A team of researchers led by Prof. Dr. Ilka Diester, professor of optophysiology and spokesperson for the BrainLinks-BrainTools research center at the University of Freiburg, has developed a formal description of internal world models and published it in the prestigious scientific journal Neuron. This formalized approach allows scientists to understand the development and functioning of internal world models in more detail. In addition to enabling systematic comparison of world models in humans, animals, and artificial intelligence (AI), it also reveals areas where AI still lags behind human intelligence and provides guidelines for further AI system development. Eleven researchers from four different faculties of the University of Freiburg participated in the creation of this interdisciplinary publication.

Internal world models and experience-based predictions
Humans and animals have the ability to abstract general laws from everyday experiences, which helps them navigate new and unfamiliar situations. They develop internal world models that allow them to predict and adapt behavior to new circumstances. For example, understanding the structure of similarly configured cities can make it easier to navigate an unfamiliar city. Similarly, experiences gained during social events, such as dinner in a restaurant, allow for appropriate behavior in similar situations.

Formalization of internal world models
To formalize internal world models in different species, the researchers in their publication distinguish three abstract spaces that are interconnected: task space, neural space, and conceptual space. Task space encompasses everything an individual experiences, including tasks and activities. Neural space describes various measurable brain states, from the molecular level to the activity of individual neurons and entire brain regions. Visualization of these activities can be achieved using functional magnetic resonance imaging (fMRI) or measurements via high-density electrodes and calcium imaging. The equivalent of neural space in artificial intelligence represents the activity of nodes within an artificial neural network. Conceptual space consists of pairs of states of task space and neural space, which together represent the individual's state and connect internal processes with external influences. These pairs constantly change, transitioning to the next state with a certain probability. The combination of individual experiences and corresponding brain activity, along with dynamic transitions, makes internal world models scientifically tangible.

Improving artificial intelligence and understanding mental illnesses
The formalized approach to internal world models allows scientists to analyze these models across various scientific disciplines and discuss their origin and development. Research results on humans and animals could be useful for improving artificial intelligence. Current AI systems, for example, are still unable to verify the credibility of their predictions. Large language models, such as ChatGPT, function as pattern recognition machines without the ability for real planning. Planning is crucial for testing and correcting strategies in unfamiliar situations before applying them to avoid potential harmful consequences. Scientists also hypothesize that deficiencies in internal world models could be associated with some mental illnesses, such as depression or schizophrenia. A deeper understanding of these models could help in the more precise use of drugs and therapies to treat these disorders.

Research on internal world models represents a significant step forward in understanding the complex interactions between the brain and the environment. With new formalized approaches, scientists can study in detail how individuals abstract information from experiences and use it for adaptation in new situations. This has far-reaching implications not only for the development of advanced AI systems but also for improving the treatment of mental disorders. Precisely defining and measuring these internal models allows for a better understanding of how the brain processes information and makes decisions.

The application of this research can be seen in various fields, from the development of autonomous vehicles that need to make decisions in real-time, to sophisticated medical diagnostic tools that can identify and suggest treatments for complex health problems. Understanding internal world models is also crucial for progress in the field of robotics, where robots need to develop the ability to adapt and learn from interaction with humans and the environment.

In the future, further research in this field could lead to revolutionary changes in how we develop technology and treat diseases. Formalized world models provide scientists with a tool for detailed study of complex systems and open new opportunities for interdisciplinary collaboration and innovation. This advancement allows us to better understand not only artificial intelligence but also the essential aspects of human cognition and behavior.

A team of scientists from Freiburg has shown how interdisciplinary collaboration can lead to significant discoveries that have the potential to transform various aspects of science and technology. Their research on internal world models lays the foundation for future studies that will continue to explore the complex interactions between the brain, behavior, and environment. Through such efforts, science is getting closer to understanding how to develop intelligent systems that can better serve humanity.

Source: Universität Freiburg