UROP Proceedings 2020-21

School of Science Department of Mathematics 59 Research in AI and Machine Learning Supervisor: ZHANG Tong / MATH Student: SHUM Ka Shun / DSCT Course: UROP1100, Fall Story Completion is a challenging task of generating the missing plot for an incomplete story. In this report, we give a generative adversarial networks based on Transformer (T-CVAE-GAN) for the missing sentence generation. Our approach uses shared attention layers for Transformer encoder and decoder and a latent variable provided by a generator. Through this generator, we could generate diverse reasonable plots. BLEU eDvaluation is used to show that our model can achieve state-of-the-art results in terms of readability, diversity and coherence. Research in AI and Machine Learning Supervisor: ZHANG Tong / MATH Student: WU Hsuan-cheng / MATH-IRE Course: UROP1100, Fall The agent learns a model of the environment. The advantage of the modelbased is that agent is allowed to plan by thinking and analyzing ahead since it has a model, and then turn into a learned policy. If it works, then the speed of improvement is more efficient than those methods that don’t have a model. On the other hand, the disadvantage of modelbased is that a groundtruth model of the environment is usually not available to the agent, then the agent need to learn the model purely from its experience. It is highly possible that there will be a bias of the model and the environment as time passed, so the optimal behaviors are not the same as what the agent learns. The most famous example of the modelbased methods is AlphaZero. Research in AI and Machine Learning Supervisor: ZHANG Tong / MATH Student: ZHONG Dingyan / DSCT Course: UROP1100, Fall Object detection is one of the heated topics in computer vision nowadays. Plenty of algorithms based on deep learning have arisen over the past few years. Typically, a 2D object detection method uses either anchors or key points to estimate the locations of objects. The difference between those two kinds of methods is only the level of abstraction. 3D object detection yields more difficulties as it concerns depth and rotation angles, which are hard to predict from one image, especially in a monocular case. This project is organized by reviewing and reproducing existing algorithms, together with proposing solutions to the remaining problems.