A simulated memristor architecture of neural networks of human memory

Abstract

The project presents a hybrid approach between artificial intelligence and neuroscience as a more common framework to investigate the function-structure relationship, emphasizing the computational properties of neural networks. The human connectome will be reconstructed using electrophysiological studies, implemented as an artificial reservoir, and trained to perform memory tasks. By comparing connectome-informed reservoirs with arbitrary architectures, the computational properties of the human connectome will be optimized at a unique macroscale network topology and its mesoscale modular organization under critical network dynamics, assumed to perform optimal information processing. The hypothesis is that regardless of global network dynamics, the human connectome maximizes memory capacity by minimizing metabolic and material costs. The idea that the interplay of network dynamics and structure sustains and modulates the computational capacity of connectome-informed reservoirs may explain the spectrum of computational abilities of the anatomical macroscale brain network. By combining connectomics and reservoir computing, it will be possible to implement biologically derived network architectures and connectomes as artificial neural networks in memory tasks. Opportunities to investigate novel facets of the function-structure relationship in brain neuronal networks will arise from the adaptable approach concerning task paradigm, network dynamics, and architecture. Another question is how variations in the connectome architecture give rise to different developmental cognitive abilities in information and computational processing of neural networks. Artificial reservoirs such as memristors have been proposed to explore information processing aspects of the brain by combining modern electrophysiological computing tools and those from artificial intelligence, such as spiking artificial neural (memristor) networks.