What if the AI ​​training data is missing? Nvidia: It’s all about “imagination”

NVIDIA researchers use a pair of generated confrontation networks (GANs) and some unsupervised learning to create an image-to-image translation network that reduces artificial intelligence (AI) training time.

In a blog post, the company explained how its GAN is trained on different data sets. They share a “potential spatial hypothesis” that allows images to be passed from one GAN to the next to generate an image.

The company said that the use of GAN was not new in unsupervised learning, but NVIDIA's research produced new results: Under partially cloudy skies, shadows are clearly visible from dense foliage, far more than ever before. .

The benefit of this work allows network training to require less tag data.

Nvidia said: "For autonomous self-driving driving, training data can be captured once and then simulated under various virtual conditions: sunny, cloudy, snowy, raining, nighttime, etc.

NVIDIA showed how winter photographs were “imagined” as a summer day and how a cat's image was used to generate images of lions, tigers and cougars.

NVIDIA is far from being a GPU company focused on games, but is trying to push its hardware to edge computing devices and use artificial intelligence as its tool.

Last week, the company announced an agreement with GE Healthcare to update 500,000 medical imaging devices deployed worldwide through the Revolution Fro ntier CT for better imaging in hospitals.

General Electric said that faster edge computing capabilities will be better in liver lesion detection and kidney lesion characteristics, and it is possible to reduce the number of subsequent appointments and the number of unexplained scans.

In the third quarter, Nvidia announced quarterly revenues of $2.64 billion, with data center sales doubled from last year's sales, and revenues from $240 million reached $501 million. .