@thomaskipf and @wellingmax strike back! in a new blog post w/ @Francesco_dgv@_JRowbottom et al we show that GCN-type models can be derived as gradient flows of Dirichlet-type energy and provably avoid low-frequency dominated dynamics https://t.co/IOuR2JBHFA
This is a super accessible explanation* of our gradient flow work and it's based on a recent arXiv version contanining new theoretical results and experiments
*Beware, great meme ahead
Here is last week's video of @Francesco_dgv, @_JRowbottom and @b_p_chamberlain presenting their paper "Graph Neural Networks as Gradient Flows"!
https://t.co/EnAaq0nu0a
Looking forward to giving an invited (and in-person!) talk in @jure's group tomorrow @Stanford. I'll talk about our latest take on physics-inspired learning on graphs using non-linear oscillators.
arxiv: https://t.co/bC37KD377u
code: https://t.co/HBjXUQxX3h
w/ @mmbronstein
I am happy to share a recent work on energy functionals giving rise to GNN equations via gradient flows 🧵
https://t.co/FlbFz7sVFP
This is joint work with @_JRowbottom*, @b_p_chamberlain, T. Markovich, and @mmbronstein
1/4 Hope to see friends new & old tomorrow 4:30pm GMT / 8:30am PT poster session 6 #NeurIPS2021@_JRowbottom@Francesco_dgv and I will occupy the prime virtual real estate known as Spot E3 with BLEND https://t.co/sSuIDpaiz0
#GNNs are related to PDEs governing information diffusion on graphs. In a new paper with @b_p_chamberlain James Rowbottom @migorinova@stefan_webb@emaros96 we study a new class of Neural Graph Diffusion PDEs
Blog post: https://t.co/sxVcS1pWmK
Paper: https://t.co/upMNI0EyW8