School of Engineering Department of Electronic and Computer Engineering 156 Representation Learning for Graph Neural Networks Supervisor: LI Xiaomeng / ECE Student: TILNEY-BASSETT Oktarian George / COMP Course: UROP1100, Summer This project was split into two parts, which allowed me to gain exposure to two different deep learning areas of research. During the first half, I studied graph representation learning through contrastive learning. I started by studying graph neural networks and their properties, then implementing the graph convolutional network model. I then moved onto studying contrastive learning for graphs, an area of a research that is particularly important to solve the data labeling problem. I was able to understand the core frameworks and techniques used, as well as implement a few of the main models. Finally, I studied various models that use novel augmentation schemes to improve the overall accuracy, as contrastive learning frameworks need robust schemes to generate different augmented views for the training. In the second half of the project, I moved to study medical imaging. More specifically, I learnt about transformers and how they can be used in medical imaging to conduct automatic grading of diabetic retinopathy and diabetic macular edema. Overall, this research project has greatly increased my understanding of various deep learning components, as well as their applications to real-life tasks. Robust and Generalized Methods for Medical Image Analysis Supervisor: LI Xiaomeng / ECE Student: CAINE Wilbert / COSC Course: UROP1100, Summer Fundus fluorescein angiography (FFA) is commonly used in ophthalmology diagnosis as complementary image information to obtain a map of the vascular structure of the retina. However, the exogenous dye injected into the bloodstream can cause adverse harm to the patients of angiographic imaging. Despite the characteristic differences, conventional fundus images and FFA images still share most of the prominent features. This report applied a generative adversarial network (GAN)-based method to learn the mapping features from fundus images to FFA images. The result was validated and compared with existing generative networks for fundus-to-angiography synthesis using the publicly available Isfahan MISP dataset. The metrics used for evaluation include peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and Fréchet Inception Distance (FID). Robust and Generalized Methods for Medical Image Analysis Supervisor: LI Xiaomeng / ECE Student: YANG Lin / COMP Course: UROP1100, Summer Deep learning methods have become a trend to solve various computer vision problems and have shown promising performance in recent years. Among the networks, nnU-Net (no-new-net), a convolutional neural network based on U-Net shows great performance on datasets with different contents. This article elaborates the features of nnU-Net and reviews the related works of nnU-Net, including CNNs and U-Nets, analyze their architecture, improvements, and drawbacks. In the last part, this article states a possible direction for the future study in the area of deep learning inspired by the hypothesis of nnU-Net, which is to develop a deep learning method that is universal and is easy to use.