A new approach to helping machines see mimics human vision and could help overcome the growing problem of adversarial attacks on computer vision. Instead of taking objects as a whole (a collection of pixels), computers are taught to isolate and identify features to come to a conclusion about identity. Photos of birds, for example, were broken down into wing, beak, etc. This also allows the computer to communicate why it decided the bird was, say, a lesser spotted warbler and not a great tit. More here.
Today’s Simple AI – AI made easy