Researchers from chipmaker NVIDIA have found a way to train a form of artificial intelligence called generative adversarial networks (GANs) using less data, said Venture Beat. The team was able to reduce the required data by 10 to 20 times.
Typically, GANs need over 100,000 images for training. It uses a combination of a generator network and a discriminator network to create videos and images. This technology is used to create artworks and video conferencing.
With the introduction of adaptive discriminator augmentation (ADA), the researchers were able to minimize the needed information to train GANs.
In a paper called “Training Generative Adversarial Networks with Limited Data,” the NVIDIA team was able to solve the issue of training AI with access to a limited amount of data.
In the paper, they said, “The key problem with small data sets is that the discriminator overfits the training examples; its feedback to the generator becomes meaningless and training starts to diverge.”
It added, “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.”
According to David Luebke, VP of graphics research at NVIDIA, data gathering and curation or the ETL (extract, transform, and load) pipeline consumes a significant amount of time when it comes to pragmatic data science.
Luebke told VentureBeat, “That alone takes a huge chunk of pragmatic boots-on-the-ground data science, and we think this [approach] is super helpful because you don’t need nearly as much of that [data] to get useful results.”
Luebke believes that this new data-efficient approach can be more useful when it comes to annotating.
Engadget noted that the ADA approach solves the problem posed by earlier approaches used in creating new images. ADA is able to work around issues such as GANs overfitting images and learning to mimic distortions in images instead of creating new ones.
The introduction of this new approach shows potential in training GANs to spot rare neurological brain disorders. With NVIDIA’s ADA, doctors and researchers could share their findings with the use of images created by artificial intelligence.
Aside from NVIDIA, a team of researchers from Adobe Research, Tsinghua University, and MIT was able to come up with a new approach to augmentation for GANs. More information is set to be released during the NeurIPS conference.
