4 ms·
I think apple is pretty invested in distributed ML training. Ie. your PC will go through your photo gallery/siri audio data, and use the labels and stuff to f
by londons_explore 5y ago
I think apple is pretty invested in distributed ML training. Ie. your PC will go through your photo gallery/siri audio data, and use the labels and stuff to fine-tune an apple-provided ML model. The fine-tunings are then re-uploaded to apple, added up, and distributed to other users in the next update.
The privacy implications of this are dubious. There is as-yet no good research showing how these finetuned models can be used to reveal exactly what photos you have in your photo library, but many experts believe that won't be the case for long.
- secondcoming 5y ago> The privacy implications of this are dubious. Really? Apple can essentially be asked by Law Enforcement for answers to questions like "Give me all the IMEIs of people in location X, radius 10 miles, that appear to have pictures of cats"
- londons_explore 5y agoThe model fine tunings can't immediately be used to say "this person has pictures of cats". They effectively are just a big diff of the ML weights. It's a big set of floating point numbers which is hard to get any meaning out of. But experts believe that techniques might be developed in the future to figure out based on the numbers what data must have been in the training set (and hence your photos). Nobodies managed to do that yet.
- mhitza 5y agoNo experience with NN, just basic understanding of these things. If you train a model A) cats, one for B) dogs and one for C) dogs and cats. Is there any research which asks if you can do a network subtraction C-B = D, and get a sensible network D than can loosely do what A does?
- londons_explore 5y agoTry it... but I predict you'll get a model that doesn't work at all...
- rootusrootus 5y ago> The fine-tunings are then re-uploaded to apple, added up, and distributed to other users in the next update. IMO this is the kind of statement that should have something to back it up.