Graph neural networks are uniquely suited for unstructured data like 3D meshes and social networks where traditional CNNs struggle.
Key graph theory concepts are explained, including nodes, edges, directed vs undirected graphs, and homogeneous vs heterogeneous graphs.
An adjacency matrix is suitable for small graphs, while an adjacency list is more spatially efficient for large, sparse graphs.
The core deep learning process of training and inference is reviewed, along with the distinction between supervised and unsupervised learning.
GNNs can perform predictions at the graph level (molecule properties), node level (community detection), and edge level (recommendation systems).
GNNs are trained using the message passing algorithm to create node embeddings, followed by a transductive split for training and validation sets.
Popular frameworks like DGL, PyTorch Geometric, and TensorFlow GNN simplify the implementation of graph neural networks.
The talk concludes with a recap of key takeaways, including graph modeling, data representation, prediction tasks, and the message passing algorithm.
The Q&A session addresses audience questions about data leakage in transductive splits, applying GNNs to semantic knowledge graphs, and comparing graph embeddings to word embeddings.
56:46Jodie Burchell
28:55Han Xiao
46:51
24:44Tomaz Bratanic
29:11Michael Hunger
24:42Martin O'Hanlon
29:50Ekaterina Sirazitdinova
56:55Ekaterina Sirazitdinova
Jobs that call for the skills explored in this talk.
The University Of Göttingen
Cypress Semiconductor Corporation
IU Internationale Hochschule
IU Internationale Hochschule
Cypress Semiconductor Corporation