Preview: Graph NNs
summary
Overview of graph representation learning and GNNs with concrete system results: shallow encoders are transductive lookups while deep GNN encoders are inductive and use neighborhood aggregation; an in-storage accelerator eliminates host data movement and speeds up training by an order of magnitude; a serverless CPU-only pipeline matches or exceeds GPU training; IO-reduction plus pipelining trains a billion-edge graph on one machine; a hybrid distributed method cuts communication and boosts training up to 7x while scaling to multiple billions of edges.
AI-written summary · about
chapters (11)
- 0:00Intro
- 0:30Graph Learning at OSDI/ATC '21
- 2:56Graph Learning: Node Classification
- 3:46Graph Learning: Edge Prediction
- 4:29Graph Classification
- 5:30Graph Embeddings
- 6:52Shallow Encoding
- 8:16Deep Encoding
- 11:15Graph Learning: Trends
- 11:38Graph Learning: Systems Challenges
- 15:34Graph Learning Summary
related talks
01h 01mToward Universal Graph Representations: Foundations and Frontiers of Graph Foundation Models
39mUsing Graph Neural Networks to Embrace the Dependency Within Your Data
19mNeuGraph: Parallel Deep Neural Network Computation on Large Graphs
29mGraph Neural Networks + LLM on
18mTC-GNN: Bridging Sparse GNN Computation and Dense Tensor Cores on GPUs
59mTransforming DataFrames: Intro to Graphs & GNNs with PyTorch Geometric
14mLeapGNN: Accelerating Distributed GNN Training Leveraging Feature-Centric Model Migration
01h 16mBoot camp on graph foundations models