UCLA

Within its tight power budget of 20 W, the human brain effortlessly processes a huge volume ofinformation, including high-definition images with ~60 frames per second, an acoustic signal ranging from 20 Hz to 15 kHz, pressure sensing information over the skin, and chemical compositions through the receptor cells in the taste bud. The empirical observations on the brain revealed its massively parallel synaptic connection among neurons and their local interactions, which is believed to be one of the origins of its exceptional ability to infer and learn. Inspired by the brain’s unparalleled efficiency and adaptability, neuromorphic computing has been introduced to emulate the neurobiological circuits of the brain to develop advanced artificialiii intelligence systems. To achieve this, it is crucial to develop novel electronic devices, particularly those capable of implementing synaptic plasticity with minimal power consumption. This dissertation introduces a synaptic resistor based on HfZrO ferroelectric material, offering stable multi-level, non-volatile memory. Utilizing its inherent learning rules, a circuit of HfZrObased synaptic resistors demonstrated real-time learning capabilities by navigating a drone through tree obstacles and erratic wind conditions in a simulated environment. Furthermore, an associative memory circuit based on synaptic resistors was proposed and employed to classify the MNIST handwritten dataset, illustrating its general applicability to AI and machine learning tasks. These advancements highlight the potential of HfZrO-based synaptic resistors in creating efficient and adaptive neuromorphic computing systems.