As a fifth-year Computer Science Ph.D. student at NJIT, I’m passionate about combining theoretical analysis with practical implementation. My research focuses on integrating data-driven learning with structured reasoning to build solutions to real-world problems, with experience in both supervised and reinforcement learning.
Model: VisionLight-R (uses a ResNet with spatial-temporal attention);
VisionLight-A (adopts anchor-based feature encoding inspired by YOLO);
VisionLight-T (applies a Transformer-based design for long-range traffic pattern modeling).
Combined with Traffic Flow Fluctuation Uncertainty Assessment Mechanism - Entropy Attention Mechanism
input: raw surveillance video frames
action-space: decisions to KEEP or SHIFT current traffic light phase (every 5 seconds)
pretrain: SUMO
fine-tune: Carla
metrics: average reward (AR), average stopping time (AST), average queue length (AQL)
This project traces back the key factor magnetic area that determines the duration of an intense flare.
Model: Resnet18 & ViT
Training set: all JSOC sharp_cea_720s magnetogram FITs sliced 24 hours up to the beginning of a M & X class flare, year 2010 - 2022
Model discovers that single polarity area contributes more to flare duration rather than polarity inversion line, which is counter intuitive
This project is an implementation of a mutated evolution algorithm
Target: 100 triangle with different color and location stack together to simulate target image
Next generation & mutation: relocate 3 apexes of triangle & randomized a color
Loss: summation of pixel value difference between generated image and target after each geneartion
There are RNA sequences linked as a graph, but some of them are not properly labeled.
We utilize a customized Graph Neural Network to learn the relation within graph and extrapolate these label.
Model: a customized GNN, structure shown above
Dataset: as shown above
Platform & language: Android & Java
Backend Database: Google - Firebase
This Instant Message App is enable the user to create account, login, send friend request, chat, view chatting history and account management etc.
Model: customized DQN based structure
input: queue length, number of vehicle, phase, total waitting time, vehicle position matrix
pretrain: 20,800+ steps pretrain on randomly generated fixed signal
train: 78,000+ steps on real word cross traffic flow
Computer graphic rendering from scratch practice:
p1 - Monte Carlo path tracing
p2 - Blinn-Phong Bilinear texture shader
p3 - Normal shading
p4 - Vanilla ray tracing
p5 - Ray tracing acceleration grid
Iteractive Cloth Simulation URL portal:
https://genwei-zhang.github.io/14_cloth/
This is a webased cloth simulation, click the URL above to interact with the cloth.
Model: yolov8s
training dataset: 150+ screenshot sample with manual boxes labels
Three python thread:
- First one for vertical view angle button controlling
- Second one for horrizontal view angle button controlling
- Third one to send screenshot to model and acquire box coordinates
Author: Y Abduallah, JTL Wang, P Bose, G Zhang, F Gerges, H Wang
Conference: The annual International FLAIRS Conference
URL: https://doi.org/10.32473/flairs.v35i.130564
Author: Y Abduallah, JTL Wang, P Bose, G Zhang, F Gerges, H Wang
Conference: arXiv preprint arXiv:2205.02447
URL: https://doi.org/10.32473/flairs.v35i.130564
Author: H Jiang, Q Li, Z Hu, N Liu, Y Abduallah, J Jing, G Zhang, Y Xu, W Hsu, JTL Wang, H Wang
Conference: arXiv preprint arXiv:2211.02278
URL: https://doi.org/10.32473/flairs.v35i.130564