WIRED / Deep Dream Generator
Google’s computer scientists have created an AI to tweak the company’s other AIs. The advanced machine learning system, which goes by the faintly sinister name of Google Vizier, automatically tunes algorithms right across Google’s parent company Alphabet. But to test it, the researchers used an old-fashioned metric: cookie quality in the canteen.
Modern machine learning systems are the algorithmic equivalent of Formula 1 racing cars. The systems have tremendous power, but they are extremely sensitive: to function effectively they need to be finely tuned, usually by hand.
Machine learning versus AI: what’s the difference?
In particular, the systems need carefully set ‘hyperparameters’: parameters set in advance that are adapted to the problem at hand. This isn’t easy, because machine learning systems are “black boxes”: even when you’ve made them, you can never be entirely sure how they get their results. One common method for tuning is known in the field as “grad student descent”: basically, you get a graduate student to twiddle the parameters until the algorithm works.
Google Vizier cuts short this tedious manual task by automatically optimising hyperparameters of machine learning models. According to Google’s researchers, the system is already in use across the company.
"Our implementation scales to service the entire hyperparameter tuning workload across Alphabet, which is extensive," they write in a paper released this week, citing an example where Google researchers “used Vizier to perform hyperparameter tuning studies that collectively contained millions of trials for a research project… That research project would not be possible without effective black–box optimisation.”
One process employed in Google Vizier is called ‘transfer learning,’ essentially, learning from experience. Using data from previous studies as a guide, the Vizier algorithm can suggest optimal hyperparameters for new algorithms. The method is most effective when there have been lots of studies in the area, but it also works when there is relatively little crossover: for instance "when the observed metrics and the sampling of the datasets are different".
As well as helping research, Google Vizier is being put to use at Alphabet, where, its creators write, it "has made notable improvements to production models underlying many Google products, resulting in measurably better user experiences for over a billion people".
These improvements include automated A/B testing of features of Google’s websites, including fonts, colours and the format of search results. On Google Maps, for instance, the system is being used to optimise the trade-off between the relevance of a particular search and the distance of that result from the user (with the inevitable aim of getting higher engagement for Google).
‘AI brain scans’ reveal what happens inside machine learning
Google Vizier can also be used to solve black box optimisation problems in the messy physical world. And this is where the cookies come in.
To test their system, the researchers gave cookie recipes to the contractors who make the puddings in Google’s canteen. They taste-tested the result and tracked any alterations the chefs made to improve the taste. Recipes are a kind of algorithm in their own right, with similar black box properties (because you never exactly know why your bake went wrong). And this test allowed the researchers to try out their transfer learning approach:
“Before starting to bake at large scale, we baked some recipes in a smaller scale run-through,” they write. “This provided useful data that we could transfer learn from when baking at scale.”
Even when things went slightly wrong – as, for instance, when the dough was allowed to sit longer, which "unexpectedly, and somewhat dramatically, increased the subjective spiciness of the cookies for trials that involved cayenne" – the schema was able to respond. After a number of rounds, the cookies improved significantly, the researchers say:
“Later rounds were extremely well-rated and, in the authors’ opinions, delicious."
HT: The weekly Import AI newsletter from Jack Clark of Open AI. Essential reading for anyone interested in AI
Content related to: Tech Sessions: Machine Learning