In the verification setting, here's how recognition (could perhaps!) work: - Find the face in the image. This can easily run on a telephone no sweat. Viola and

In the verification setting, here's how recognition (could perhaps!) work: - Find the face in the image. This can easily run on a telephone no sweat. Viola and Jones practically solved this problem in 2001. - Find the parts of the face (eyes, nose, mouth), and then warp the face so it looks like it's facing right at the camera. A recent paper at CVPR last year was called "Face alignment at 3000 FPS," so this is also no sweat at all anymore. - Extract the face's features. This turns the pixels into a smaller vector representation (maybe a list of 1024 numbers perhaps). This is a bit tricky --- many current systems use convolutional neural networks to do this step, which typically take a lot of compute time. Maybe Facebook could shrink them down to a telephone. - Compare that feature vector to your friends' vectors. In the verification setting, you know the "feature vectors" of your 500 friends, so you find the Euclidean distance between the image's vector and the 500 others. No sweat. - To get a recognition result, just pick the closest vector. Or, if the closest friend is too far away, return "No Match". There are many ways to build a good face recognizer. This is one of them. This might or might not be what Facebook's doing. Last year, they published a paper called "DeepFace" which works essentially this way but with some extra tricks. (The paper is here, but the focus was on comparing to academic state-of-the-art on a strictly academic dataset. It's totally marketing / PhD recruiting bait. :P I'm sure their internal system they have now must be far better. https://research.facebook.com/publications/480567225376225/deepface-closing-the-gap-to-human-level-performance-in-face-verification/ https://research.facebook.com/publications/480567225376225/d... ) The AI and ML people at Facebook are darn smart. I would be impressed (but probably not surprised) if the wizards found some clever way to shrink the pipeline to fit on a phone without too much accuracy loss. So it's possible. I have no clue though.