• Subscribe

The (smart) cookie machine

Speed read
  • Google Vizier service optimizes other machine-learning models
  • Experimental technology tested on real cookie recipes
  • Goal for Vizier is to solve AI problems with limited human intervention

What makes your grandmother’s chocolate chip cookies the best in the world? Is it the special butter she uses? The dash of cinnamon? Or is there some unexplainable technique that she’s learned over decades of experience?

<strong>The secret ingredient</strong> in this cookie recipe is a computer that trains other computers. Courtesy Google.Humans have an innate ability to acquire knowledge over our lifetimes — it’s one of the many aspects that separate us from machines. However, a recent experiment from Google shows that this quality may not remain unique to humans much longer. In fact, the Google Vizier platform used machine learning to make delicious cookies.

We talked with Google senior staff software engineer Daniel Golovin about this experiment. He told us that it all started with his team’s interest in active learning, which is how a machine acquires the data that help it make the right decisions in an assigned scenario.

According to Golovin, you can think of this like the game “20 questions” where one person imagines an object and another person asks questions until they can guess what it is. The machine proposes designs (or questions) to a rater (the person imagining the object.) The rater gives the design a numeric score indicating its quality, thereby allowing the machine to learn what a good design is.

This is all a part of what’s referred to as black-box optimization. In computing, a black box is any machine whose function is not understandable by humans. Therefore, the most efficient way to optimize such a system is by using a separate machine that is understood to play “20 questions” with the black box.

<strong>Real-world experimentation</strong>. Computers may have developed the recipes, but humans still had to bake them. And taste test them. Courtesy Google.“More broadly, you can think about this research area as working to imbue these systems with a form of curiosity,” Golovin said.

Since a major goal of Google Vizier is to optimize physical processes such as airfoil design, the team decided that a fun real-world experiment would be to design the best chocolate chip cookie possible. Golovin and his research team created a machine that could devise cookie recipes while also learning from its mistakes.

A research paper written on the subject by Google discussed the necessary parameters for the project, which included ingredients such as brown sugar, butter, vanilla, chip type and quantity, salt, etc. Baking time and temperature were also important. It’s impossible to make a cookie without these items, but the question remained as to exactly how much of each should be used.

The researchers also wanted to test a specific feature of Vizier. When setting up an AI algorithm, programmers have to manually tune their black box machines on a trial-and-error basis until they get the result they want. This is an inefficient method. Google Vizier, on the other hand, uses a second black box machine to do these adjustments itself in a process called hyperparameter tuning.  

<strong>Moment of truth.</strong> John Karbowski, Jeanette Harris and Daniel Golovin sample their smart cookies. Courtesy Google.While parameters are values derived from the machine learning process, hyperparameters are values that are generally set before a machine can start its learning.

One easy-to-understand hyperparameter is the rate at which a machine learning system “abandons old beliefs for new ones.” If this rate is too low, the machine won’t pick up vital patterns. If it’s too high, the machine will return coincidences that aren’t relevant. Until now, the best way to tune hyperparameters like this was to do so manually.

To test automated hyperparameter tuning, Google turned to the cookie. Vizier’s job was to create recipes, after which caterers baked them for tens of thousands of Google employees. The workers then filled out a survey about the cookie to provide feedback to the machine about what worked and what didn’t. In doing so, the machine could readjust hyperparameters based on which cookies scored highest.

<strong>The tastiest of them all.</strong> Google employees filled out a survey about the cookies to provide feedback to the machine about what worked and what didn’t, providing a fluid dynamic between humans and machines to solve a problem. Courtesy Clem Onojeghuo/Unsplash.Although this may sound like an excuse to scarf down tasty treats, Google actually proved Vizier’s viability with this experiment. The system helped avoid impractical trials –such as low amounts of butter making the dough crumbly – and it also proved Vizier’s aptitude for transfer learning. A major step forward for the industry, transfer learning allows a machine to use data and outcomes from previous studies to more quickly process a current study.

According to Golovin, the most important insight he gleaned from this experiment was the usefulness of having a fluid dynamic between human and machines when finding solutions to problems. He says that machines should behave like a student asking questions and proposing solutions, guided by a human mentor. Automated hyperparameter tuning allows for this. 

While the idea of creating the perfect cookie may not seem world-changing, the possibilities of a system like Google Vizier are far-reaching. By using computers to solve problems posed by computers, we increase not only what we can do, but how fast we can do it. 

Join the conversation

Do you have story ideas or something to contribute? Let us know!

Copyright © 2018 Science Node ™  |  Privacy Notice  |  Sitemap

Disclaimer: While Science Node ™ does its best to provide complete and up-to-date information, it does not warrant that the information is error-free and disclaims all liability with respect to results from the use of the information.

Republish

We encourage you to republish this article online and in print, it’s free under our creative commons attribution license, but please follow some simple guidelines:
  1. You have to credit our authors.
  2. You have to credit ScienceNode.org — where possible include our logo with a link back to the original article.
  3. You can simply run the first few lines of the article and then add: “Read the full article on ScienceNode.org” containing a link back to the original article.
  4. The easiest way to get the article on your site is to embed the code below.