Starting from conception, our brains develop throughout our lifetimes. They generate bounded, somewhat rational intelligence, dependent on its cognitive capacity and the structure and quality of its knowledge structures.
The Neuro-Cognitive Modeling Group addresses the following main questions:
- How do our minds develop from a computational perspective?
- How is knowledge and memory structured?
- Which processes invoke learning, cognition, and behavior within?
- How can these structures and processes be modeled with both Bayesian models and artificial neural networks (ANNs)?
- And, based on these insights, how can artificial intelligence (AI) be made human-compatible; ideally self-explaining its knowledge and reasoning.
We try to steer our research towards supporting healthy, sustainable, and enjoyable world development. Naturally, thus, it is necessary to also address the following question actively:
- How can AI support sustainable development and human resilience?
Our main premise is that our brain is an inference system that dynamically learns, develops, and maintains context-conditioned, event-predictive, conceptual, probabilistic structures. Via these structures its inference processes actively infer – and thus generate and control – current attention, thought, and behavior. The inference processes are driven by the goal to maintain internal homeostasis while minimizing cognitive effort.
To corroborate evidence, and to shed light on the details behind, we conduct behavioral studies in the real world as well as in virtual realities, including language production and interpretation studies. Moreover, we are building artificial, Bayesian and deep – typically recurrent – generative neural network models. We use the models to probe our integrative theoretical assumptions and to develop useful, truly artificially intelligent systems. Meanwhile, we advance computational theories of machine learning. We model the development of dynamically unfolding resource-aware learning and decision-making processes (concurrently). And we study inductive learning and processing biases to focus learning progress and representation effort on task-critical environmental structures and causal interactions.
We integrate our research into further-reaching aspects concerning human cognition, AI, and the more general impact of this research on our societies and our world. On the cognition side, these include the (reflexive and reflective) self, action decision making, planning, reasoning, social interaction, education, and language. On the AI side, these include the development of crucial compositional components for grounding AI in actual sensorimotor experiences (typically in simulators). On the social side, these include the analysis and prevention of spreading false beliefs and other forms of manipulation as well as considerations about how a progressively strong AI may support a healthy, sustainable, and resilient development of our world -- a world that offers long-term sustainable, worth-while lives to all of us including future generations of us.
Over the last four years we have also invested time and effort into understanding, modeling, and predicting natural dynamics in the geosciences – including water discharge, erosion, substance advection and diffusion, weather, and climate. In this respect, we have uncovered overlooked similarity relationships in vertical flux dynamics, and we have developed hybrid physics-informed AI and global forecasting ANNs.
Martin Butz
