Join my FREE course Basics of Graph Neural Networks ([ Ссылка ])! This video dives into the paper "Learning to Simulate Complex Physics with Graph Networks" from DeepMind and interviews one of its authors, Jonathan Godwin.

Original Paper: [ Ссылка ]
Simulator video source: [ Ссылка ]
Project Code & Datasets: [ Ссылка ]

Mailing List: [ Ссылка ]
Discord Server: [ Ссылка ]
Blog: [ Ссылка ]
Patreon: [ Ссылка ]

References:
- Daniel Holden's talk from UbiSoft: [ Ссылка ]
- SPlisHSPlasH project: [ Ссылка ]
- "Data-driven Fluid Simulations using Regression Forests": [ Ссылка ]
- "Latent Space Physics: Towards Learning the Temporal Evolution of Fluid Flow": [ Ссылка ]
- "Learning to Predict the Cosmological Structure Formation": [ Ссылка ]
- "Graph Networks as Learnable Physics Engines for Inference and Control": [ Ссылка ]
- "Relational inductive biases, deep learning, and graph networks": [ Ссылка ]


Chapters
- 00:00 - Intro
- 02:24 - Why learnable physics engines?
- 03:15 - Literature survey
- 05:51 - High level overview of learning process
- 09:04 - Understanding the role of Graph Networks
- 13:15 - Interview with Jonathan Godwin introduction
- 14:26 - What are the key contributions of this paper?
- 16:40 - Why does this generalize so well?
- 18:23 - What about the "butterfly effect"?
- 21:08 - Possible application areas
- 25:35 - What framework for implementing/scaling this?
- 28:47 - Open questions and challenges
- 32:35 - What other research areas excite you, outside of GNNs?

свернуть