Avedillo de Juan, María JoséNúñez Martínez, Juan2025-06-272025-06-272025-03-07Jiménez Través, M. (2025). Hardware Demonstrators of Oscillatory Neural Networks: A Versatile Approach to Brain-Inspired Computing, from Embedded Systems to Phase-Transition Materials. (Tesis Doctoral Inédita). Universidad de Sevilla, Sevilla.https://hdl.handle.net/11441/174712The ever-increasing amount of information that computing platforms are required to handle presents a tremendous challenge, particularly concerning power consumption, data rate, and processing speed. The separation of processing and memory constitutes a significant efficiency bottleneck in well-established systems based on the von Neumann architecture. For example, the rise of Edge Computing and the Internet of Things underscores the need for more efficient and decentralized computing architectures. As the demand for real-time processing, reduced latency, and low power consumption intensifies, traditional CMOS-based architectures are approaching their inherent limits, and innovative perspectives at different levels are gaining considerable interest. Research covering from materials and devices to architectures and systems, as well as novel computing paradigms, aims to provide medium-term solutions. Specifically, computation that resorts to dynamical systems, such as coupled electrical oscillators, has emerged as a fast and energy-efficient computing paradigm. In this approach, the problem solution is encoded within the system energy function and achieved through its inherent dynamics in a collective-intelligent, parallel manner. Following a brain-inspired approach to this computing paradigm, the Oscillatory Neural Networks (ONNs) consist of many oscillator (neuron) circuits interconnected by electrical elements (synapses). Recent advances in materials technology and the emergence of novel devices have enabled the implementation of compact oscillators with very low power consumption. In particular, oscillators built with the phase-transition material vanadium dioxide (VO2) have received significant interest, and their exploitation in ONNs has become an active area of research. This Doctoral Thesis presents the development of different hardware demonstrators of ONNs and their application as Associative Memories for pattern recognition tasks and as Ising Machines to solve combinatorial optimization problems. First, a fully digital ONN was designed as a proof of concept of the oscillatory-based computing paradigm. It was implemented in FPGA technology and two prototypes were experimentally validated: a digit-recognition application and a responsive obstacle-avoidance system in a mobile robot. Later, an analog CMOS-ONN ASIC was designed and fabricated in a commercial TSMC 65nm technology, integrating a CMOS emulator of the electrical behavior of the VO2 material. This ASIC enabled an early evaluation of the VO2-based ONN response. Finally, demonstrators built with fabricated VO2 oscillators and discrete components were evaluated, showcasing the utility of VO2-based ONNs to solve computationally hard problems in very few cycles. Alongside the development of these demonstrators, a comprehensive review of suitable learning rules was conducted to enhance the capabilities of the ONN as Associative Memory. This led to the conception of a novel iterative algorithm, IRPUSH, which outperforms all the existing learning rules compatible with ONN. It ensures the learning of highly correlated patterns and achieved high retrieval accuracy in pattern recognition tasks, even with reduced weight precision, showing a minimal loss of accuracy. These efforts have significantly contributed to validating and deepening our understanding of the oscillatory-computing paradigm using VO2-based ONNs. The insights derived from the development of the demonstrators, regarding the fast and parallel convergence of the ONNs to the problem solution, have proven the potential for energy-efficient computation. Furthermore, a solid foundation has been established on how electrical ONNs should be operated and configured considering the specific application while exploring methods to overcome the intrinsic constraints and limitations of such a complex dynamical system.application/pdf147 p.engAttribution-NonCommercial-NoDerivatives 4.0 Internationalhttp://creativecommons.org/licenses/by-nc-nd/4.0/info:eu-repo/semantics/doctoralThesisinfo:eu-repo/semantics/openAccess