Abstract: Motivated by Generative Adversarial Networks (GANs) – a powerful class of generative models – we will investigate a special type of optimisation problems, so-called minimax or saddle point problems.
First, we will address certain issues when using gradient descent, the most popular solution method for many minimisation problems, in a naïve way, without taking into account the specific minimax structure. Next, we will discuss possible remedies to cope with unwanted behaviour of gradient descent and introduce first-order methods with convergence guarantees. Finally, we will talk about various notions of how to reasonably define solutions in different settings as well as methods to obtain such.
Der Link zum Vortrag wird am Donnerstag auf der Website unter www.vsmath.at/academics/phd-colloquia/ veröffentlicht.