Hauptinhalt

FB Kolloq SoSe2025: Susanne Trick

„Bayesian Fusion of Probabilistic Forecasts - From Cue In-tegration to Human-Robot Interaction”

Veranstaltungsdaten

Dr. Susanne Trick, Technische Universität Darmstadt
„Bayesian Fusion of Probabilistic Forecasts - From Cue Integration to Hu-man-Robot Interaction”

In order to handle perceptual uncertainty, humans integrate information from multiple sources, such as visual and auditive or haptic sensory input. There is much evidence that humans integrate these multimodal sensory inputs in a statistically optimal way, which can be explained using Bayes’ rule. To deal with our inherently uncertain world, combining information from multiple sources is crucial not only for humans but also for AI systems, which take over more and more tasks. In my talk, I will present a series of Bayesian models for combining uncertain probabilistic forecasts, inspired by how humans integrate sensory information based on uncertainty. This includes Bayesian models for the combination of probabilistic classifiers, which are able to consider the in-dividual classifiers’ uncertainty, bias, variance, and correlation. I demonstrate these models’ advantages in practical applications, focusing on multimodal human-robot interaction: recognizing human intentions, learning from mul-timodal human advice, and detecting when a person wants to start an interac-tion with a robot. I further present Bayesian models for aggregating subjective probability estimates from human forecasters. Across these use cases, the pro-posed methods reduce uncertainty in a normative way and improve interaction quality and decision support in uncertain environments.