Keynote: Machine Learning & A.I. At Uber
summary
Uber runs a production ML stack called Michelangelo platform that serves online and offline workloads, trains ~36,000 models per month, deploys ~12,000,000 models per month and sustains over a million QPS; core contributions include open source projects ludwig, pyro and a distributed deep system named harvard. Autotune is a Bayesian/model‑based optimizer that tunes thousands of models and yields measurable lifts. Emphasized limitations include data hunger, poor uncertainty quantification, opacity and vulnerability to adversarial examples.
AI-written summary · about
chapters (18)
- 0:00Intro
- 1:00Machine Learning
- 1:14Deep Learning
- 2:43Deep Learning Literature
- 5:32Deep Learning Limitations
- 6:40What is Machine Learning
- 7:33Uber is a Giant Machine Intelligence Problem
- 7:50Ubers Machine Learning Challenges
- 8:48AI Expertise
- 10:02Scale of Use
- 10:41Models
- 11:19Open Source
- 12:54Pyro
- 13:27Bayesian Optimization
- 14:49Autotune
- 15:18Application Area
- 15:52Safety
- 16:08Conclusion
related talks
26mUber’s Intelligent Insights Assistant
19mMichelangelo ML Platform at Uber: Past, Present and Future
25mAccelerating ML at Uber with the Palette Feature Store
33mUber's GenAI Leap: Batch Predictions Using Ray and vLLM
31mUse-Case Driven ML Platforms
38mFantastic ML apps and how to build them
18mAI Platform Q&A
35mUber's Journey in Building Multi-Modal Foundation Models