Repository of DUTH

Design of deficient neuromorphic systems for the study and development of chimeric synchronization phenomena and bioinspired calculations (Doctoral thesis)

Τσάκαλος, Κάρολος Αλέξανδρος/ Tsakalos, Karolos Alexandros

The human brain’s complexity presents significant challenges for understanding cognitive processes and emergent functionalities. Recent advancements in empirical brain networks elucidate the relationship between network topology and physiological functions utilizing advanced neuroimaging technologies to provide a more comprehensible and simplified yet effective representation of brain networks. Despite their slow and nonideal processing units, biological brains manage complex functions through highly adaptive and robust network-level synchronization and the processing of noisy, nonlinear data. Chimera states, a form of neural synchronization, are particularly intriguing due to their experimental relation with cognitive functions and neurological disorders. They not only model certain brain functionalities and vulnerabilities but also have the potential to be used for the development of bioinspired computing systems. However, the computational challenges in simulating chimera states in complex neuromorphic networks in software necessitate the development of scalable and specialized neuromorphic systems that can emulate these states with high fidelity. Neuromorphic computing provides a versatile platform, which involves the design of analog and digital-level neuromorphic circuits, that simulate neuron and synapse models. Advanced neuromorphic chips aim to address both scalability and power efficiency challenges. However, they not only reproduce ideal neuromorphic features, but also follow the Von Neumann architecture. Adopting a top-down approach, this doctoral dissertation develops digital, analog and molecular neuromorphic networks to progressively incorporate and evaluate the effects of intrinsic device and system imperfections on chimera states. Within the scope of this doctoral dissertation digital neuromorphic networks with reconfigurable Field Programmable Gate Arrays (FPGAs) are designed, focusing on simple, biologically plausible neuron implementations. The design of digital neurons was focused on easy configuration, hardware resource minimization, and asynchronous communication, while providing the simulation of neuromorphic responses of cortical neurons. Then, exploiting parallel processing capabilities of FPGA, a digital neuromorphic network was designed to study synchronization and demonstrate chimera states by evaluating the effect of asynchronous communication between neurons. The results demonstrated the versatility of digital neuromorphic networks in reproducing chimera states, confirming the feasibility of using FPGAs to study and reproduce chimera states, providing a platform for further experimental exploration. However, for a more realistic simulation of neuromorphic responses with higher fidelity and accuracy, we had to pursue another alternative. Subsequently, the dissertation shifted towards analog neuromorphic networks using nanoelectronic nonvolatile memristors in crossbar arrays. These arrays more closely mimic the synaptic functionality of biological brains. The circuits progressively integrated intrinsic mechanisms and device imperfections, enhancing our understanding of how physical imperfections can induce chimera states under specific conditions in both neuromorphic networks of Fitzhugh-Nagumo neurons and Chua circuit networks. Starting from ideal devices, we evaluated the effect of switching threshold mechanisms, device-to-device variability and sneak-path currents on the emergence of chimera states. Extensive circuit-level simulations in LTSpice confirmed the synchronization phenomena and the potential for controlling synchronization by adjusting the states of the memristors through crossbar reprogramming is investigated. Memristor devices are then experimentally validated, with particular focus on the silicon nitride (SiN) devices. We performed endurance tests on properly tuned SiN devices, which are then calibrated to examine the nonlinearities effect on synchronization. Our results demonstrated chimera states despite these conditions. However, recognizing the problem of extensive scalability and integration, of both digital and analog circuits, we proceeded with an alternative system-level approach. Finally, the dissertation expanded to molecular neuromorphic networks, utilizing the three-dimensional structure of Verotoxin proteins, to further scale up and simulate neuromorphic functions at the system level. These proteins, providing a biological and highly adaptable structure, resembling biological neural networks in self-organization, offering biocompatibility, and allowing the configuration of complex and dense networks at the atomic level, enabled the emergence of synchronized and desynchronized domains, essential for chimera states. The study further explored the potential of molecular networks in bioinspired computation, demonstrating their effectiveness in classification tasks within reservoir computing frameworks.