School of Engineering Department of Electronic and Computer Engineering 139 Deep Learning for Magnetic Domain Image Denoising and Super-resolution Supervisor: SHAO, Qiming / ECE Co-supervisor: LI, Xiaomeng / ECE Student: DUAN, Qinkai / SENG Course: UROP2100, Fall I joined in competition TinyML with Prof.Shao’s urop students. My work was to design the model and try different algorithms related to super-resolution. Our team got a good grade at last, we ranked 14. I learned SRCNN, res-net, data vision, Generative Adversarial Network. I also did a test for basic super-resolution model on medical machines, using ESRGAN. I runed the model on retina dataset STARE. I focused on midterm and final exams in the second half of this term so I will do more experiments on skin cancer, brain tumor, cardiac ultrasound datasets and tries to improve the model performance. Deep Learning for Magnetic Domain Image Denoising and Super-resolution Supervisor: SHAO, Qiming / ECE Co-supervisor: LI, Xiaomeng / ECE Student: PENG, Yiyan / SENG Course: UROP1100, Fall This report is about what I learned during this semester, and it mainly covers two parts – TinyML Competition and UROP project progress. First, “TinyML” refers to the “2022 ACM/IEEE TinyML Design Contest at ICCAD”. There were six members in our group working on this competition in September 2022. We finally got a rather good result and ranked 14 out of 150 worldwide. Second, during the rest of this semester, I was working on the UROP project “Embedded System for AI hardware” and reproduced the out-of-plane Anomalous Hall hysteresis loops of the Tantalum (Ta), following the article “Asymmetric Spin-Orbit-Torque-Induced Magnetization Switching with a Noncollinear In-Plane Assisting Magnetic Field” by Fan, W. et al. (2019). Deep Learning for Magnetic Domain Image Denoising and Super-resolution Supervisor: SHAO, Qiming / ECE Co-supervisor: LI, Xiaomeng / ECE Student: TAN, Tianshu / CPEG Course: UROP1100, Fall Machine Learning (ML) is one of the central attractions of Artificial Intelligence (AI), while most ML and training happened on powerful servers over the last few decades. However, recently, the experiments have led us to a point where AI devices can be performed on the edge, right where the data source is. The concept of Tiny Machine Learning (TinyML) has been raised due to its improvement in privacy, energy efficiency, affordability, and reliability of the applied Artificial Intelligence (AI) devices. The purpose of this thesis is to report an TinyML project on discriminating the life-threatening VAs (i.e., Ventricular Fibrillation and Ventricular Tachycardia) from single-lead (RVA-Bi) IEGM recordings, under the requirements of the ACM/IEEE TinyML Design Contest at ICCAD. Our primary target is to construct an optimized neural network for processing ECG signals and identifying dangerous VA signature patterns, and eventually incorporate it into microdevices to provide early warnings for patients during their asymptomatic period. Models will be evaluated based on criteria like detection precision and memory occupation. We’ve improved 9.7% on F_beta, 87.72% on Latency, and 67.98% on Flash.
RkJQdWJsaXNoZXIy NDk5Njg=