Slashdot Mirror
Apple Publishes Its First AI Research Paper (engadget.com)
When Apple said it would publish its artificial intelligence research, it raised at least a couple of big questions. When would we see the first paper? And would the public data be important, or would the company keep potential trade secrets close to the vest? At last, we have answers. Apple researchers have published their first AI paper, and the findings could clearly be useful for computer vision technology. From a report on Engadget: The paper tackles the problem of teaching AI to recognize objects using simulated images, which are easier to use than photos (since you don't need a human to tag items) but poor for adapting to real-world situations. The trick, Apple says, is to use the increasingly popular technique of pitting neural networks against each other: one network trains itself to improve the realism of simulated images (in this case, using photo examples) until they're good enough to fool a rival "discriminator" network. Ideally, this pre-training would save massive amounts of time and account for hard-to-predict situations that don't always turn up in photos.
The likes of Microsoft, IBM, Google, and more publish amazing research every week. Apple rolls in with one fucking paper and the tech blogs are all covering it like breathless teenagers. Sometimes I really hate my industry.
True that. On IBM Bluemix you can get face detection and train Watson for image classification, and it's free for up to 250 images per day (after that it's about $0.01 per 5 images).
IBM is not only years ahead of Apple on this, they already commoditized visual recognition and they're making money with it.
lucm, indeed.
Why exactly would this discriminator be more authoritative than the original neural net?
The goal is not image detection but rather improving the way computers can generate fake images that could be used to train other computers to recognize objects in images. Basically they want to generate images on-the-fly that are convincing enough to be used in machine learning, that way they can crank up the volume and velocity of training sets without having to deal with the constant issue of managing a gigantic image inventory.
That's why they use two neural nets: one provides images and the other one tries to guess if they're fake or not, that way they can adjust algorithms and whatnot until the second neural net can't tell real from fake images.
lucm, indeed.
What did you expect, a workable AGI on their first paper? It's a decent CV paper. They used GANs (generative adversarial networks) which are state of the art in deep learning today to transform game images into more realistic images. It has at least two interesting applications: to generate cheap, synthetic data for training other CV systems (it's another way to automatically supervise training), and to improve game graphics (for fun).
Did they just download an old MSResearch paper and scratch out "By Microsoft" and crayon "By Apple" over it?
No. They just got a pair of scissors and rounded the page corners. They just hope that everyone looks at the stylish paper and ignores the "By Microsoft".
I know humour is a big tradition around here but you guys should really consider getting some new jokes once in a while.
Machine vision is weak AI by definition.
Nobody is going to stand up and proclaim they are implementing a strong AI. The near future will be a steady match of commercially applicable narrow AI implementagins (digital assistants, autonomous cars, asteroid assayers and miners). Eventually we will find we have backed into strong AI at the overlaps among such systems.