Facebook Uses Artificial Intelligence to Generate Realistic Virtual Imagery
This has been a big few weeks for machine learning and imagery.(See links below). Now Facebook just published new research outlining a novel approach for generating realistic, artificial images of scenery and things like dogs, planes, deer, ships, trucks, horses, and, of course, cats.
I’ve tried parsing through the research, but the details are beyond me. What is interesting, however is their use of an approach called “Generative Adversarial Networks” (GAN). Essentially, what they’ve done is create a kind of feedback loop between two networks, where the first, the “generative network” generates an image from noise. Then the other, “discriminative network,” takes that resulting image, and essentially compares it to training data that is based on real images (note: this is a slight simplification). The result is that with each iteration, the generative network is ‘tricked’ into generating increasingly realistic looking imagery.
The researchers then testing the resulting images with a group of volunteers and found that 40% of the images were realistic enough to fool a human into thinking they are real images.
What is Facebook likely to do with the results of this research? That’s unclear, but with their Oculus Virtual Reality acquisition, it seems reasonable to assume that they are going to need cost-effective methods for generating a massive scale of virtual scenery and objects. Could this research represent early forays into that work?
Deep Generative Image Models using a Laplacian Pyramid of Adversarial Networks
Looking Inside the Image Recognition of Artificial Intelligence:
Is This the First Computational Imagination?
#artificialintelligence #machinelearning #virtualreality #facebook