School of Engineering Department of Computer Science and Engineering 92 Data-efficient, Domain Generalizable and Interpretable Deep Learning Supervisor: CHEN Hao / CSE Student: WAN Hanzhe / MATH-PMA Course: UROP 1100, Summer In this report, I summarize the concept whitening framework which transforms an arbitrary deep learning model into an interpretable model by whitening transformation. I will also introduce a new method that disentangles the latent space while keeping the correlation of different concepts. This method, however, is flawed and needs further modification. Deep Learning for Medical Image Analysis Supervisor: CHEN Hao / CSE Student: GU Yi / COMP Course: UROP 1100, Fall Anomaly detection (AD) in medical image analysis is crucial for early disease detection and pathological localization. Current methods, such as reconstruction-based anomaly detection using Autoencoders and Generative Adversarial Networks, face challenges due to inaccurate reconstructions or inadequately calibrated likelihoods. This paper introduces Maverick Learners, a novel ensemble-based uncertainty estimates framework for medical AD. The proposed method enforces learners to disagree on anomalies through an explicit constraint loss and a bidirectional discrepancy, which combines both forward and backward discrepancy information. Extensive experiments on five medical benchmarks demonstrate the effectiveness of our approach. Deep Learning for Medical Image Analysis Supervisor: CHEN Hao / CSE Student: KANG Wooyoung / COMP Course: UROP 1100, Fall In the rapidly evolving domain of artificial intelligence, deep learning has emerged as a transformative force, especially in the field of machine learning. Characterized by its ability to learn and make inferences from large volumes of data, deep learning is particularly significant in areas where traditional analytical approaches fall short. One such area is medical image analysis, where the complexity and variety of data require sophisticated, nuanced interpretation. The application of deep learning in medicine has opened new horizons for diagnostic methodologies. By leveraging complex neural network architectures, deep learning algorithms have shown remarkable proficiency in interpreting medical images, such as X-rays, MRI scans, and histopathology slides. These advancements are not just technical achievements; they represent substantial strides in enhancing diagnostic accuracy, personalizing treatment plans, and ultimately improving patient outcomes.
RkJQdWJsaXNoZXIy NDk5Njg=