Feiyan Zhou 周菲嫣

Northeastern University

MS in Khoury College of Computer Science

Email: zhoufeiyn AT gmail DOT com

    

Biography

I am currently an MSCS student at Northeastern University. I previously interned with Meta's Core Ads Generation team, where my research focused on multimodal media generation, including audio foley and joint audio-video generation.

Before that, I spent four years as a Senior Research Engineer at the China Electric Power Research Institute (CEPRI). My research focused on deep learning-based fault diagnosis for distribution networks and vision-based defect detection and localization for power equipment.

I received both my bachelor's and master's degrees in Electrical Engineering from Xi'an Jiaotong University.

Projects (Published)

WavFlow
WavFlow: Audio Generation in Raw Waveform Space
Developed a high-fidelity audio foley generation model without codec- or VAE-based latent compression, using an MMDiT architecture with waveform patchification, amplitude lifting, and direct x-prediction. Trained from scratch on 5M text-video-audio clips, it achieves competitive performance with latent-space methods.

[project page]

GameGen demo
Learning to Play: Real-Time Interactive Game Generation with Diffusion Forcing
Built and trained a real-time generative game model using Diffusion Forcing in VAE latent space, with recurrent latent-state memory for long-horizon temporal consistency. The model conditions on player actions and propagates latent state across generation steps, enabling interactively consistent Super Mario gameplay at 16 FPS with continuous rollouts exceeding 2 minutes.

[project page]

QueryMinds: LLM-Powered Interactive SQL Learning and Assessment System
Built and deployed a full-stack AI-assisted SQL learning platform using Django and MySQL on Google Cloud (App Engine + Cloud SQL), integrating GPT-4 for adaptive SQL question generation, semantic answer grading, personalized feedback, and instructor analytics.

[project page]

Dynamic Simulation Testing Systems for Urban Intelligent Distribution Power Networks
Developed a hybrid digital-physical testbed for next-generation intelligent distribution networks, integrating physical equipment with real-time simulation and cloud-based monitoring. Built automated scenario testing for system-level validation across fault detection, device interaction, and remote control. The platform supported 96 test scenarios and was deployed across 12 provincial power companies.

[project page]

Vision-Based Overheating Detection and Localization in Power Transmission Lines
Proposed a three-stage cascaded DL detection method: cropping component image patch from whole image, segmenting component region in component image patch and locating overheating defect in component region, which removes the interference background step by step and finally realizes the heating diagnosis of insulator and drainage plate in the transmission line.
3rd Price in 6th China State Grid Corporation’s Youth Innovation and Creativity Competition

[project page] [code and dataset]


Publications (show selected / show all)

Honors

Academic Service


© Feiyan Zhou | Last update: April 2024