Concerned about Apple's face sync / machine learning approach

[mark-vdw]

Listening to John Gruber's The Talk Show interview with Federighi and Schiller at WWDC 2016, Federighi said:

That is true right now. So each device does its own processessing. In the future, we could share the results of, like, the first one who does the work, just share, just to make that work go along for the ride. But today, it's gonna be each device doing it independently.

A year later when the face sync feature was actually introduced, he said:

But when you say yes to one face, we're actually saying "Okay, well, there are a thousand other photos we saw that we think are the same person as this, so we're going to count those all as John — but all you really told us is about that one photo. 
And so what we actually sync, and support, and so that we get this right — we sync only the specific photos you told us about, and then each… 'cos that's really the truth, y'know, that's what the user really told us.

And then we sync that information, and then each device, because they're independently doing photo recognition are talking that, and re-deriving the — coming to the final result on all your machines to synchronize. So synchronize is actually very little data, and it's all just the truth data, and we won't hold back the right solution.

But getting that architecture completely right was a — took a little longer than we would have liked, and so we're super excited to have it done now.

I'm a bit concerned that that is a VERY different approach to syncing machine learning info about photos, especially if not all devices have access to the same set of photos, because some will have my full library and some only part of it. I was expecting that my Mac, which holds the full library, would do the most detailed analysis, having access to the most data points for running the clustering algorithms and training the neural networks or SVM's or whatever it is that they use. And the iPhones and iPads would then get an easy copy of those results.

If anyone with some more insight in machine learning techniques knows arguments why they chose to only sync ground truth data rather than the analysis results, I would love to hear more.

 

[cswifx]

Privacy?