DataTalks #18: The Power Couple: Deep Learning + Graphs

Our 18th meetup is hosted by 𝗧𝗮𝗯𝗼𝗼𝗹𝗮 and it's going to be the deepest, graphiest, most awesome one yet!
Uriel Singer is going to give us a marvelous intro to learning on graphs.
Next, 𝗬𝗮𝗺 𝗣𝗲𝗹𝗲𝗴, the one and only, will blow our mind by extending and developing the message passing network to a universal graph network model for accurate property prediction on nearly any graph dataset 🤯
(mind=blown)

𝗔𝗴𝗲𝗻𝗱𝗮:

🍕 18:30 - 19:00 - Gathering, snacks & mingling

📖 19:00 - 20:00 - First talk:
Learning on Graphs
by Uriel Singer
(Hebrew)

💪 20:10 - 21:10 - Second talk:
The unimaginable power of Graph Message Passing Neural Networks
by Yam Peleg
(Hebrew)

𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴 𝗼𝗻 𝗚𝗿𝗮𝗽𝗵𝘀 - Uriel Singer (Hebrew)

Images, video, speech, audio, text, time-series, tableau – tried all these domains? Have you tried the graph domain?
Graphs are a general language for describing complex systems. By looking around, you can realize that they are all around us: social networks, communication systems, internet, citations, economics, diseases, roads, and even neurons in our brains. In many systems, even if there is no formal graph, there is a network that defines the interactions between the components.
In this talk, we’ll go through the basics of graphs, understand the tasks we can tackle and learn the methods that are used in order to answer these questions.
By the end of the talk, one should be able to understand when graphs are the right domain to use, be able to formulate his task and use the right techniques.

𝗧𝗵𝗲 𝘂𝗻𝗶𝗺𝗮𝗴𝗶𝗻𝗮𝗯𝗹𝗲 𝗽𝗼𝘄𝗲𝗿 𝗼𝗳 𝗚𝗿𝗮𝗽𝗵 𝗠𝗲𝘀𝘀𝗮𝗴𝗲 𝗣𝗮𝘀𝘀𝗶𝗻𝗴 𝗡𝗲𝘂𝗿𝗮𝗹 𝗡𝗲𝘁𝘄𝗼𝗿𝗸𝘀 - Yam Peleg (Hebrew)

Graph message passing neural networks (MPNN) are connectionist models that capture the dependence of graphs via message passing between the nodes of graphs. Unlike standard graph neural networks, MPNN retains a state that can represent information from its neighborhood with arbitrary depth.
In this talk, we extend and develop the message passing network to a universal graph network model for accurate property prediction on nearly any graph dataset.
We will demonstrate (Live Coding) that this method outperforms any prior DL models on nearly any task given (We will test on QM9).
We will also review state-of-the-art graph neural networks of four domains, recurrent graph neural networks, convolutional graph neural networks, graph autoencoders, and spatial-temporal graph neural networks.

𝗡𝗼𝘁𝗲: This session will be practical, You are invited to bring your own laptop and join.