NVIDIA has achieved a breakthrough in training AI with a limited dataset. Its latest generative adversarial network (GAN) can learn complex skills such as emulating renowned painters with as little as 1,500 images.
Researchers at NVIDIA applied a new technque called adaptive discriminator augmentation (ADA) to reduce the number of training images by up to 20x to achieve the results.
“These results mean people can use GANs to tackle problems where vast quantities of data are too time-consuming or difficult to obtain. I can’t wait to see what artists, medical experts and researchers use it for,” said David Luebke, Vice President of Graphics Research at NVIDIA.
AI vs AI
GAN works by pitting AI against AI. A generator neural network challenges a discriminator neural network to generate and discern images until the desired outcome is achieved. This pixel by pixel process improves the realism of the synthetic images.
Naturally, the more images the training is based on, the better the results. It typically takes up to 100,000 images to train a high quality GAN. However, it is not always possible to has a large dataset, as in the case of art.
One way to get around this is to use data augementation where existing images are modified, such as via rotating, cropping or flipping, to foce the model to generalise better.
NVIDIA’s ADA method applies data augmentations adaptively, meaning the amount of data augmentation is adjusted at different points in the training process to avoid overfitting. This enables models such as StyleGAN2 to achieve equally amazing results using an order of magnitude fewer training images.
Different editions of StyleGAN have been used by artists to create exhibits and produce a new manga based on the style of legendary illustrator Osamu Tezuka. It’s even been adopted by Adobe to power Photoshop’s new Neural Filters AI tool.
The technique can also be applied in healthcare, where medical images of rare diseases can be limited because most tests come back normal. Amassing a useful dataset of abnormal pathology slides would require many hours of painstaking labelling by medical experts.
Synthetic images created with a GAN using ADA could generate training data for another AI model that helps pathologists or radiologists spot rare conditions.