Modularity and multitasking in neuro-memristive reservoir networks. Issue 1 (23rd August 2021)

• Record Type: Journal Article
• Authors: Loeffler, Alon; Zhu, Ruomin; Hochstetter, Joel; Diaz-Alvarez, Adrian; Nakayama, Tomonobu; Shine, James M; Kuncic, Zdenka
• Is Part Of: Neuromorphic computing and engineering. Volume 1:Issue 1(2021)
• Publication Date: 2021-08-23
• Subjects: modularity -- multitasking -- neuromemristive -- nanowire networks -- reservoir computing -- neuromorphic networks; Neural networks (Computer science) -- Periodicals; Neural computers -- Periodicals; Neuromorphics -- Periodicals; 006.3
• DOI: 10.1088/2634-4386/ac156f
• Languages: English
• ISSNs: 2634-4386

Abstract

Abstract: The human brain seemingly effortlessly performs multiple concurrent and elaborate tasks in response to complex, dynamic sensory input from our environment. This capability has been attributed to the highly modular structure of the brain, enabling specific task assignment among different regions and limiting interference between them. Here, we compare the structure and functional capabilities of different bio-physically inspired and biological networks. We then focus on the influence of topological properties on the functional performance of highly modular, bio-physically inspired neuro-memristive nanowire networks (NWNs). We perform two benchmark reservoir computing tasks (memory capacity and nonlinear transformation) on simulated networks and show that while random networks outperform NWNs on independent tasks, NWNs with highly segregated modules achieve the best performance on simultaneous tasks. Conversely, networks that share too many resources, such as networks with random structure, perform poorly in multitasking. Overall, our results show that structural properties such as modularity play a critical role in trafficking information flow, preventing information from spreading indiscriminately throughout NWNs.