Nvidia researchers have used a pair of generative adversarial networks (GANs) along with some unsupervised learning to create an image-to-image translation network that could allow for artificial intelligence (AI) training times to be reduced.
In a blog post, the company explained how its GANs are trained on different data sets, but share a “latent space assumption” that allows for the generation of images by passing the image representation from one GAN to the next.
“The use of GANs isn’t novel in unsupervised learning, but the Nvidia research produced results — with shadows peeking through thick foliage under partly cloudy skies — far ahead of anything seen before,” the company said.
The benefits of this work could allow for network training to require less labelled data, it said.
“For self-driving cars alone, training data could be captured once and then simulated across a variety of virtual conditions: Sunny, cloudy, snowy, rainy, nighttime, etc,” Nvidia said.
Nvidia showed how a picture taken in winter could be “imagined” as a summer’s day, and how an image of a house cat could be used to generate images of lions, tigers, and cougars.
Far from a GPU company focused solely on gaming, Nvidia has been attempting to push its hardware into edge devices and using artificial intelligence as the vehicle for it.
Last week, the company announced that it had struck a deal with GE Healthcare to update 500,000 medical imaging devices deployed across the globe with Revolution Frontier CT, which will allow for better imaging at hospitals.
GE said the speedier Revolution Frontier would be better at liver lesion detection and kidney lesion characterisation, and has the potential to reduce the number of follow-up appointments and the number of non-interpretable scans.
For its third quarter, Nvidia posted record quarterly revenue of $2.64 billion, of which its datacentre segment nearly doubled the $240 million in sales posted at the same time last year, to report $501 million in revenue.