Graph neural networks: Variations and applications

Many real-world tasks require understanding interactions between a set of entities. Examples include interacting atoms in chemical molecules, people in social networks and even syntactic interactions between tokens in program source code. Graph structured data types are a natural representation for such systems, and several architectures have been proposed for applying deep learning methods to these structured objects. I will give an overview of the research directions inside Microsoft that have explored different architectures and applications for deep learning on graph structured data.

[Mostafa-cvjc]

Give this guy a cookie. Clearly explained, make my life easier as I now refer people to this instead of explaining them for hours and hours

[kennys]

Links or no links, this video is far more clear than even the talks of the ones who published the paper...

[beizhou]

On the eve of watching this presentation, I gave it a thump up because all the comments say it is phenomenal.

[jimmyshenmusic]

Thanks so much. This is very clear explained. A MUST SEE for GNN beginners.

[ethanjyx]

Great intro to a bunch of useful resources

[longliangqu]

can you share the great slides? it's so vivid!

[ShikhaMallick]

Does this method work for dynamic graphs? Since we need information about neighbours of every node, the adjacent nodes should be known prior. Also, in what format the graph is given as input? Is it an adjacency matrix or list?

[aixueerever]

The animation of message passing is so cool. Where can I steal the slides? lol

[atwinemugume]

Thank you for the talk, I have good info on this as of now...

[taku]

I have question in nlp applications, We all know there is graph relationship in a sentence, but we do not know what the relationship(edges) is, so how can we use it in nlp?

[bharathram]

How do I make such cool presentations? (Also which tool did he use to make this presentation slides?)

[crwhhx]

A bit confused about the networks representing edges. Which of the following is true? 1. Each edge is represented by a unique network, or 2. Edges of the same *type* are represented by the same network, each *type* is represented by a unique network?

[xianchen]

Wait how are they directed? Aren't they bidirectional? If the adjacency matrix is symmetric it is not a directed graph.. but a bidirectional one isn't it?

[vcool]

Microsoft has no related software to offer. As of 2018, look at "deepmind/graph_nets" and "dmlc/dgl" instead.

[Jirayu.Kaewprateep]

I am not good about those function parameters and the historical of it but one me interesting I rectangular shape he create to contains object!
Anyof shape is diagonal symmetry or they are horizontal and vertical symmetry⁉️
Otherwise people need to do like this all the time they taken pictures 🤸

[tomburns]

A few thoughts:
- Their integration algorithm seems like a bad inspiration of what SNNs do. It would make more sense to have time constants and more dynamics.
- They should not always look from the "god's eye view" considering in many cases it highly specific features or subgraphs of graphs which are important for a function/
- I notice they didn't compare their method to established graph theoretic methods. They should.

[marat]

As usual very exiting video with no links to recent papers. Very shame microsoft

[cy-tiln]

Why do not you talk a bit clearly ? Has talking been monetized a time ago ? Very bad presentation