance, and energy efficiency.