Department of Physiology

Research Groups

Computational Neuroscience

Common wisdom often holds that the brain remains poorly understood. However, over the past quarter-century, during which our group has actively witnessed the growth of computational neuroscience, tremendous progress has been made. From early work on motor unit recruitment and spinal cord pattern generators, we have advanced an understanding of dendritic processing in cortical pyramidal neurons, cortical error formation and propagation, and spike-timing dependent synaptic plasticity in dendrites that minimize somato-dendritic prediction errors.

Drawing inspiration from theoretical physics, the Neuronal Least Action principle offers a theory of cortical computation wherein behavioural errors are minimized jointly with dendritic prediction errors. Furthermore, guided by the principles of artificial intelligence, we are beginning to uncover the computational roles of sleep and dreams, dendritic attention mechanisms and cortical microcircuits in language generation. Ultimately, computational neuroscience is now approaching its most profound endeavour: tackling the human consciousness.

Postodc position opening: Description, Job Portal UniBE

Walter Senn

Google Scholar

News

Preprint (bioRxiv)

Cortical multi-head self-attention

We propose a biologically constrained computational framework demonstrating that corticothalamic circuits naturally implement the multi-head self-attention mechanisms that power modern transformer networks. By mapping specific cortical layers to key, value, and query computations, this architecture provides a gradient-based, mechanistic substrate for mammalian cognition that directly aligns with human neural recordings.

11. August 2026

Preprint (arXiv)

Variational latent equilibrium for learning in neuronal circuits

Online-learning of the forward network (W, red) by reducing the back-propagating dendritic prediction error (e, blue). The backward weights (B, blue) are themselves adapted to improve the prediction error. The variational theory merges ideas from the Neuronal Least Action Principle (https://elifesciences.org/reviewed-preprints/89674) and the Generalised Latent Equilibrium (https://www.nature.com/articles/s41467-025-66666-z).

10. March 2026

Roadmap

EBRAINS 10-year roadmap: MechAIBrain: mechanistic AI brain models for behavioral, neural, and biophysical predictions

This proposal introduces a paradigm shift for EBRAINS through a hybrid modeling strategy designed to create a mechanistic foundation model of the human brain and mind. For enabling new medical applications, we need to merge mechanistic models of the brain with the computational power of AI, and to optimize these hybrid brain models for behavioral performance and alignment with neural data.

26. March 2026

Publication

Increased perceptual reliability reduces membrane potential variability

According to Jordan et al. (2024, https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1012047), individual cortical neurons integrate novel (bottom-up) input with prior information (top-down) in a Bayes-optimal way at the level of somatic membrane potentials. Traces illustrate membrane potentials of each compartment. This theoretical prediction is now experimentally confirmed.

7. January 2026

Interview

What if AI were to develop its own consciousness?

Interview with Prof. Dr.phil.nat. Walter Senn (in german)

5. November 2025