Studying how minds learn conceptual structures from sensorimotor experiences, how self-referential encodings develop, how these structures and encodings interact with language, and how all this depends on and develops via progressively complex object and social interactions, cooperation, and communication.
The Cognitive Modeling Group takes an ecological stance in all these respects, that is, the neural learning processes develop for the purpose of maintaining internal homeostasis of the organism. As a result, all involved structures, encodings, and neural processes serve the purpose to optimize behavior. In the case of our highly interactive and social minds, it is necessary to be able to plan, decide upon, and execute versatile, adaptive, highly flexible, context-dependent behavior to enable deep, communicative, cooperative, social interactions as well as tool usage. It appears that evolution has thus evolved brains that develop, distributed, well-structured, event-predictive structures and, meanwhile, control the inference processes that dynamically unfold within.
As a result, out theoretical considerations focus on the principle of event-predictive inference, which integrates formulations of free energy minimization and active inference with event-integrative common encodings, event segmentations, and event schema-theoretic approaches.
Four focused research areas:
We develop and run behavioral experiments – mostly in Virtual Reality Settings including eye- and motion tracking equipment – to assess far and in which manner our brain explores the future while, for example, interacting with objects or tools, but also while interacting socially with others. Moreover, we are modeling the behavioral results by means of variational free energy formulations of the interaction events and event sequences that the participants of the psychological experiments underwent.
We are working on modeling the unfolding dynamics during conversations – including the inference of preferences – by means of enhanced rational speech act models. Current model enhancements go beyond the processing of instructions towards deeper intentional speech production and consequent observation inferences. In addition, we work on the BrainControl program developing artificial intelligent agents, which learn to converse about their world given their current event-predictive knowledge about it. For example, one current goal is to use the available event-predictive structures to assess which observations about the world may be of interest to the listener. Third, we relate the developing structures to episodic memory, where memorization is expected to focus on those events and event successions that were experienced as particularly “surprising”.
Our deep generative RNN group focusses on developing artificial creatures that first learn to control their own bodies by building predictive models about them. These models are suitably compressed into motor primitives, which constitute particular event codes. Progressively, the event codes will enable the creature to think along them, and thus in a temporally more abstract manner than along actual sensorimotor interaction codes. This process will thus enable the creatures to mentally explore the future in an extended manner.
It is very common for words and sentences to have more than one meaning. If speakers behave rationally, they should try to reduce the amount of ambiguity in their speech to avoid confusing their listeners. In this project, we explore situations when speakers remain ambiguous intentionally. We build models of how watching ambiguity being resolved in conversation can lead to social learning. Using data from behavioral experiments, we show that humans can reason about hidden beliefs and preferences of their conversation partners. This ability is an important component of social cognition: building accurate predictive models of other people allows us to interact with them in an adaptive and efficient manner.