Stanford CS224W: ML with Graphs | 2021 | Lecture 19.3 - Design Space of Graph Neural Networks
ฝัง
- เผยแพร่เมื่อ 10 ก.ค. 2024
- For more information about Stanford’s Artificial Intelligence professional and graduate programs, visit: stanford.io/2ZvXEBp
Jure Leskovec
Computer Science, PhD
In previous lectures, we have discussed the important topic of GNN architecture design. Here, we introduce 3 key aspects in GNN design: (1) a general GNN design space, which includes intra-layer design, inter-layer design and learning configurations; (2) a GNN task space with similarity metrics so that we can characterize different GNN tasks and, therefore, transfer the best GNN models across tasks; (3) an effective GNN evaluation technique so that we can convincingly evaluate any GNN design question, such as “Is BatchNorm generally useful for GNNs?”. Overall, we provide the first systematic investigation of general guidelines for GNN design, understandings of GNN tasks, and how to transfer the best GNN designs across tasks. We release GraphGym as an easy-to-use code platform for GNN architectural design. More information can be found in the paper: Design Space for Graph Neural Networks: arxiv.org/abs/2011.08843
To follow along with the course schedule and syllabus, visit:
web.stanford.edu/class/cs224w/
0:00 Introduction
0:24 Key Questions for GNN Design
1:30 Background: Terminology
2:27 Recap: GNN Design Space
5:03 Summary: GNN Design Space
6:01 A General GNN Task Space
10:10 Evaluating GNN Designs
12:47 A Guideline for GNN Design
14:37 Understanding GNN Tasks
16:23 Transfer to Novel Tasks
17:44 GNN Design Space: Summary
#machinelearning #machinelearningcourse
Very nice lectures of this entire course and very interesting research of this video. Thank you very much for the course.
One of the best design space approach to experiment quickly, cover more models, and share meaningful results.
Pretty impressive.
Not sure if the experimental design approach has been used in design space search ?
The whole class keeps presenting new concepts without clear introduction, or giving out some intuitions that are not convincing enough. I am so disappointed to stanford, especially online.