Logic FET based physical reservoir for accelerating event stream analysis

• Main Author: Lin, XiaoYu
• Other Authors: Ang Diing Shenp
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Engineering
• Online Access: https://hdl.handle.net/10356/173496
• Institution: Nanyang Technological University

Artificial intelligence is a significant development in computer science. Complex problems are automatically solved by utilizing specialized algorithms after training. However, this technology has reached a stage that requires high power consumption and long training periods to generate a highly efficient model. In this context, we propose a hybrid physical reservoir system, combining a physical reservoir made up of logic n-MOSFETs with a software-based readout layer. Through the capture and release of carriers by oxide vacancy defects, the HfO2 bulk beneath the gate imparts memory properties to the logic n-MOSFET, condensing the input time series data into an analog value. These inherent physical properties are crucial for temporal correlation. Therefore, in processing dynamic spatial-temporal relationships, the software computational paradigm eliminates the need for complex architectural design. In addition to model simplification and reduced parameter usage, this approach requires ultra-low training time and power consumption. The highly parallel architecture and information density also minimize waste. We applied two readout layers, Logistic Regression and ResNet, to analyze the output generated by the reservoir for the DVS 128 dataset. After nearly 700 seconds of training, the accuracies reached 73.5% and 91.5% (92% for 6 seconds event stream) respectively, using only a 2-second event stream to classify 10 classes of gestures. This dissertation not only demonstrates the great potential of physical reservoirs as accelerators for artificial intelligence, it also shows the feasibility of utilizing current logic FETs as building blocks for the physical reservoir. Relying on the reservoir's output, CNNs and simple machine learning algorithms can be used to recognize movements in videos without any additional modifications. The system exhibits near-real-time processing and has a broad range of applications in miniaturized and portable devices under various conditions due to its low power consumption and fast training time.