Singer Lab

Research Topics at Singer Lab at ESI

The working hypothesis of the Singer Lab is that the processing mechanisms in the cerebral cortex utilize the high-dimensional dynamics that arise in recurrent networks whose nodes have the properties of damped harmonic oscillators.

Years ago, researchers discovered that feature-specific neurons in the visual cortex synchronize their responses when these features belong to the same visual object. This mechanism is discussed as a possible solution to the “binding problem” and for the formation of Hebbian assemblies.

This synchronization is often accompanied by oscillations in the gamma frequency range, raising the question of whether the observed oscillations have functional meaning or are merely an epiphenomenon of complex interactions in recurrent networks.

This question is difficult to answer in neurophysiological experiments because it is almost impossible to manipulate the oscillations in a targeted manner without also changing other variables of the system dynamics. Therefore, researchers at the Singer Lab at ESI conducted simulation experiments that allowed them to control the oscillations in a targeted manner. To do this, they programmed a simulation of a recurrent network and configured the nodes as damped harmonic oscillators.

By varying the damping factor, they were able to configure the nodes to function either as conventional integrators (leaky integrators) or as damped oscillators. To quantify the performance of the networks, the scientists trained them to recognize handwritten letters and spoken words.

It was found that the networks performed significantly better at recognition when the nodes were enabled to oscillate. The learning speed and resilience of the networks to disturbances increased dramatically compared to non-oscillating networks. The implementation of additional properties of the cerebral cortex (heterogeneous oscillation frequencies, varying conduction speeds of the recurrent connections, and modularity of the connection architecture) led to further increases in the performance of the networks without increasing the number of parameters to be specified. In addition, the oscillating networks reproduced a variety of structural and functional features of the cerebral cortex.

This suggests that the information processing processes based on oscillations are also realized in the cerebral cortex and that the oscillations should not be considered an epiphenomenon.

Much of the current work at the Singer Lab at ESI aims to understand the computational operations that make oscillating recurrent networks so unusually efficient.

The researchers are investigating this question with extensive simulation experiments. They are focusing on three objectives:

1. The researchers hope to gain deeper insights into the functioning of the cerebral cortex and to find explanations for phenomena that have been discovered in physiological experiments but whose functional role is still unclear. These include, in particular, the complex dynamics of the cerebral cortex and the characteristics of the architecture of neural networks.
2. By analyzing the learning processes in oscillating networks, they hope to gain insights into the mechanisms by which natural systems build an internal model of their environment during their development, which they then use to organize and interpret sensory signals (predictive coding).
3. The researchers want to examine the extent to which the learning processes and computational strategies identified in the simulation experiments can be transferred to technical systems in order to increase their performance and energy efficiency. The focus here is on analog information processing options.