2

An article claims that a neural network was trained with over a billion images. Assume that the majority of images contain a face.

The infrared camera takes an image and the dot projector uses around 30,000 IR dots to create a 3D dot pattern of your face. This data is sent to the the new A11 Bionic chip processor, trained with a neural network of over a billion images.

What organizations besides Facebook has access to a billion facial images?

forest
  • 64,616
  • 20
  • 206
  • 257
gatorback
  • 1,541
  • 2
  • 12
  • 17
  • I suppose they could have used 1 billion images in the form of short videos of people. There is evidence of [Apple doing machine learning with pictures](https://www.theverge.com/2017/10/30/16575600/apple-iphone-photos-brassiere-machine-learning), it's not impossible that they've taken the models generated locally and combined them somehow? – JonRB Apr 26 '18 at 22:42
  • I don't think 1B images are necessarily 1B images of faces only. You need negative training data as well so that the network can tell apart faces from non-faces. – Lie Ryan Jul 26 '18 at 04:18

1 Answers1

0

Almost certainly Google. There are probably others. At the end of the day, you can just make a crawler and scrape images of the web. It is not necessarily 100% legal (grey area IMO), but it is not something anyone could prove in court anyway.

Peter Harmann
  • 7,728
  • 5
  • 20
  • 28
  • I don't think you can crawl the web the to find pictures of faces to train picture of faces. All that will do is overfit the training data with the search engine's existing recognition algorithm. Getting 1B images is not too difficult, what's difficult is getting images that are suitably and reliably tagged for training purpose. – Lie Ryan Jul 26 '18 at 04:13