Biological constraints on neural network models of cognitive function
News vom 12.07.2021
"Neural network models are potential tools for improving our understanding of complex brain functions. To address this goal, these models need to be neurobiologically realistic". Following this approach, Pulvermüller and colleagues highlight a novel research strategy, called brain-constrained modelling in their paper "Biological constraints on neural network models of cognitive function": Neural networks bridging between the microscopic level of nerve cell function, the mesoscopic level of interactions in local neuron clusters and the macroscopic level of interplay between these clusters and even larger brain parts are used to approximate human brain structure and function at different levels.
The paper »Biological Constraints on Neural Network Models of Cognitive Function« is one of the first key outputs from a recently initiated Advanced Grant funded by the European Research Council called Material Constraints Enabling Human Cognition (MatCo, ERC-2019-ADG 883811). In this project and in the Cluster of Excellence »Matters of Activity. Image Space Material«, Pulvermüller and his team are now systematically approaching material-based answers to questions such as the following: How can humans learn a vocabulary of 10,000s of words whereas our closest relatives are normally stuck with 10s? How is it possible that little children quickly interlink signs with meanings, upon only one experience in the extreme, although our closest relatives have great difficulty building such links and neural networks require excessive time for learning them? By which mechanisms can we build abstract concepts and what contribution (if any) makes language to this process? (… and many others)
Pulvermüller, F., Tomasello, R., Henningsen-Schomers, M. R., & Wennekers, T. (2021). Biological constraints on neural network models of cognitive function. Nature Reviews Neuroscience, doi: 10.1038/s41583-021-00473-5.
More information about this paper
More information about this paper: German Version
MatCo Project Website