AdityaChaudharyCaseStudy0nGraphNeuralNetwork

It is a work od A.Choudhary .

1. Aditya Chaudhary
IUSSTF-VITERBI Program 2023 Presentation
July 13, 2023
Faculty Supervisor: Prof. Ajitesh Srivastava
Graduate Collaborator: Majd Al Aawar
Electrical and Computer Engineering
University of Southern California


Graph Neural Networks for Point
Cloud Classification

2. Basic Graph Adj Matrix Node Features
Graphs?

3. Graph Neural Networks vs Convolutional Neural Networks

4. Problem & Motivation
●Problem Statement: Segmentation and Detection using point clouds
●Brain Imaging
○Early-stage Alzheimer's Disease
○Traumatic brain injury
○Autism Spectrum Disorder

●Social networks, biological networks, and brain connectivity networks.

●Manufacturing
○Deviations
○Minimize defects

●Aiding in the identification of potential therapeutic targets and
predicting molecular properties.

5. Background
●Existing work
○CNNs
○LSTMs
○SVMs

●How to leverage the growth in GNNs in the last decade ?
○Transform point clouds to a graph
○Apply GNNs on the graph for Node classification (segmentation)
and Graph classification (detection)

6. Background: Graph Construction
●Existing methods include
○Epsilon-neighborhood Method
○KNN method

●Drawback:
○Our approach preserves all distances using a
sparse graph.
○We drastically reduce the number of edges
required to represent the data while maintaining
the rigidity of the structure

7. Background: Graph Classification
●Graph Neural Networks (GNNs)
○Excel at analyzing graph-structured data
○Capture intricate relationships and patterns of complex networks.
○Types
■Graph Isomorphic Networks
■Graph Convolutional Networks
■Graph Attention Network

8. Approach: Graph Construction
Tetrahedralization-based
Algorithm to generate a graph
using KNN from point cloud for
graph identification

9. Approach: Graph Construction (Summary)
Tetrahedralization-based
Algorithm to generate a graph
using KNN from point cloud for
graph identification

10. Approach: Graph Classification

11. Results: Graph Classification
Ε- neighborhood MethodDPT Algorithm

12. Results: Graph Classification

13. Future Work
●Implement the GIN Model to actual brain models to predict
early stage alzheimer's disease

●Use more edge attributes and custom node features to
boost the model

●Improve graph creation algorithm to capture more intrinsic
and complex relation between nodes and edges

●Explore more GNN models (GAT’s, GraphSAGE, GAGAN) to
better fit our custom algorithm and compare results.

14. Acknowledgement
Ajitesh Srivastava Majd Al Aawar

15. Acknowledgement
Dr. Raghavendra Andy Jones-Liang

16. Q & A
Thank you
Time for