Education

• 

  Carnegie Mellon University

  Aug 2023 – Aug 2024

  Master of Educational Technology and Applied Learning Sciences

  Pittsburgh, PA

• 

  University of Michigan

  Aug 2021 – Apr 2023

  BSc in Computer Science

  Ann Arbor, MI

• 

  Shanghai Jiao Tong University

  Sep 2019 – Aug 2023

  BSc in Electrical and Computer Engineering

  Shanghai, China

Research Projects

PS: Some projects may not be fully visible here. Please reach out if you are interested in more details.

Towards Understanding the Trade-off between Sense of Agency and Performance in Assisted Target Selection in Virtual Reality

Muzhe Wu, Byungjoo Lee, David Lindlbauer

Available Soon

Abstract

Virtual Reality (VR) systems enable immersive experiences while providing users with efficient interaction techniques for tasks such as selection or manipulation. With the increased integration of AI-powered intelligent assistants, these interaction techniques continue to improve users' performance by anticipating their goals and actions. It is yet unclear, however, if such advanced interaction techniques negatively influence users' sense of agency. We explore this trade-off through a user study (N=12) on target selection tasks while varying the levels of assistance and task difficulty. We measured the sense of agency with the intentional binding paradigm and questionnaire, and compared this data with quantitative performance metrics. Our results reveal that while a higher level of assistance improved performance, the sense of agency remained stable in difficult task settings and increased in easy task settings, which might be due to participants becoming more proactive and attentive under high-assistance conditions. We provide guidelines for future intelligent interaction techniques that aim to balance performance with the sense of agency.

New Ears: An Exploratory Study of Audio Interaction Techniques for Performing Search in a Virtual Reality Environment

Muzhe Wu*, Yi-Fei Cheng*, David Lindlbauer

To Appear at IEEE ISMAR 2024

Abstract

Efficiently searching and navigating virtual scenes is essential for performing various downstream tasks and ensuring a positive user experience in VR. Prior VR interaction techniques for such scenarios predominantly rely on users' visual perception, which contrasts with physical reality, where people typically rely on multimodal information, especially auditory cues, to guide their spatial awareness. In this work, we explore the potential of leveraging auditory interaction techniques to enhance spatial navigation in virtual environments. We drew inspiration from prior distant interaction techniques and developed four approaches to augmenting how users hear in the virtual environment: Audio Teleportation, Audio Cone, Ninja Ears, and Boom Mic. In a comparative user study (N = 25), we evaluated these approaches against a baseline teleportation technique in a search task, where participants traversed a virtual environment to locate target items. Our results suggest that several of our audio interaction techniques may enable more efficient search behaviors while enhancing overall user experience. However, not all techniques were appreciated equally, suggesting that careful attention to their design is critical for ensuring their effectiveness. We conclude by discussing the potential implications of our results for future audio interaction technique designs.

On "AR-Based Intelligent Tutoring for Physical Task Learning"

Muzhe Wu*, Haocheng Ren*, Gregory Croisdale, Anhong Guo, Xu Wang

Michigan AI Symposium 2022 (Best Demo )

In Submission

Poster

Learning to solve a Rubik's Cube requires the learners to repeatedly practice a skill component, e.g., identifying a misplaced square and putting it back. However, for 3D physical tasks such as the Rubik's Cube, generating sufficient repeated practice opportunities for learners can be challenging, in part because repeated configuration of physical objects is strenuous. We propose Rubikon, an intelligent tutoring system for learning to solve the Rubik's Cube. Rubikon reduces the necessity for repeated manual configurations of the Rubik's Cube without compromising the tactile experience of handling a physical cube. The foundational design of Rubikon is an AR setup, where learners manipulate a physical cube while seeing an AR-rendered cube on the screen. Rubikon automatically generates configurations of the Rubik's Cube to target learners' weaknesses and help them exercise diverse knowledge components. A between-subjects experiment showed that Rubikon learners scored 25% higher on a post-test compared to baselines.

On "Early-Stage Risk Identification in Responsible AI"

Muzhe Wu*, Yanzhi Zhao*, Shuyi Han, Michael Xieyang Liu, Hong Shen

In Submission

Website

ActiveAI: The Effectiveness of an Interactive Tutoring System in Developing K-12 AI Literacy

Ying-Jui Tseng, Gautam Yadav, Xinying Hou, Muzhe Wu, Yun-Shuo Chou, Claire Che Chen, Chia-Chia Wu, Shi-Gang Chen, Yi-Jo Lin, Guanze Liao, and Kenneth R. Koedinger

ECTEL 2024

Abstract DOI PDF

As we witness groundbreaking advancements in Artificial Intelligence (AI), it is clear that the next generation must be equipped with AI literacy: the skill to interact, evaluate, and collaborate with AI systems. This study introduces ActiveAI, a scalable web-based tutoring system aligned with AI4K12's five big ideas in AI, designed to foster AI literacy among K-12 students through active learning and interaction with intelligent agents. A controlled classroom study involving 171 middle school learners was conducted to assess the effectiveness of ActiveAI in fostering AI literacy skills and competency toward AI. Results showed that, compared to students in the tell-and-practice control condition, students who used ActiveAI exhibited higher post-test performance in the module about how next-word prediction and temperature work in large language models. Students also developed higher self-reported competence toward AI after using ActiveAI than in the control condition. We conclude by suggesting assessment designs that promote deeper engagement with AI concepts by addressing students' common misconceptions, like “AI thinks just like humans”, in K-12 AI literacy education.

Prototypes

How Do You See The World

Simulate visual impairments in VR environments to educate people about these conditions and increase awareness.

Video Code

Mini-World

Replicate the sense of co-presence in web browsing with Chrome Extension.

XR Goal Tracker

Prototype immersive XR goal tracking experience exploring concepts of personification and fading.

Code

Wiki-Learn

Enhance informal learning on Wikipedia with LLM-based instruction and assessment generation with Chrome Extension.

Code

Other Projects

Why Antiwork: A RoBERTa-Based System for Work-Related Stress Identification and Leading Factor Analysis

Leverage a subreddit as a data source, create and train a model to detect antiwork sentiments, reveal that lack of authority, frustrating recruitment experiences, and unfair compensation are major contributors to employee dissatisfaction and antiwork sentiments.

Arxiv

Towards Understanding the Relationship between Misinformation and Engagement for Online Medical Videos

Examine the feasibility of computationally understanding the relationship between misinformation and engagement from different modalities in online medical videos.

Slides

Auxiliary Variables Improve Group Accuracy without Group Information

Validate the effectiveness of auxiliary variables in the first stage of the JTT/BAM algorithm resolving the spurious correlation problem with fine-grained datasets.

Paper Code

FAD: Feature Alignment Discriminator for Text Summarization

Building block useful in the fine-tuning process for text generators like BART, which addresses problems of discreteness in adversarial learning for NLP, better captures the word distribution, and achieves SOTA ROUGE score of abstractive text summarization in DailyMail/CNN dataset.

Paper

Retro Game API for Reinforcement Learning

Jim-Team

Reinforcement learning simulation API for retro games (e.g., Pokemon Gold), built on OpenAI Gym-Retro package, featuring utility classes (recorder, interactor, and dataset), vision transform classes (random cropping, random convolution, and gaussian noise) and a GUI visualizing observation space and state information.

Mask Distribution Simulator

C++ program that simulates the distribution of masks among the cities with a certain mask production capacity during the COVID-19 pandemic.

Code