1. Neural Markov Logic Networks https://openreview.net/pdf?id=SkeGvaEtPr … motivated by markov logic networks, but define an exponential family for modeling distributions. Of more general interest, but also used for molecular graph generation. 2. Graph Warp Module: https://openreview.net/pdf?id=S1l66nNFvB …
-
-
Prikaži ovu nit
-
.. this adds an extra module to the GNN which has three components (including virtual nodes). Some of these ideas were already in the original MPNN paper I think. 3. Generative model for molecular distance geometry https://openreview.net/pdf?id=S1x1IkHtPr … this has been a pretty active area
Prikaži ovu nit -
4) https://openreview.net/pdf?id=BkxthxHYvr … here they use disentangled representations to give a conditional generative model for molecules. The properties are input separately and disentangled from the structural representations.
Prikaži ovu nit -
Staying with generation is 5) https://openreview.net/pdf?id=ryxQ6T4YwB … dubbed GraphNVP, as should be clear the name comes from RealNVP and gives a flow based invertible model for graph generation. 6) https://openreview.net/pdf?id=S1esMkHYPr … called GraphAF is also a flow based model, but autoregressive.
Prikaži ovu nit -
There are many more that look interesting, but would take some effort to go through. For GNNs, but not chemistry specific, here are two very interesting submissions that try to leverage more of the geometry 7) https://openreview.net/pdf?id=BJg73xHtvr … GNNs for hyperbolic space representations.
Prikaži ovu nit -
8) Geometric GCNs https://openreview.net/pdf?id=S1e2agrFvS … try to inject the underlying geometric information while aggregating messages.
Prikaži ovu nit -
I made a note of this, but forgot to post. This one I actually really enjoyed 9) Directional message passing https://openreview.net/pdf?id=B1eWbxStPH … messages also incorporate directional info such as bond angles and rotations; while being equivariant.
Prikaži ovu nit
Kraj razgovora
Novi razgovor -
Čini se da učitavanje traje već neko vrijeme.
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.