Master

Computer Games and Games related formats are an essential branch of the media industry with sales exceeding those of the music or the movie industry. In many games, it is necessary to build up a dynamic environment with autonomously acting entities. This comprises any types of mobile objects, non-player characters, computer opponents or the dynamics of the environment itself. To model these elements, techniques from the area of Artificial Intelligence allow for modelling adaptive environments with interesting dynamics. From the point of view of AI Research, games currently provide multiple environments which allow to develop breakthrough technology in Artificial Intelligence and Deep Learning. Projects like OpenAIGym, AlphaGo, OpenAI5 or Alpha-Star earned a lot of attention in the AI research community as well as in the broad public. The reason for the importance of games for developing autonomous systems is that games provide environments usually allowing fast throughputs and provide clearly defined tasks for a learning agent to accomplish. The lecture provides an overview of techniques for building up environment engines and making these suitable for largescale, high-throughput games and simulations. Furthermore, we will discuss the foundations of modelling agent behaviour and how to evaluate it in deterministic and non-deterministic settings. Based on this formalisms, we will discuss how to analyse and predict agent or player behaviour. Finally, we will introduce various techniques for optimizing agent behaviour such as sequential planning and reinforcement learning.

This lecture focuses on deep learning approaches in computer vision with a particular emphasis on generative approaches that not only analyze, but in particular synthesize novel images and video.

Modern deep learning has fundamentally changed artificial intelligence. Computer vision was at the forefront of many of these developments and has tremendously benefited over the last decade from this progress. Novel applications as well as significant improvements to old problems continue to appear at a staggering rate. Especially the areas of image and video synthesis and understanding have seen previously unthinkable improvements – and provided astounding visual results with wide-ranging implications (trustworthiness of AI, deep fakes).

We will discuss how a computer can learn to understand images and videos based on deep neural networks. The lecture will briefly review the necessary foundations of deep learning and computer vision and then cover the latest works from this quickly developing field. The practical exercises that accompany this course will provide hands-on experience and allow attendees to practice while building and experimenting with powerful image generation architectures.

Topics include but are not limited to:

• Image & video synthesis

• Visual superresolution and Image completion

• Artistic style transfer

• Interpretability, trustworthyness of deep models

• Self-supervised learning

• Modern deep learning approaches, such as transformers and self-attention, invertible neural networks, diffusion models, etc.
  
  Registration here on Moodle
  Einschreibeschlüssel/Registration Key: Neural Networks