Articles: Profiling AI at Facebook, Apple & Google (Backchannel)
Three extended profiles of the teams driving the machine learning bus from Steven Levy
If you’re reading this, you already know that Facebook, Apple and Google are placing massive priority on machine learning. New startup acquisitions, sexy personalization features and research partnerships are really only symptoms of a (justifiable) obsession with AI.
Backchannel’s three profiles explore their commitment in-depth, with a focus on the people evangelizing, teaching and applying ML. That said, these profiles don’t reveal identical ideologies or pursuits, so reading one isn’t necessarily reading them all.
This most recent of the three profiles was published a couple weeks ago, so it’s likely the most accurate snapshot of the three profiles.
The story centers not on AI-rockstar Yann LeCun, but Joaquin Candela’s applied machine learning group. Beginning with applications to Facebook’s ad serving functions, Candela has led the effort to build AI into everything, including Instagram and Messenger.
My big takeaway here is that Facebook is really aiming at a “content understanding engine”. Today Candela categorizes core applications into vision, language, speech and camera effects, but generally, the amount of ML models being used to understand comments, pictures, audio, videos, links, likes and more will soon converge. The key is nailing transfer learning, meaning the ability for models to take past experiences and apply them to somewhat related problems. As I read more, I keep seeing this as a high-priority research area — OpenAI’s Universe is a good example of facilitating this general problem solving ability in gaming.
At Apple, it’s also clear that AI’s been applied everywhere, but the feeling here is more ‘measured excitement’ than the ‘all-in’ tone at Facebook and Google. Apple really views machine learning as a tool to facilitate their curated product experience, but doesn’t think it’s any more transformative than previous technology breakthroughs:
In Apple’s view, machine learning isn’t the final frontier, despite what other companies say. “It’s not like there weren’t other technologies over the years that have been instrumental in changing the way we interact with devices,” says Cue.
A few other things from this piece:
- Unlike Facebook and Google, internal ML experts are a relatively distributed rather than centralized.
- Apple is sensitive about the perception that its stance on customer privacy materially limits its access to training data. They assert they have plenty of data and sophisticated means of anonymizing personal data that don’t handcuff their machine learning talent.
- Apple prizes researchers with a product focus more than a need to publish papers. It only published it’s first at the end of 2016. Seems like they’re necessarily softening their stance here to remain attractive to talent.
The Google profile is from June of last year, which in AI terms, means it’s practically outdated. I’ve already posted several examples of breakthroughs made since then, but it’s likely still a good articulation of the corporate philosophy and efforts to make Google synonymous with machine learning.
Both Steven Levy Apple and Google articles reference the phrase ‘bear-hugging’ AI — but it’s very clear from this piece and others that Google is hugging with comparably more enthusiasm.
To a great extent, this piece explores their leader’s efforts to make machine learning ubiquitous, creating curricula, research, platforms, robots, media, hardware and internal applications to that end. Google wants to grow the pie and be the unquestioned leader. All indications would say they’re succeeding.
