Graph Convolutional Networks with Random Weights
Independent implementation of a published GCN approach using randomly initialized (untrained) weights.
GCN
PyTorch
AI Researcher
Zeinab Rahbar
I develop graph representation learning methods for multimodal intelligence and AI for healthcare.
I study how to learn the hidden graph structure connecting multimodal data — text, images, and clinical measurements — instead of assuming a fixed one, so models can propagate information the way the underlying system actually works.
This matters most in healthcare, where the true relationships between a patient's scans, tests, and records are rarely known in advance. My approach combines graph neural networks with geometric deep learning to make that structure-learning both theoretically grounded and computationally efficient.
§ 01 — Focus
Research Interests
Core areas of ongoing work and research interest, spanning graph-based representation learning and its application to multimodal and biomedical data.
Graph Neural Networks
Learning over graph-structured data, including information propagation and generalization behavior in GNN-based models.
Geometric Deep Learning
Theory-driven, efficient processing of graph and non-Euclidean structured data.
Graph Representation Learning
Inferring latent graph topologies directly from heterogeneous, non-graph data.
Multimodal Learning
Unifying textual, visual, and structured signals within graph-based representations.
AI for Healthcare
Biomedical imaging and multimodal disease prediction from clinical and imaging data.
Thesis · Featured Research
Learning Latent Graph Structures for Multimodal Text and Vision Tasks
Graph structure learning aims to discover the hidden relationships within data automatically, rather than relying on a human-defined graph — this matters most for multimodal data, where complex interactions run across textual, visual, and biological signals. This thesis develops two graph structure learning methods for non-graph multimodal data: one at the node level, one at the graph level.
Step 1 — Multimodal Disease Prediction
A node-level, dynamic, task-oriented graph structure learning model for predicting Alzheimer's and Autism from multimodal medical data. Outperforms the baseline on both datasets, with the best reported result on ABIDE against existing state-of-the-art methods.
PyTorch Geometric
GNN
Biomedical
ABIDE
TADPOLE
Step 2 — Graph-based Visual Question Answering
The first multimodal approach to apply graph structure learning at the graph level, discovering the relational topology between textual and visual components. Competitive performance against the strongest existing graph-based methods on VQA v2.0.
GNN
Multimodal
VQA v2.0
- Degree
- M.Sc., Computer Engineering — Artificial Intelligence, University of Isfahan (2022–2025)
- Supervisors
- Dr. Peyman Adibi · Dr. Alireza Darvishy (ZHAW, Switzerland)
- Keywords
- Multimodal Learning, Graph Neural Networks, Graph Structure Learning
- Code
- Not yet public
§ 02 — Record
Experience
Professional, teaching, and mentorship roles.
-
HoopadVision, Isfahan
- Improved real-time fire detection accuracy using time-series analysis.
- Built a human action recognition system (running, falling) processing four camera streams at 15 FPS on a single GPU.
-
University of Isfahan — Supervisor: Dr. Peyman Adibi
- Developed course assignments and evaluated final projects.
-
University of Isfahan — Supervisor: Dr. Saeed Ehsani
- Conducted complementary online and on-site classes. Designed and evaluated assignments and final exams.
-
Arak University — Supervisor: Dr. Maryam Amiri
- Supported course delivery, assignment design, and student evaluation.
-
Yasan Academy
- Taught Python programming fundamentals to beginner-level students.
§ 03 — Applied work
Additional Project
Independent, self-directed work outside the thesis.
§ 04 — Background
Education
University of Isfahan
- Thesis
- Learning Latent Graph Structures for Multimodal Text and Vision Tasks
- Supervisors
- Dr. Peyman Adibi, Dr. Alireza Darvishy (ZHAW, Switzerland)
Arak University
§ 05 — Toolset
Skills
Programming
- Python
- C++
Deep Learning
- PyTorch
- TensorFlow
Graph Learning
- PyTorch Geometric
- DGL
Computer Vision & Deployment
- Ultralytics
- FastAPI
Tools
- Docker
- Git
- Linux
- VS Code
Spoken Languages
- Persian — Native
- English — IELTS 7.0
- French — A1