Apple has published the first work on AI

Training competitive neural networks based on annotated synthetic and non-annotated real images. Illustration from Apple's first scientific work on AI

Imagine a future in which the artificial intelligence systems of large corporations will compete with each other, earning money for their shareholders. Of course, corporate AI will have to maximize profit. Well, if the shareholders of the company are people and the profits will get them. AI programs can analyze the market, identify the most promising market niches, set the tasks for employees to develop new products. Perhaps the AI ​​can generate products on its own, but it is necessary to check them in public, so that living employees still need corporations to test.

Such a cyberpunk future is getting a little closer.

Google, Facebook, Microsoft and other large corporations are actively developing in the field of weak AI systems with limited functionality. Apple is also leading its development, but so far this company has not published a single scientific work with its characteristic closeness. No one knew what she was working on. And Apple itself lost the most from such secrecy.
')
Now Apple has finally published its first work in the field of artificial intelligence. And let the topic of the first work is not so global - just an improvement in the quality of training for competitive neural networks without a teacher on synthetic images ("Learning from Simulated and Unsupervised Images through Adversarial Training"). What is important is the fact that in the "apple company" there are research and development units that work in this promising area. First, neural networks are trained to create beautiful images, but then they will also solve more complex tasks.

Apple has been working on AI systems for several years. Visual confirmation was the Siri voice assistant, which appeared on the iPhone in 2011 - chatbot with additional features. He looked quite interesting for his time. But over the past years, scientists from Apple have not published a single scientific work .

According to some experts, this closeness is the reason that Apple has lagged behind competitors in this area. It's clear. Having shut down from colleagues, researchers from Apple did not receive feedback, comments, and comments from colleagues that were so necessary for scientific work. No one could use their work and improve it, no one referred to them. After all, after all, the scientific community is a global open environment, where each work relies on a powerful foundation from the many works of colleagues. Here, the isolation method does not work, as in the familiar Apple world of computer ecosystems.

In December 2016, Apple’s policy in this area has finally changed. On December 5, 2016, Russ Salakhutdinov, director of artificial intelligence at Apple, spoke at the Neural Information Processing Systems conference in Barcelona, ​​who took up this position in October after coming from Carnegie Mellon University. He told the public that Apple will now publish their work on AI, as other companies do.


Slide from Salakhutdinov’s presentation on December 5, 2016

Apple hired an AI research director and changed policy, because many cool experts in this field simply refused to work in the Apple Company because of its closed policy. What is the point in a big salary if nobody knows about your work?

To compensate for the problems with hiring scientists, Apple has focused on buying startups. This year, she bought Turi Inc. for $ 200 million and another half a dozen other startups , including Indian Tuplejump Software Pvt Ltd (specializes in fast data mining), Emotient (developing AI systems that recognize and respond to emotions on a person’s face). From Apple you can expect the most "humane" developments in the field of AI, that is, those that interact directly with the person. Recognizing emotions and responding to them, this is what a newly purchased startup does.

Apple's first scientific paper describes technology to improve the quality of neural network learning. Recently, synthetic images are often used for training neural networks (in competitive neural networks, the generator is engaged in generating images, and the discriminator determines their similarity to reality). So it is easier to make the necessary large base for learning neural network So, researchers from Apple have found a way to make such training more effective by adding photorealism to synthetic images using the real photo database and the refiner (enhancer).



Learning neural networks on synthetic images is effective because they are already annotated and labeled. But the neural network trained on them hardly transfers its knowledge to the real world, the view of which is far from synthetic pictures. Apple development solves this problem.

The scientific article was published on December 22, 2016 in open access on the site of preprints arXiv.org (arXiv: 1612.07828). The authors of the research are six Apple employees, including Josh Susskind, co-founder of the emotion-recognizing startup Emotient, which Apple recently bought.