NVIDIA is closing out 2020 on a strong note with a new method for training GANs that requires significantly less data than current methods. Instead of using hundreds of thousands of images to train efficient GANs with high rates of accuracy, their new technique, adaptive discriminator augmentation (ADA), requires only a few thousand images.
Although NVIDIA’s method requires 10 to 20 times fewer images to train GANs, the time spent training the networks was not significantly reduced compared to methods that use more data. With tens of thousands, if not hundreds of thousands, fewer images required to be collected and labeled, however, there is still a considerable amount of time saved. As a result, this method speeds up the total process of compiling the data and training highly accurate GANs.
The primary obstacle that NVIDIA had to overcome when utilizing so few images to train their GANs was overfitting. Traditionally, a training set that consisted of only a few thousand images would result in overfitting and a neural network that diverges. However, by augmenting the data the team was able to create additional images on which to train the network. And by limiting the ways in which the data was altered, they were able to prevent augmentation leaks from corrupting the generated images.
ADA has the potential to enable the use of GANs in industries and fields that have traditionally not had enough data available to take advantage of this tool. Particularly, medicine, where available images, and the time to label them, are limited resources. Other fields of research where a large repository of images are not available, could also benefit from this technique.
If you are interested in learning more about GANs in their influence on a variety of industries, check out the Ai+ Training Platform. With a monthly subscription, you can watch on-demand training videos about the latest advances in data science topics and tools. For training sessions on essential data science topics and tools check out Dr. Jon Krohn’s 14-part machine learning mini-bootcamp, The Foundations for Machine Learning, free with an annual Premium subscription.