Predicting Structure-Function Relationships in Cortex via Artificial Neural Networks
1 Max Planck Institute for Neurobiology of Behavior
The coupling of information streams is a hallmark feature of cortical function, and it is believed that network architecture is key to this. Neuroanatomical studies have shown that the specificity of projections of particular cell types both to and from the cortex facilitate the formation of characterizable networks. However, determining the impact these structures have on how incoming information streams are coupled and processed is not well understood and remains challenging to study. Here, we propose a computational approach to investigating this by informing artificial neural network models with increasing detail from neuroanatomical reconstructions of the cortex. We demonstrate that by training such cortically-inspired networks on a battery of machine learning tasks, we obtain concrete predictions on how network architecture and wiring specificity therein could facilitate function. We explore such structure-function relationships with respect to biologically-relevant tasks like generalization and show how these networks compare to other possible architectures. Our approach provides promising results and empirically testable predictions which we hope will shed new light on how interareal connectivity patterns facilitate the manner by which information streams are coupled.