Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster F in Poster Session F: Thursday, August 6, 1:45 – 3:30 pm, Kimmel Center, Shorin & Rosenthal Rooms

A Modular Framework for Embedding Brain-like Topology in Recurrent Neural Networks

Maroš Rovný¹, Danyal Akarca², Jascha Achterberg³, Iva Ilioska¹, John Duncan¹, Duncan Astle¹; ¹University of Cambridge, ²Imperial College London, ³University of Oxford

Presenter: Maroš Rovný

Spatially embedded recurrent neural networks (seRNNs) provide a framework for investigating how the brain’s spatial organisation shapes computation. Previous work extended this framework to include topological constraints, using the Earth Mover’s Distance between communicability distributions to drive artificial networks toward brain-like connectivity structure. A proof of concept demonstrated that such networks not only converge toward empirical communicability distributions but also develop spatially structured mixed selectivity for distinct cognitive demands — a functional signature absent in unconstrained networks. However, this initial demonstration required manual tuning and was limited to a single topological measure and task. Here we present a modular framework designed for systematic investigation of how different structural properties of biological neural networks affect learning and computation. We present functional results from the original proof of concept, describe the framework’s design, and discuss how this approach enables controlled investigation of how neural topology affects function and behaviour.

Topic Area: Methods, Tools, Theory & Neural Coding