Michelangelo, Horovod & Pyro: ML and AI at Uber

At Uber, we ignite opportunity by setting the world in motion. We are increasingly investing in machine learning and artificial intelligence to help build seamless, impactful experiences. Come join us at this meetup to learn about the three frameworks and services Uber has developed to fulfill this vision!

Agenda
6:00PM - Doors Open, Food & Drink
6:30PM - Introduction
6:35PM - Michelangelo (MA) Learners and Transformers - Solutions for Uber's Machine Learning Applications
7:05PM - Horovod: Distributed Deep Learning on Spark
7:35PM - Long-Term Rider Behavior Modeling using Pyro
8:05PM - Q&A and Networkingg

More on the talks:
Michelangelo (MA) Learners and Transformers - Solutions for Uber's Machine Learning Applications [Mingshi Wang]

A Michelangelo (MA) Learner is a workflow authoring framework that allows Uber data scientists, researchers, and engineers to solve complex machine learning problems with customized workflows. An MA Learner provides Python SDK's that make it easy to write MA models on Jupiter notebooks while hiding the underlying complexities of distributing the machine learning jobs to different computing environments. At Uber, machine learning models are represented by pipelines composed of MA transformers. The data preparation and training-- processes that involve one or more estimators-- produce a trained pipeline with MA transformers. The trained pipelines are persisted for subsequent usage by batch and online predictions. In this tech talk, we discuss the architectures of MA learners and transformers.

Horovod: Distributed Deep Learning on Spark [Travis Addair]

Horovod is a distributed training framework for TensorFlow, PyTorch, Keras, and MXNet. Scaling to hundreds of GPUs, Horovod can take your training time from hours to minutes with just a handful of lines added to your existing single-GPU training process. In this talk, we'll introduce the concepts that make Horovod work, and show you how you can make use of Horovod on Spark to add distributed training to your machine learning pipelines.

Long-Term Rider Behavior Modeling using Pyro [Hesen Peng]

Censored time-to-event data (data where the outcome variable is the time until the occurrence of an event of interest) is critical to the proper modeling and understanding of customer engagement on the Uber platform. In this article, we demonstrate an easier way to model censored time-to-event data using Pyro.

Note: We must collect your first and last name for security reasons prior to the event. You will also need to sign an NDA when entering the building.

We are collaborating with Seattle Spark+AI Meetup on this event!
https://www.meetup.com/Seattle-Spark-Meetup/