The Good, the Bad, and the Ugly of ML for Networked Systems

Опубликовано 6 сентября 2018, 22:14

We are witnessing a huge surge in efforts and interest in developing machine-learning (ML) based solutions for optimizing large-scale networked systems. With the unprecedented availability of data and computing power, the early evidence on using ML for systems has been promising.

Microsoft has been at the forefront of applying data at massive scales to address pressing issues in large-scale systems. At the same time, there is a growing voice in the systems and networking community for more “principled” solutions. They are concerned that building solutions we don’t fully understand will ultimately come back to bite us.

This session intends to wade into this storm! What are the networked systems problems for which ML-based techniques are appropriate? Does the fact that the network is often a black box, with much uncertainty with regard to its state, make ML more or less appropriate for networked system than for systems in general? What is the right mix of ML and traditional modeling and algorithms? Do we really need a full understanding of solutions as the “traditionalists” insist?

Leading researchers in the field will discuss how cutting-edge ML advances can be applied to networked systems and lay out principles for ML-based networking and systems research in the coming years.

See more at microsoft.com/en-us/research/v...