Earlier this month, Apple announced that it would allow its artificial intelligence researchers to publish research papers — a major shift in the notoriously secretive company’s policy. Now, just a few weeks later, the first of these papers has been made public on the preprint server arXiv.
THE PAPER TITLED LEARNING FROM SIMULATED
The paper — titled “Learning from Simulated and Unsupervised Images through Adversarial Training”— deals with intelligent image recognition technology. Specifically, it describes a technique that would enable a program to recognize and decipher computer-generated images.
So far, this has not been possible because, as the researchers from Apple note, “synthetic data is often not realistic enough,” and increasing the realism is “computationally expensive.”
WITH RECENT PROGRESS IN GRAPHICS, IT HAS BECOME MORE TRACTABLE
“With recent progress in graphics, it has become more tractable to train models on synthetic images, potentially avoiding the need for expensive annotations. However, learning from synthetic images may not achieve the desired performance due to a gap between synthetic and real image distributions,” the researchers wrote. “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.”
In order to do so, the researchers used a technique known as adversarial learning, wherein two competing neural networks basically try to outsmart each other. In this case, the two neural networks are the generator, which, as the name suggests, generates realistic images, and the discriminator, whose function is to distinguish between generated and real images.
“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 an S+U learning algorithm – we train our refiner network using an adversarial loss.”
The paper is likely to be the first of many to be authored and published by Apple’s machine learning researchers in the coming weeks and months, now that the tech giant has lifted its restrictions on sharing research. Reports suggest that this decision was taken primarily to recover the ground Apple has lost in recent years to the likes of Google, Microsoft, Facebook and the Elon Musk-backed startup OpenAI, insofar as AI-related research is concerned.