Six researchers from Apple, working at a recently formed machine learning group, published a paper that describes a method for simulated and unsupervised learning. This news comes after Apple’s Director of AI Research, Russ Salakhutdinov announced earlier this month that Apple would be publishing research.
Apple’s research and development on artificial intelligence were kept in secret all over these years, despite the well-known advances made on the field by big tech companies such as Google. Now everything has changed after Apple decided to publish its first paper on artificial intelligence.
“We develop a method for S+U learning, which we term SimGAN, that refines synthetic images from a simulator using a neural network which we call the ‘refiner network,’” the researchers wrote in the study. “To add realism – the first requirement of a S+U learning algorithm – we train our refiner network using an adversarial loss.”
The paper is titled: Learning from Simulated and Unsupervised Images through Adversarial Training
Earlier this month, Apple announced it would permit its researchers to publish their papers, which is a major change in Apple’s policy. Apple research has always been kept for themselves. And now, the first paper is out. It was made by six researchers at the machine learning research group. It was called “Learning from Simulated and Unsupervised Images through Adversarial Training”.
The paper is about artificial intelligence and it deals specifically with intelligent image recognition technology, describing a technique that would allow a program to recognize computer-generated images. Their aim was to increase the quality of synthetic training images.
The tech-establishment, including Facebook, Microsoft, and Google, has been focusing its efforts to know more about artificial intelligence in the last years, increasing their staff and budget for their machine learning research groups.
Recently, Facebook CEO and Founder Mark Zuckerberg showed off its new AI device, Jarvis, which he adapted for his own house. These companies have tons of publications that are well documented. It seems that Apple wants to be more visible when it comes to Artificial Intelligence technology.
Apple proposes using Generative Adversarial Networks
Synthetic images and videos have been used to train machine learning devices. Generated images are less expensive than real-world imagery. But the generated images don’t have the same quality standards as real images. Therefore, though they are readily available it is risky to use them because small imperfections in synthetic training material can have serious consequences in the final product. This is the technique Apple is trying to improve.
What apple proposes to meet this challenge is to use Generative Adversarial Networks or GANs in order to improve the synthetic training images since GANs take advantage of the adversarial relationship between competing neural networks that try to outsmart one another. They are not something new.
Apple is trying to make some modifications to GANs. Apple wants to use a simulator that generates synthetic images which are going to be pass through a refiner. Then these images are sent to a discriminator that can distinguish between real and synthetic images.
Source: International Business Times