Thursday, December 7, 2017

Learning With Privacy at Scale

Davey Alba (tweet):

BuzzFeed News interviews with a dozen AI experts paint a picture of Apple’s artificial intelligence research that shows the company is opening up a bit more — but there is still a disconnect between the academic AI community’s values and Apple’s way of doing business. The company’s obsessive focus on the AI applications in Apple products can make working for the company less desirable to some talented experts who have no shortage of options, researchers said. And that’s bad news for Apple, which faces an uphill battle in attracting the people it needs to become a true frontrunner in AI among the giants of tech.


“That blog is completely useless,” an AI professor of an elite university, who asked to remain anonymous because they did not want their name attached to criticisms of an influential tech company, told BuzzFeed News a few weeks ago. “There are absolutely no details, for example, in Apple’s post about AI in handwriting recognition. It amounts to bragging and it is impossible to actually learn anything from it. It feels like they realized most big-name institutions have blogs and created one, but didn't do it in a way that adds any value. I would contrast it with Google’s post about neural networks for language understanding, which has many more details and points to public code along with walkthrough explanations.”


Doubling down on its commitment to privacy, Apple also keeps most user data on the phone itself and deletes it after a few months. But Eugenio Culurciello, a professor at Purdue University who works on machine learning hardware, said that while AI processing on a chip is better than it has been before, limitations on power and memory bandwidth still make a mobile device no match for cloud AI — which is what Google and Amazon use. […] Essentially, Skymind’s Nicholson added, Apple is accepting a commercial disadvantage based on its business model. “AI at Apple is hobbled by the way they handle information,” said Nicholson.

Apple’s Differential Privacy Team:

Given the popularity of emojis across our user base, we want to determine which specific emojis are most used by our customers and the relative distribution of these characters. To that end, we deploy our algorithms to understand the distribution of emojis used across keyboard locales.


Some websites are exceedingly resource-intensive, and we wish to identify these sites in order to ensure a better user experience. We consider two types of domains: those that cause high memory usage and those that cause excessive energy drain from CPU usage. In iOS 11 and macOS High Sierra, Safari can automatically detect these exceptional domains and report them using differential privacy.


We want to learn words that are not present in the lexicons included on the device in order to improve auto-correction. To discover new words, we deploy the Sequence Fragment Puzzle (SFP) algorithm described above.

Previously: Apple’s Machine Learning Journal/Blog, iOS 11 Autocorrect Bug, Why Little Bugs Need to Get Fixed.

2 Comments RSS · Twitter

Something tells me that Sequence Fragment Puzzle is responsible for two recent auto-correction bugs.

Leave a Comment