Center for Artificial Intelligence in Medicine (CAIM)

"I want to understand how the brain learns and makes predictions."

February 2024

Katharina Wilmes is fascinated by the question of how humans learn in an uncertain world. To find out more, she conducts postdoctoral research at the “Computational Neuroscience Group” of the Department of Physiology, University of Bern. Her research creates neuroscientific models, combining mathematics, neural theory, and computational neural network simulations with experimental testing of hypotheses linking sensory experience and learning.

Katharina uses abstract mathematical models to try to explain how our brain works and learns (© CAIM, University of Bern).

Katharina, what are you currently working on?
I am interested in how we learn from sensory experiences. My project focuses on learning in a noisy and uncertain world. More specifically, I investigate how prediction errors can be used as a signal for learning and how we can model this on the computer. The current research suggests that we make predictions about what we are going to see and experience which allows us to compare these predictions with what actually happens.
Imagine taking a bus. You expect the bus to arrive at a certain time plus minus some minutes. It is important that you learn about the variability of the bus, because if the bus is only a few minutes late, this small prediction error should not change your behaviour for the next time. However, if the bus is usually quite punctual, but one day it is 20 minutes late, this large prediction error can indicate that the bus schedule changed. Prediction errors are very useful for learning as they tell us how we need to adapt our model of the world so that it works again. However, it is important to also know the variability when learning from prediction errors.
For my project, I develop theories about how this prediction error is represented in the brain and develop models of neural circuits to better understand the neural correlates of prediction errors. This ties in nicely with the goal of our research group. We develop abstract mathematical models of neural circuits and neural networks. With this we want to explain the neuronal substrate of learning.

Katharina enjoys developing theoretical models of how the brain is dealing with uncertainty and how we learn from this. In computational models, she studies prediction errors and how they are used to create a largely realistic model of the world (here with teammate Federico Benitez) (© CAIM, University of Bern).

How did you find your research topic?
I have always been interested in how the brain can learn. It is quite astonishing that it changes so much with every sensory experience. We humans are so flexible! Artificial neural networks are still having trouble to reach the level of human flexibility. I really want to find out how our brains can do this!
My background is in biophysical modeling. For my current project we hypothesized that uncertainty should modulate prediction errors based on a mathematical theory called Bayesian inference. We then translated that into a neural theory how this could be implemented in the brain. That resulted in our model: a simulation of a network of neurons on the computer. This simulation gives us the predictions of neural activity that our experimental colleagues can then test in an experiment. Their results then feed back into our theory.
In the future, I plan to work on understanding how changes in neural circuits affect prediction errors and learning. More knowledge on the involved cell types could be important to understanding psychiatric disorders such as autism spectrum disorder which includes a difficulty dealing with uncertainty. Theoretical models could help bridge the gap between behavioral and neuroscience research.

Katharina is at the moment the only female team member in her group – not an unusual phenomenon in computational neuroscience. Here, she is explaining her theory to her colleagues (Federico Benitez, Prof. Walter Senn, Mihai Petrovici) (© CAIM, University of Bern).

With our research we want to understand the neuronal substrate of learning.

What are the most exciting developments in your field?
The development of optogenetics in experimental neuroscience enabled scientists to activate or silence individual cell types with light. This allows to study the activity and role of individual cell types in a circuit. We can also see what changes when this cell type is manipulated. This allows my neuronal models, which make predictions for individual cell types, to be tested in experiments to better understand what is happening in the brain during behavior.
I came to Bern because in this lab I have both the theoretical underpinnings and the direct connection to experimental neuroscience. Even though I arrived during the pandemic I have had a very nice start and could establish a good network with other postdocs – also through the Mittelbauvereinigung of the University.

(© CAIM, University of Bern)

Dr. Katharina Wilmes is a postdoctoral researcher in Walter Senn's lab at the University of Bern, Switzerland, working on the theoretical side and collaborating with experimentalists. She studied Cognitive Science (BSc) at Osnabrück University, and Brain and Cognitive Sciences (MSc) at the University of Amsterdam. During her PhD studies with Susanne Schreiber and Henning Sprekeler at the Institute for Theoretical Biology, Humboldt University of Berlin and the Bernstein Center for Computational Neuroscience in Berlin she received a PhD fellowship from the German Research Foundation (DFG) within the graduate school GRK 1589 “Sensory Computation in Neural Systems”.

From February 2017 to April 2020, Katharina was a postdoc with Claudia Clopath at Imperial College London, UK, funded by a DFG postdoctoral research fellowship. In 2020, she joined the Department of Physiology of the University of Bern. Her current and future research aim is to understand how cortical circuits enable us to learn from sensory experiences by predicting them.