Exploiting Learning Capabilities of Neuromorphic Nanowire Networks

Abstract

As the Moore’s Law era approaches its end and the constraints of conventional computing paradigms become starkly evident, the pursuit of alternative, efficient computational models has emerged as a critical necessity. The human brain, with its remarkable capability and energy efficiency, stands as a compelling model from which we can draw inspiration to engineer intelligence into machines. This thesis explores the potential of Neuromorphic Nanowire Networks (NWNs), aiming to exploit their processing power and learning capa- bilities similar to biological neural systems. With their brain-like structures and electrical junctions resembling synapses, NWNs stand as a promising physical substrate for creating hardware systems that capture the brain’s adaptability and efficiency. This thesis investigates both simulated and physical NWNs, with a concentrated focus on understanding and leveraging their learning potential. Through in-depth investigations into the information dynamics of NWNs during learning tasks, and exploration of their learning capability under a reservoir computing framework, this research unveils new insights into unleashing the potentials of NWNs. By employing NWNs under a reservoir computing framework, it was found that their optimal learning performance correlates with specific information-theoretic measures. Additionally, drawing inspiration from the machine learning community, transfer learning, convolutional kernels, meta learning, and recursive least square were implemented to exploit the inherent learning and memory capacities of NWNs. The findings presented in this thesis underscore the significant potential of NWNs in neuromorphic computing, paving the way for future research and applications in crafting energy-efficient, learningcapable computational systems.