Search Papers | Poster Sessions | All Posters
Poster A77 in Poster Session A - Tuesday, August 6, 2024, 4:15 – 6:15 pm, Johnson Ice Rink
A taxonomy of multifunctionality in connectome-constrained neuromorphic computers
Jacob Morra1 (); 1Department of Computer Science, Western University, 2Centre for Vision Research, York University
The relationship between biological neural network (BNN) structure and artificial neural network (ANN) function is poorly understood. In the brain, structure and function are positively correlated, but far from unity. And yet, structure acts as a guide for function, while disruptions in structure can lead to dysfunction. As a NeuroAI approach, finding topological rules can help us to curtail ANN behaviours. Thanks to recent efforts in connectomics, or the construction of brain wiring diagrams, this strategy is now viable. We are well-poised to begin to explore 'what' questions, before proceeding towards the 'how' or 'why'. The 'what' in this paper builds on previous work constructing a taxonomy of mesoscale connectomes across species. We sample from this dataset, comparing 18 connectomes to their rewired counterparts within a neuromorphic machine learning framework (reservoir computing) on their capacity to exhibit multifunctionality (MF). We observe a dramatic difference between connectomes and their rewired variants, suggesting a link between BNN structure and MF. We furthermore identify shared features across 'successful' networks. Future work will ablate these features and vary known MF-related parameters (e.g. the spectral radius) in order to analyze model prediction dynamics.
Keywords: Connectome-constrained machine learning Network neuroscience Reservoir computing