GANs train two competing networks to generate images, but convergence is tricky. GLO takes a simpler approach: optimize a single latent space to synthesize faces and perform arithmetic like adding sunglasses or removing features. The results are low-res, but the technique opens new possibilities.
Dear fellow scholars, this is two minute papers with in many previous episodes we talked about Generative Adversarial Networks, a recent new line in machine learning research with some absolutely fantastic results in a variety of areas. They can synthesize new images of animals, create 3D models from photos, or dream up new products based on our edits of an image. A Generative Adversarial network means that we have two neural networks battling each other in an arms race. The generator network tries to create more and more realistic images and these are passed to the discriminator network which tries to learn the difference between real photographs and fake forged images. During this process, the two neural networks learn and improve together until they become experts at their own craft, and as you can see, the results are fantastic. However, training these networks against each other is anything but roses and sunshine. We don't know if the process converges or if we reach Nash Equilibrium. Nash Equilibrium is a state where both actors believe they have found an optimal strategy while taking into account the other actors possible decisions, and neither of them have any interest in changing their strategy. This is a classical scenario in game theory where two convicted criminals are pondering whether they should snitch on each other without knowing how the other decided to act. If you wish to hear more about the Nash Equilibrium, I've put a link to Khan Academy's video in the description. Make sure to check it out. You'll love it. I I find it highly exciting that there are parallels in AI and game theory. However, the even cooler thing is that here we try to build a system where we don't have to deal with such a situation. This is called Generative Latent Optimization GLO in short, and it is about introducing tricks to do this by only using a generator network. If you have ever read up on font design, you know that it is a highly complex field. However, if we'd like to create a new font type, what we are typically interested in is only a few features like how curvy they are or whether we are dealing with a serif kind of font and simple descriptions like that. The same principle can be applied to human faces, animals and most topics you can imagine. This means that there are many complex concepts that contain a ton of information, most of which can be captured by a simple description with only a few features. This is done by projecting this high dimensional data onto a low dimensional latent space. This latent space helps eliminating adversarial optimization which makes this system much easier to train. And the main selling point is that it still retains the attractive properties of generative adversarial networks. This means that it can synthesize new samples from the learned dataset. If it had learned the concept of birds, it will be able to synthesize new bird species. It can perform continuous interpolation between data points. This means that, for instance, we can produce intermediate states between two chosen furniture types or light fixtures. It is also able to perform simple arithmetic operations between any number of data points. For instance, if A is males with sunglasses, B are males without sunglasses and C are females, then A, B C is going to generate females in sunglasses. It can also do super resolution and much, much more. Make sure to have a look at the paper in the video description now before we go, we shall address the elephant in the room. These images are tiny. Our seasoned fellow scholars know that for generative adversarial networks there are plenty of works on how to synthesize high resolution images with more details. This means that this is a piece of work that opens up exciting new horizons, but it is not to be measured against the 10th follow up work on top of a more established line of research. Two minute papers will be here for you to keep you updated on the progress which is, as we know, staggeringly quick in machine learning research. Don't forget to subscribe and click the bell icon to never miss an episode. Thanks for watching and for your generous support and I'll see you next time. Ra.
