Listen to this story
In 2017, Jeanette Harris’ name was splashed across several newspapers. A baker and the owner of Gluten Free Bakery & Cafe in Pittsburgh produced a batch of chocolate chips and cardamom cookies with a small tweak in the preparation method. Chef John Karbowski from Google’s teaching kitchen invited Harris to take part in an experiment to make a ‘smart cookie’ with a little help from Google’s black-box optimization technology.
This Google system made the testing and refining process of even an incomplete recipe much easier. Harris and Google’s team collaborated on several works and tested multiple batches of cookies; different recipes were tested and ranked by Google employees on a scale of negative three to three. Based on the rating, the scale was adjusted. The result—a perfect cookie. Performing this process manually would have taken a considerable amount of time (there were ten ingredients of different proportions), however, Google’s Vizier cut down the time exponentially.
Introduced in 2017, Google Vizier is the company’s internal service for performing black-box optimization. Over time, this system became the de-facto parameter tuning engine at Google. After five years, Google has open-sourced Vizier.
AIM Daily XO
Join our editors every weekday evening as they steer you through the most significant news of the day, introduce you to fresh perspectives, and provide unexpected moments of joy
Google introduces OSS Vizier
Blackbox optimization is a process of optimising an objective function in a case where the only available information about the objective is the output. Blackbox optimization is applied to a variety of applications like drug discovery, hyperparameter optimization, industrial engineering, and reinforcement learning.
Since its release in 2017, Google’s black box optimization system Vizier has witnessed thousands of monthly users on both the research and production sides at the company. As per Google’s estimates, the tool has run millions of black box optimization tasks since its inception and has helped save a lot of computing and human resources.
Download our Mobile App
In July this year, Google released the Open Source (OSS) Vizier, a standalone Python implementation of Google Vizier’s API. This includes a user API for users to optimise their objective function and a developer API for implementing new optimization algorithms. Both APIs have Remote Procedure Call (RPC) protocols to set up a scalable, fault-tolerant and customisable black box optimization system and Python libraries to abstract corresponding RPC protocols.
Google has developed OSS Vizier as a service. The service architecture specifies a stable API for obtaining suggestions for evaluating and reporting results as Trails, instead of assuming how Trails are evaluated. Users can determine when to evaluate Trails and report back results.
The service architecture can also collect data and metrics over time, and users can track usage patterns to plan the research agenda. Google’s extensive database of runs serves as a valuable dataset for research into multitask transfer learning and meta-learning—allowing users to benefit from the resulting improvement.
When compared to the 2017 version of Google Vizier, the OSS Vizier features has an evolved backend design for algorithm implementations and new functionalities like conditional search and multi-objective optimization.
All said, despite introducing OSS Vizier, Google has announced that it won’t be open-sourcing the default algorithms used in Google Vizier and Cloud Vizier, citing proprietary and legal concerns. Google also cautions that OSS Vizier may not be suitable for all the problems in the entire scope of black box optimization. OSS Vizier may even be a wrong choice for problems that require a large number of parameters and evaluations, like training a large neural network with gradientless methods.
In news for more than one reason
Google Vizier made headlines this year when HR analytics software company Visier Inc sued Google for trademark infringement in the San Francisco federal court. The latter claims that Google Vizier is likely to confuse Visier’s customers. Visier told the US District Court that the similarity of the company and product names could mislead consumers into thinking that Visier is reselling or repackaging Google’s technology.
Visier said it reached out to Google regarding infringement concerns multiple times but was unable to resolve the claims. The company seeks an unspecified amount of money in damages and an order to block Google from using the Vizier name.