A short explanatory video (Two Minute Papers) describing NVIDIA researchers’ Progressive GAN method for generating high-resolution synthetic celebrity faces. The presenter explains GAN basics, progressive growing of generator and discriminator for more stable training and higher resolution outputs, demonstrates latent-space interpolation and nearest-neighbor checks to detect overfitting, and points viewers to source code, pretrained models, and a Patreon link.
Chapters
Dear fellow scholars, this is two minute papers with hold on to your papers because these results are completely out of this world. You'll soon see why. In this work, high resolution images of imaginary celebrities are generated via a generative adversarial network. This is an architecture where two neural networks battle each other. The generator network is the artist who tries to create convincing real looking images. And the discriminator network, the critic, tries to tell a fake image from a real real one. The artist learns from the feedback of the critic and will improve itself to come up with better quality images. And in the meantime, the critic also develops a sharp eye for fake images. These two adversaries push each other until they are both adept at their tasks. A classical drawback of this architecture is that it is typically extremely slow to train. And these networks are often quite shallow, which means that we get low resolution images that are devoid of sharp details. However, as you can see here, these are high resolution images with tons of details. So how is that possible? So here comes the solution from scientists at Nvidia. Initially, they start out with tiny, shallow neural networks for both the artist and the critic. And as time goes by, both of these neural networks are progressively grown. They get deeper and deeper over time. This way, the training process is more stable than using deeper neural networks. From scratch, it not only generates pictures, but it can also compute high resolution intermediate images via latent space interpolation. It can also learn object categories from a bunch of training data and generate new samples. And if you take a look at the roster of scientists on this project, you will see that they are computer graphics researchers who recently set foot in the world of machine learning. And man, do they know their stuff and how to present a piece of work. And now comes something that is the absolute most important part of the evaluation that should be a must for every single paper in this area. These neural networks were trained on a bunch of images of celebrities and are now generating new ones. However, if all we are shown is a new image, we don't know how close it is to the closest image in the training set. If the network is severely overfitting, it would essentially copy paste samples from there. Like a student in class who hasn't learned a single thing, just memorized the textbook. Actually, what is even worse is that this would mean that the worst learning algorithm that hasn't learned anything but memorized the whole database would look the best. That's not useful knowledge. And here you see the nearest neighbors, the images that are the closest in this database to the newly synthesized images. It shows really well that the AI has learned the concept of a human face extremely well and can synthesize convincing looking new images that are not just copy pasted from the training set. The source code, pre trained network and one hour of imaginary celebrities are also available in the description. Check them out. Premium quality service right there. And if you feel that eight of these videos a month is worth a dollar, please consider supporting us on Patreon. You can also get really cool additional perks like early access and it helps us make better videos grow and tell these incredible stories to a larger audience. Details are available in the description. Thanks for watching and for your generous support and I'll see you next time.
