Search Papers | Poster Sessions | All Posters
Poster B106 in Poster Session B - Thursday, August 8, 2024, 1:30 – 3:30 pm, Johnson Ice Rink
Connectome-constrained spatially embedded recurrent neural networks
Maroš Rovný1 (), Danyal Akarca1, Jascha Achterberg1, Duncan Astle1; 1MRC Cognition and Brain Sciences Unit, University of Cambridge
Deep learning models are one way to address the reliance of cognitive neuroscience on association studies, by providing a sandbox for systematically testing causality in complex systems. However, only a fraction of these allow for the modelling of the entire human brain. Expanding upon a recent innovation of spatially embedded recurrent neural networks, our approach introduces a novel regularisation term based on the optimal transport problem. This enriches the training process with information about the distance between the distributions of network measures describing the artificial and empirical topology. Exemplified through communicability, our approach unlocks future avenues for exploring the impact of different topological properties of the human brain on its performance across varied contexts by allowing these to be included in training and tested on different tasks.
Keywords: topological embedding spatial embedding regularisation recurrent neural networks