The team has found a way to train Generative Adversarial Networks – a form of AI that pits a generator network against a discriminator network to create images or videos – that significantly speeds up the ETL (extract, transform, and load) pipeline
“The key problem with small datasets is that the discriminator overfits to the training examples; its feedback to the generator becomes meaningless and training starts to diverge. We demonstrate, on several datasets, that good results are now possible using only a few thousand images, often matching StyleGAN2 results with an order of magnitude fewer images.”
Extract from ‘Training Generative Adversarial Networks with Limited Data’ paper
Usually, 100,000-plus images are required to train a GAN, but the new approach, Adaptive Discriminator Augmentation (ADA) can produce results using only 5-10% of the data that would have previously been used. The details of NVIDIA’s research can be found in the Training Generative Adversarial Networks with Limited Data research paper. The paper’s authors believe that removing data constraints in this way will encourage researchers to inspect new imaging data use cases for GANs, particularly in the areas of curation and archiving.
Source: VentureBeat