Large-scale deep learning training and tuning with Ray at Uber

At Uber, machine learning powers many business decisions to bring about experiences that delight our customers. To empower our product development teams with ML technology, we have built the Michelangelo Machine Learning Platform as an end-to-end solution for standardizing ML application development from exploration to production. In Michelangelo's early years, we centered our technology stack on Apache Spark, a large-scale data processing and classical ML training platform. Recent mass adoption of deep learning by our product teams has driven us to rethink our Spark-based technical architecture. Compared to classical ML techniques such as linear/logistics regression and tree-based methods, deep learning poses significant new infrastructure challenges in compute, network, and storage. In this talk, we will present case studies to highlight some of the specific challenges in large-scale deep learning training, and how we address them by leveraging Ray, a distributed compute platform.
See all Ray Summit content @ anyscale.com/ray-summit-2022