Study on memristive intelligent system for neuroscience-inspired AI

Abstract

With the rapid development of artificial intelligence (AI) algorithms, it has provided a breakthrough in intelligent systems, such as the Internet of Things (IoT), neuro-robotics, and autonomous vehicles. Since the fields of AI and neuroscience have been intertwined, neuroscience has become a rich source of inspiration and validation for various algorithms and architectures in AI, leading to the rapid development of artificial intelligence. However, due to the diversity and complexity of the biological nervous system, neuroscience-inspired AI faces a challenge in hardware implementation using conventional digital computing. Therefore, it is essential to develop neuromorphic hardware for neuroscience-inspired AI. Memristors can mimic the brain's synapses and neurons, providing a neuromorphic hardware platform. In this dissertation, I develop an intelligent system platform for artificial intelligence by developing memristor-based integrated devices and network structure-based memristor devices. First, using memristor-based integrated devices, I propose neuromorphic vision hardware for motion detection and explainable artificial intelligence hardware that provides explanations of artificial intelligence actions. Taking inspiration from the structure and function of the insect's visual system, a memristor-based neuromorphic vision system is proposed. This is incorporated with a neural network algorithm and used in a machine vision system to predict lane change maneuvers. In addition, inspired by peekaboo in infants, a memristor-based explainable artificial intelligence hardware is proposed to explain and interpret the AI decision-making process. Second, I propose self-organized neuromorphic devices and neuromorphic sensor hardware using network structure-based devices. A self-organized neuromorphic device is proposed to mimic the stochastic structure of the biological nervous system. I emulated the behavior of the biological neural network resulting from its structure. In addition, taking inspiration from the function of a bird’s multisensory system, neuromorphic multisensory sensor hardware is proposed. I examined its noise-tolerance performance in noisy and dim environments for object and motion perception. Each of the four studies performs computations for sensation, perception, decision, and explanation in biological neural networks, and integrates each function into a memristor-based intelligent platform. The feasibility and effectiveness of the memristor-based intelligent platform are confirmed through experimental verification and simulation-based large-scale system scalability verification.