Apple’s highly secretive and closed door ecosystem means that we are often kept from seeing what folks over at the Cupertino giant are working upon before they actually release a finished product before us. However, that trend is slowly changing. Early this month, Apple announced that its researchers will now make their findings public and along the same, a research paper on intelligent image recognition has now been published.
The paper is titled Learning from Simulated and Unsupervised Images through Adversarial Training and it elaborates over a brand new program that possesses similar image recognition capabilities as Siri Intelligence. You can also think of a more advanced version of the facial recognition techniques that were introduced in iOS 10.
The paper also discusses the advantages and disadvantages of using a simulated image as compared to a real image in training the program. While the latter do offer a more clear picture — no pun intended — to the viewer, the former come automatically labeled and save a lot of humans from having to spend their time on labeling the images. That said though, the program will mostly be spending its time in recognizing real images once its out in the field and as such, Apple is using a combination of real and synthetic images to train its program.
In this paper, we propose Simulated+Unsupervised (S+U) learning, where the goal is to improve the realism of synthetic images from a simulator using unlabeled real data. The improved realism enables the training of better machine learning models on large datasets without any data collection or human annotation effort.
Known as Adversial training methodology, this leads to a highly intelligent image recognizing program.
We show that this enables generation of highly realistic images, which we demonstrate both qualitatively and with a user study.
Apple is likely to deploy this program to improve its systems in the future. Image recognition is a hot topic right now and most corporate entities that deal with AI and intelligent programs, including Google and Facebook, are making efforts in the direction. Apple’s AI, which has been somewhat lagging of late, might just get better after being integrated with this program.
Of course, Apple might never do that and instead use the program separately in its Photos app, to better recognize and label images.
The research team behind the findings of the paper include Ashish Shrivastava, Tomas Pfister, Oncel Tuzel, Josh Susskind, Wenda Wang, and Russ Webb.