Apple’s secret AI sauce gets a new ingredient

By

Apple's secret AI sauce gets a new ingredient
Apple’s special AI creates smarter devices without sharing your data.
Photo: Charlie Sorrel/Cult of Mac

Today’s machine learning AI technologies frequently rely on the free flow of user data. That puts Apple in a tough spot.

While it wants to stay on the cutting edge of artificial intelligence, it doesn’t want to do anything to get in the way of its privacy-first approach to technology. Fortunately, there’s a way around that issue.

In a talk given at this year’s Neural Processing Information Systems (NIPS) conference, Apple’s head of privacy, Julien Freudiger, discussed Apple’s approach. It uses a privacy-preserving machine learning method called federated learning.

Federated learning trains machine learning algorithms on multiple local datasets without exchanging data samples. This allows Apple to do things like get Siri to recognize your voice (and only your voice) as a wake word. More importantly, though, it does this without sharing that data globally. All that gets shared are the updated neural networks, which are used for improving the overall master network. It’s a smart way of ensuring the benefits of data-driven AI, without taking things like the raw audio of Siri requests off devices.

Apple also uses differential privacy. This adds a small amount of noise into raw data so that it’s harder to reverse-engineer audio files from a trained model.

Apple has been employing differential privacy since 2017. It has only recently combined it with federated learning, however. Apple also uses the AI approach for things like its QuickType keyboard.

Apple’s AI approach takes a new leap forward

Apple’s focus on privacy is one that existed under Steve Jobs. However, it has become a more public focus of Apple’s under Tim Cook’s leadership. That’s largely because developments such as large scale hacks and massive data mining has become more widespread.

Tim Cook has also increased Apple’s focus on artificial intelligence. This is an area Apple was previously lagging behind on. In 2017, Apple launched its own semi-academic blog, the Apple Machine Learning Journal. This let it talk about how it uses machine learning to, “help build innovative products for millions of people around the world.”

Apple additionally hired Ian Goodfellow as its new director of machine learning. The former Google AI expert created generative adversarial networks (GANs), one of the most exciting modern AI developments.

Not everything has been positive on the privacy front, though. Earlier this year, it was revealed that Apple contractors had access to recordings of people talking to Siri. These were used to help improve Siri’s functionality. The practice was temporarily stopped, and the contractors let go.

“We realize we haven’t been fully living up to our high ideals, and for that we apologize,” Apple said in a statement.

It seems the company is keen to show how it’s working hard to avoid making that mistake again.

Source: MIT Technology Review