Last week, the tech giant Apple published its first artificial intelligence (AI) research paper covering improving recognition in computer vision systems.
Apple's AI Research
According to Apple Insider, before seeing publication through the Cornell University Library on Dec. 22, the research paper was submitted in mid-November. The paper titled "Learning from Simulated and Unsupervised Images through Adversarial Training," arrives less than a month after the company announced a change in its policy that it would no longer stop employees from publishing artificial intelligence-related research.
Forbes reported that the first Apple public research paper focuses on describing various techniques of training computer vision algorithms to better recognize objects by using computer generated, images. Those leveraging synthetic data are often more efficient compared to training models based solely on real-world images because computer generated images are usually labeled. For instance, real-world images are unknown to the algorithm and need to be described by a human operator, while a synthetic image of a hand is annotated as such.
As an accurate learning set cannot be achieved only through computer-generated content that is sometimes not realistic enough, relying completely on simulated images might yield unsatisfactory results. Apple proposes a system of refining a simulator's output through a method of "Simulated+Unsupervised learning," in order to help bridge the gap.
This particular flavor of Simulated+Unsupervised learning combines in practice some annotated synthetic images with unlabeled real image data. The method of machine learning is based in large part on Generative Adversarial Networks (GANs). This technique applies two competing neural networks against each other, one generator and one discriminator. This way is possible to better discern generated data from real data. This is a pretty recent development that has been successful in the generation of photorealistic "super-resolution" images.
Apple applies its modified GAN method to the evaluation of hand poses and gaze estimation. It is unclear yet whether this technique will be applied by Apple in upcoming consumer technology. However, it seems that the company aims to move S+U learning beyond static images to video input.
Some Consequences Of Publishing The Research Paper
Apple's first public AI research paper focusing on computer vision was penned by vision expert Ashish Shrivastava and a team of engineers including Apple Director of Artificial Intelligence Research Josh Susskind, Oncel Tuzel, Tomas Pfister, Russ Webb and Wenda Wang. According to Fortune, the publication of this research paper by Apple is notable. Up to date, scientists around the world have long criticized the high-tech company for not publishing research about artificial intelligence.
In the past Apple had tried to keep secret anything that may be used in future products, previously banning its researchers from publishing their findings. The company's ban was in line with its overall affinity for secrecy, but it contrary to the tradition of sharing knowledge among the scientific community.
Apple's previous policy hurt the company's ability to attract top talent and also made some questions if the tech giant was on the cutting edge of artificial intelligence. Earlier this month, Apple announced in response to these questions that that from now on it would allow its scientists to publish artificial intelligence research papers.
