Google used cholate chip okie recipes to train an AI

Google has turned to an unusual source to train its high-tech AI: chocolate chip cookie recipes.

Tasking programmers with helping an AI learn from data through trial-and-error is tedious and time-consuming, so the company has employed a neural network called Vizier to help another neural networks learn via a type of training automation called hyperparameter tuning.

To teach Vizier, Google tasked it with formatting the perfect chocolate chip cookie recipe, considering taste-tester feedback until it for the recipe just right.

Scroll down for video 

Google has turned to an unusual source to train its high-tech AI: chocolate chip cookie recipes. To teach VIzier, Google tasked it with formatting the perfect chocolate chip cookie recipe, considering taste-tester feedback until it for the recipe just right

THE COOKIE TEST

Google tasked its AI Vizer with figuring out the perfect chocolate chip cookie recipe.

The company gave chefs recipes to bake for taste-testers, who then provided feedback via a survey. 

The results were aggregated and sent to Vizier, which would analyze the results to fine-tune the recipe for the next round.

Google found this to be a successful training method for teaching Vizier on how to advise other systems on the best way to do a task. 

'The cookies improved significantly over time; later rounds were extremely well-rated and, in the authors’ opinions, delicious,' reads a paper on the test.  

Google began by providing contractors responsible for providing desserts for Google employees with recipes. 

Ingredients and baking conditions acted as the parameters, allowing only specific temperatures, bake time and following ingredients: baking soda, brown sugar, white sugar, butter, vanilla, egg, flour, chocolate, chip type, salt, cayenne, orange extract.

The head chefs baked the cookies with exact precision, making no alterations to the recipes unless absolutely necessary (and then carefully noting the change).

The chefs would bake the cookie for taste-testers, who would then provide feedback via a survey. 

The results were aggregated and sent to Vizier, which would analyze the results to fine-tune the recipe for the next round - a new round of 'machine learning cookies' was backed and distrusted twice a week for 'several' weeks. 

Google found this to be a successful training method for teaching Vizier on how to advise other systems on the best way to do a task.

'The cookies improved significantly over time; later rounds were extremely well-rated and, in the authors’ opinions, delicious,' reads a paper on the test. 

The architecture of Vizier service, a neural network Google is training to help other neural  networks learn via hyperparameter tuning, is illustrated above

'The cookies improved significantly over time; later rounds were extremely well-rated and, in the authors’ opinions, delicious,' reads a paper on the test

The experiment also gave insight into a few capabilities of Vizier, such as the way it was able to quickly understand the difference between what's needed for different batch sizes and the fact that too little butter would result in a crumbly cookie.

It has already proven to be a valuable platform for research and development, and we expect it will only grow more so as the area of black–box optimization grows in importance,' reads the paper. 

'Also, it designs excellent cookies, which is a very rare capability among computational systems.' 

0 Comment

NO COMMENTS

LEAVE A REPLY

Captcha image