Open Source Vizier: Google’s Blackbox and Hyperparameter Optimization at Scale
Date: Tuesday, Sep 12, 2023, 11:00 - 12:30
- Xingyou Song, Google Deepmind
- Qiuyi Zhang, Google Deepmind
- Sagi Perel, Google Deepmind
Vizier is the de-facto system for blackbox optimization and hyperparameter tuning at Google, having serviced some of Google’s largest products and research efforts. It has tuned millions of machine learning models and regularly services thousands of users over production-critical systems at Google. To operate reliably at such a scale, Vizier required solving key technical challenges in distributed systems, outside of regular algorithm development.
In this hands-on tutorial, I will present Open-Source (OSS) Vizier, a standalone Python-based package based on Google Vizier, to optimize experiments for users and benchmark algorithms for researchers. I will demonstrate how to use OSS Vizier’s distributed system, its numerous integrations with Google’s AutoML ecosystem, and its default Bayesian Optimization algorithm alongside its collection of benchmarks.