UROP Proceedings 2022-23

School of Engineering Department of Computer Science and Engineering 125 Research on Mining Course Structure Supervisor: WONG, Raymond Chi Wing / CSE Student: LIM, Hyungmin / DSCT Course: UROP1100, Spring Dimensionality reduction, a technique that transforms a high-dimensional data to a reduced representation of original data, such that the number of dimensions is smaller. This knowledge is useful in many fields of data, such as speech signals, photographs, and scanned images which implies a large number of dimensions. For such datasets, it is often advisable to perform dimensionality reduction prior to applying K-nearest neighbors (KNN) algorithm, in order to prevent curse of dimensionality from ruining the result of KNN. One such way that is widely accepted is to apply principal component analysis. In this report, the effects of PCA and random dropping of columns are examined on a dataset which features are completely uncorrelated with each other. Research on Mining Course Structure Supervisor: WONG, Raymond Chi Wing / CSE Student: ZHANG, Jinming / DSCT Course: UROP1000, Summer Music generation has witnessed significant advancements with the advent of deep learning techniques. Popular music generation models are constructed by LSTM or transformer, each has its own limitation. We propose a novel model for music generation based on Graph Neural Networks (GNNs). Our model leverages the inherent structure and relationships present in musical data by representing it as a graph. By incorporating GNNs, our model captures the dependencies and interactions between musical elements, enabling it to generate coherent and expressive musical compositions. Language-Guided Dense Prediction for Scene Understanding Supervisor: XU, Dan / CSE Student: LU, Weiqi / COSC Course: UROP1100, Fall Open-world semantic segmentation is a significant and challenging task in Computer Vision. It aims to leverage a dataset with a small portion of classes having ground truth segmentation masks to do the semantic segmentation of classes without pixel-level annotations. One approach is incremental learning, which detects novel classes based on the base classes in the dataset. However, this approach is often biased to the class types of the training data. To address this problem, CLIP-guided learning is introduced to leverage the rich semantic information from the image encoder and text encoder of Contrastive Language-Image Pretraining (CLIP). Following this idea, we propose OpenCLIP that is expected to do open-world semantic segmentation of a large number of novel classes.