Advantages of Streaming Architecture in Management of Machine Learning Models
Details
Dr. Ellen Friedman and Dr. Ted Dunning, MapR Technologies
• 6:30 arrive, register, network, food
• 7:00 ACM upcoming events, general announcements
• 7:10 to about 8:30 Presentation
• 9:00 hard stop, all out.
Description:
There is a massive shift underway, a change in the best way to build things, and a key aspect of that change is a widespread and fundamental use of streaming architecture. The advantages of an event-stream-based approach go far beyond just the need to address real-time or low latency applications at huge scale, although that’s a big part of why people work with streaming data. Beyond low latency, additional drivers include the need for flexibility and agility, so a streaming microservices style is rapidly gaining in adoption. In the first part of this talk we will examine the advantages and issues involved in working with streaming data, take a look at emerging technologies for streaming and describe best practices for this style of work, using use world examples from a variety of industries such as telecommunications, manufacturing and web-based services. From there, we will move into a more advanced example from machine learning. One of the common surprises that practitioners face in machine learning is that the logistical difficulties of fielding a machine learning system - such as managing multiple models through evaluation and into production - typically outweigh the actual machine learning itself. The audience will learn how streaming architecture can substantially simplify those logistics, including:
- How to design a rendezvous server to make model deployment easier and more flexible
- How to build a more reliable machine learning system that meets service-level agreements
- How to do better model-to-model comparisons
- How to integrate data engineers and ops into a machine learning team. This presentation will be self-contained and will not require any particular knowledge of streaming architecture or machine learning. We assume that attendees have a basic knowledge of software and familiarity with basic software engineering will be helpful. While accessible, this talk should also be useful for more advanced practitioners of the trade.
Bios:
Ellen Friedman is Principal Technologist for MapR. Her journey has taken her from being a research scientist with a PhD in biochemistry to now being a committer on the Apache Drill and Apache Mahout projects and coauthor of a number of short data books on computer science. Books by Ellen published by O’Reilly Media include Machine Learning Logistics, Streaming Architecture, the Practical Machine Learning series, and Introduction to Apache Flink. Ellen has been an invited speaker at Strata conferences, Big Data London, Big Data Paris, Berlin Buzzwords, Nike Tech Talks, the University of Sheffield Methods Institute in the UK, and NoSQL Matters Barcelona.
Sample blog: https://mapr.com/blog/tensorflow-mxnet-caffe-h2o-which-ml-best/
Ted Dunning is Chief Application Architect for MapR. Since earning his PhD from the University of Sheffield, he has been involved in a series of startups that based on applying machine learning to human behavior. The applications of his work have ranged from ad targeting, to music and video recommendation, to finding fraud and abuse. He has co-authored 10 books on computer science and machine learning and is named as an inventor on over 20 patents. In addition to his work at MapR, Ted serves on the Board of Directors of the Apache Software.
