End User Narrative: Automated Video and Reinforcement Learning

• Titel: End User Narrative: Automated Video and Reinforcement Learning
• Referent: Dr. Robert Rapoport
• December 5th, 2023 / C40.320 / 12-2 p.m

Large language models go through a socialization process called RLHF (Reinforcement Learning with Human Feedback). In the case of text and still image generation, the reward functions that trainers use are clearer (e.g. broadly “understanding” the prompt). Due to the semantic density of sequential images, the reward functions for RLHF in videos are less clear. We will discuss previous attempts in film theory to establish a syntax of images. With this in mind, we will ask what is at stake in current attempts to apply RLHF to generative video.