Timezone: »

Workshop
CANCELED: Automated Machine Learning For Networks and Distributed Systems
Behnaz Arzani · Bita Darvish Rouhani

Wed Mar 04 07:00 AM -- 03:30 PM (PST) @ Level 3 Room 10

The first workshop on "Towards A Domain-Customized Automated Machine Learning Framework For Networks and Systems" aims at creating a coalition of researchers who aim to build an AutoML platform for network operators. The platform helps network operators bridge the expertise gap when using ML to solve challenging networking problems.

Researchers at this workshop will discuss how we, as a community, can build a framework that: enables users to use ML to solve problems in networked systems without having in-depth ML expertise and that, similarly, enables ML experts to contribute to solving problems in networked systems without having expertise in these domains. We will discuss whether existing AutoML frameworks, as-is can be used by network operators? If not, what needs to change? Can domain customization help? If yes, what are the components of a domain-customized AutoML framework and how are they different from traditional AutoML solutions? What are the important criteria that such a system needs to meet? What are the techniques we can use to build such a framework? What are the collaborations we can initiate across industry and academia to make headway on solving this problem?

 Wed 7:00 a.m. - 7:15 a.m. Introduction by Workshop Organizers (Talk) Wed 7:15 a.m. - 7:45 a.m. The agony and the ecstasy of machine learning over the Internet (KeyNote: Keith Winstein) » Many networking settings ask us, or really our computers, to make tough decisions from partial information: congestion control, traffic engineering, provisioning, channel characterization, scheduling, query planning, spam filtering, video streaming, predicting real-life indicators from crowdsourced information, etc. This suggests a natural setting for machine learning, which has shown great success in adjacent areas of computer science. And yet -- and yet! The Internet has turned out to be a particularly challenging setting for ML. We don't know how to simulate it, which makes it challenging to learn reliable control algorithms. It's hard to measure loss functions on a distributed network where each node receives only partial information. And we have a hard time learning algorithms that are robust to adversarial input. I'll present findings from two multi-year deployments of ML over the Internet, for congestion control and video streaming, and discuss where I think there's cause for optimism and caution. Wed 7:45 a.m. - 8:30 a.m. Panelists: Victor Bahl (Microsoft Research), Nicolo Fusi (Microsoft Research Cambridge), Ranjita Bhagwan (Microsoft Research India) (Discussion Panel) Wed 9:00 a.m. - 10:00 a.m. Panel, Part 2 (Panel) Wed 10:00 a.m. - 11:30 a.m. Lunch (break) Wed 11:30 a.m. - 12:30 p.m. Accepted Talks (talks) » Taurus: An Intelligent Data Plane Tushar Swamy (Stanford University), Alexander Rucker (Stanford University), Muhammad Shahbaz (Stanford University), and Kunle Olukotun (Stanford University) Automating Botnet Detection with Graph Neural Networks Jiawei Zhou (Harvard University), Zhiying Xu (Harvard University), Alexander Rush (Cornell University), Minlan Yu (Harvard University) Wed 1:00 p.m. - 2:00 p.m. Invited talks (talks) » Daniel Berger (Experience with developing ML for distributed caching systems) Amar Phanishayee (Project Fiddle: Fast & Efficient Infrastructure for Distributed Deep Learning) Wed 2:00 p.m. - 3:00 p.m. Round table discussions (discussion)

#### Author Information

##### behnazarzani Arzani (Microsoft Research)

{Behnaz Arzani} is a senior researcher at Microsoft Research in the Mobility and Networking group. She got her PhD in computer science from the University of Pennsylvania in 2017. She completed her dual masters degree in computer science and electrical engineering at the University of Pennsylvania in the same year. Behnaz leads the effort of using machine learning for data center networking problems at Microsoft Research. She has also built various network diagnosis systems which are currently being used in Microsoft Azure. Through her work she has improved the efficiency of diagnosing problems in the data center by over $10\times$. Her work has been published in influential conferences in networking, e.g. SigComm and NSDI. Behnaz has received various awards throughout her career, these include: the N2Women rising stars in networking, the University of Pennsylvania Rubinoff dissertation award, and the Microsoft Research research collaboration award.