A novel machine learning model enabled by Perovskite-memristor-based hardware

Abstract

Machine learning (ML) offers the automation of routine yet highly complex decision making processes, previously regarded as only a human job. However, ML is still limited by the amount of computational resources it requires to operate. As such, employing machine learning for decision-making on the spot, where the input data originates, is challenging due to resource constraints associated with such settings. In this thesis, I present my contribution to the solution of this problem in the form of a novel ML algorithm, Fuzlearn, a model of a new electrical component, namely memristor, that can be utilised for efficient hardware acceleration of Fuzlearn, and, finally, a circuit simulation of the said hardware accelerator. In the first part of the thesis, I review the drawbacks of already existing ML algorithms, namely neural networks (NNs) and Tsetlin machines (TMs), which limit their applicabil ity for in-situ ML. Additionally, I explore the potential of memristors for ML hardware implementation/acceleration, with Halide Perovskites emerging as the most promising candidate due to Perovskites’ inherent memristive nature and ease of fabrication. In Chapter 3, I introduce Fuzlearn - a novel ML algorithm, based on a Tsetlin machine architecture. I present its working mechanism, which performs analogue-to-Boolean clas sification by learning appropriate input class boundaries. I demonstrate its effectiveness through extensive testing on various real-world ML problems, highlighting its capability to process analogue inputs effectively, as well as demonstrating potential caveats that can be encountered while using Fuzlearn. In Chapter 4, I delve into the development of Perovskite memristors, presenting and comparing different device configurations for optimal usage within ML hardware. Through rigorous testing, I confirm that filament-formation silver-electrode design yields the most promising memristive properties. As a major contribution to this chapter, I created and verified a simplistic yet powerful model of Perovskite memristors suitable for rapid circuit design applications. Finally, Chapter 5 focuses on integrating Fuzlearn into hardware by leveraging my Per ovskite memristor model. I present my circuit design, which successfully demonstrates the algorithm’s functionality within circuit simulation software, proving its potential for prac tical implementation as a hardware-accelerated architecture for machine learning tasks.