Despite a meteoric rise in data volumes within modern enterprises, enabling nontechnical users to put this data to work in diagnostic and predictive tasks remains a fundamental challenge. Peter Bailis (Sisu | Stanford University) details the lessons learned in building new systems and interfaces to help users quickly and easily leverage the data at their disposal with production experience from Facebook, Microsoft, and the Stanford DAWN project.

Drawing on his research and startup experience, Peter examines why deep networks aren’t a panacea for most organizations’ data; how usability and speed are the best path to better models; why Facebook, Apple, Amazon, Netflix, and Google (FAANG) likely won’t (and can’t) dominate every vertical; and why automating feature selection is more practical than AutoML.

Prerequisite knowledge

• General knowledge of how machine learning models are built, trained, and deployed

What you'll learn

• Learn practical principles informed by recent theory and actual production deployments
• Understand how to use the existing body of structured enterprise data as a source of training data, focus on augmentation and not automation of common workflows, and deploy models quickly and rapidly iterate

This session is from the 2019 O'Reilly Strata Conference in New York, NY.