Work

Principles of neural network organization that enable the emergence of complex functions in spatially defined regions of the brain

Year: 2025
Type: dissertation
Abstract: The human brain begins as a relatively uniform collection of neural tissue, yet develops into a highly specialized organ with distinct functional regions. While neuroscience has made tremendous progress in mapping these specialized areas, the fundamental question of how functional organization emerges from this homogeneous substrate remains largely unanswered. This dissertation addresses this question through two complementary approaches. Study 1 investigates how cortical features shape functional brain organization by examining whether connectivity, microstructure, morphology, and geometry can predict task activations across seven cognitive tasks. To provide stronger support for the link between cortical features and functional activation patterns, we examined this correspondence at the level of individual brains, reasoning that if a given feature reliably traces functional activation locations in individuals, there should be a strong association between the feature and implemented function. We found that resting-state functional connectivity and cortical microstructure emerged as key predictors, with higher-order connectivity components proving crucial for task-specific predictions. The predictive components were typically shared across multiple cognitive tasks and some could be linked to neurobiological features, which indicates an intrinsic scaffold underlying functional specialization in specific brain regions. Study 2 examined factors that enable the emergence of brain-like function in computational models by systematically comparing representations from diverse pretrained neural networks to corresponding brain activity during naturalistic movie watching. We quantified brain-network similarity using encoding models that predict brain activity from representations of a range of neural networks from language, auditory, and visual domains. We further assessed whether network training leads to region-specific prediction improvements by comparing trained networks to their untrained counterparts. Networks from the same modality consistently showed high predictivity for identical modality-specific brain regions, indicating stimulus-driven organization independent of network architecture. Neural network training enhanced the predictive accuracy for network-modality matching sensory regions, and language pretraining specifically enhanced neural network similarity with regions of the default mode network (DMN). These findings demonstrate that sensory and DMN-like representations can emerge in artificial neural networks through stimulus and learning-driven processes. Together, these findings provide evidence that is consistent with the idea that the brain's functional organization may involve both an intrinsic scaffold and learning-driven processes that lead to refined and functionally meaningful representations.
Institution: Leipzig University
Language: English
Open Access status: green
OpenAlex ID: W7157510856
Cites: 193
Cited by:
Related to: 10