Towards Brain-Inspired Computing: Memristor-Based Single Layer Neural Networks

This master thesis delves into the realm of neuromorphic computing, a promising paradigm shift poised to overcome the limitations of traditional von Neumann computer architectures in the field of neural network algorithms. The focus is squarely placed on the memristor, a non-volatile memory device that emulates the behavior of synapses in the human brain, offering a scalable and efficient alternative for computing systems. Through rigorous exploration of design, fabrication, characterization, and simulation aspects, this dissertation evaluates the practicality of incorporating memristive devices into neuromorphic computing systems.

The research is framed around a multi-faceted approach that includes the innovation of tools and computational frameworks, with a highlight being the development of a pioneering software platform that synergistically integrates the capabilities of LT-Spice and Python. A characterization software, "PyCharMem", has been developed to streamline the characterization process, making it more accessible and scalable. Additionally, a portable probe station was conceived and manufactured, addressing space constraints.

A total of 19 distinct memristive devices were subjected to electrical characterization, facilitating the extraction of their critical properties. This exhaustive analysis laid the groundwork for their emulation using SPICE models, navigating the complexities through a genetic optimization algorithm tailored to extract intricate parameters. The research culminates in the execution of single-layer neural network simulations, employing a gradient descent learning algorithm through a reputable SPICE simulator, ensuring a comprehensive and reliable evaluation of sneak paths in memristor crossbars.