Re:Archivists are uploading hundreds of random VHS tapes to ...
by tialaramex on Sunday March 01, 2020 @11:47PM
(#22461750)
Yes, phones are now powerful enough that if the resulting ML core can be crushed down small enough you can just send that to the phone. My Pixel is set to passively identify every song it hears and display it on the lock screen. No Clown network service nonsense involved, Google built it for other reasons and went "Oh, this would fit on a phone. Cool, might as well".
In iOS 13 (which I am running), it is indeed my phone that does this.
https://www.apple.com/ios/photos/pdf/Photos_Tech_Brief_Sept_... (page 4) says:
> Photos is enabled by powerful machine learning to deliver unique features like Memories, Search Suggestions, and For You. Photos analyzes every photo in a user’s photo library using on-device machine learning that delivers a personalized experience for each user. And this analysis is designed from the ground up with privacy in mind, with all of the processing done on device—and the results of this analysis are not shared with anyone, not even Apple.
> Photos uses on-device processing to analyze each photo and video in a number of ways, including:
> • Scene classification
> Identifies objects, like an airplane or a bike, and scenes, like a cityscape or a zoo, that visually appear in a photo, using a multilabel network with over a thousand classes.
> [...]
The training itself as you point out, happens not on the local device but on the servers that Apple own. So that part you are right about, but that is to be expected. Otherwise, manual tagging by each individual user (as well as significantly more processing power) would be required after all in order to train the models in the first place.