UROP Proceedings 2020-21

School of Engineering Department of Computer Science and Engineering 125 Human in the Loop Machine Learning Supervisor: HUI Pan / CSE Student: ZHAO Yizhe / COSC Course: UROP1100, Fall Human-in-the-loop means that humans train, test, or tune the AI system to help it achieve more reliable results. Human-in-the-loop can be performed to teach the system what features to look for and guide the system to get more accurate answers. To introduce a method to interact with Machine Learning Model, we learned principles of Reinforcement Learning and Imitation Learning. In this project, we did experiments on DQN and Behavior Cloning in OpenAI Gym environment to better understand the strategies and parameters of these methods. We read the paper of InfoGAIL to capture the basic idea of this model and get a preliminary understanding of Human-in-the-loop in Reinforcement Learning. Next step of this project is to do more experiments on different models and read more cutting-edge papers to further explore this area. Human in the Loop Machine Learning Supervisor: HUI Pan / CSE Student: ZHOU Xinrui / COGBM Course: UROP1100, Fall This project aims at introducing a method to interact with Machine Learning Model in a Web/Virtual Reality environment. The main tasks for this semester are to study GAN and algorithms for its different variations, also we reviewed and reproduced the model of some related papers that committed to this field. The research assistant Mr. Arthur Yau led weekly meetings where students presented what they have learned and implemented for the week. The main task for students were studying the materials provided by the research assistant and also implemented and reproduced some paper results and do some literature reviews. Each person in the group has a different level of understanding of the technology involved, and the research assistant distributed different tasks to us according to each person's actual situation and progress. In addition, the research assistant grasped a good schedule and arrangement, and used Asana as a collaboration tool so that we could all keep up with the progress. Human in the Loop Machine Learning Supervisor: HUI Pan / CSE Student: TAI Sung Chit / COSC Course: UROP1100, Spring Data poisoning attacks pollute the training data of a machine learning model, impeding model accuracy and producing biased results. Label flipping and outliers are examples of such attacks. In this study, we attempted to remove outliers and flip mislabelled data by utilizing human feedback in the training process of a pretrained model after identifying the unclean data using an outlier metric. Our results show that this new method sees slight improvement in the accuracy of the model. This method of training may see use in crowdsourcing projects where data is easily mislabelled and may be unclean for model training purposes.