Deep Graph Networks Reading Group

To join the reading group mailing list follow this link https://groups.google.com/forum/#!forum/graph-nets-reading-group/join

The group is managed by Francesco Lomonaco, Ph.D. candidate at the CSEE department in the University of Essex. If you want to present paper or have any questions, please email to: francesco.lomonaco@essex.ac.uk

The Reading Group

While Graph Network techniques are gaining importance and good results in practical applications there is also the necessity for dissemination activities that allows researchers to confront each other in the possible applications but most importantly in understanding the workflow of Graph Networks to accelerate the capability to use in practical applications as one of the state-of-the-art.

The goal of the reading group is to disseminate the most relevant techniques, applications, and papers about Graph Networks. We explicit choose to involve researchers from different domains to stimulate the exchange different perspectives of the same tool. The Target of the Reading group are Ph.D. student's as well as other researchers, interested in understanding methodology as well as in finding useful resources about this emerging topic.

The group involves people from different countries, universities, and departments. Usually each session is led by one speaker that present a paper, but we also arranged applications-focused sessions where multiple speakers present demo-code and discuss results and issues regarding domain and model specific problems.

The Rise of Graphs (why we are doing this)

During the last years, Deep Neural architectures brought an impressive increasing in the quality of models and predictions in computer vision. Convolutional networks, in fact, leverage local connection around pixels and share these weights to optimize computational cost [1].

This domain is mainly characterized by Euclidean structures or regular grids, which are defined by the fact the all the pixels respect certain rules about position and distances between them.

This positionality requirements impact some domains where the grid could be irregular or distances between points (i.e. pixels in images) are not fixed.

Social Networks, paper’s citation network, brain’s connections are examples of non-regular domain. The question is how and whether apply deep neural techniques in these domains but also how to represent the information that is encoded in this structure for various task. As highlighted by [2], graphs are typically locally connected structure and are theoretically suitable to represent hierarchical patterns.

Graphs in the end find their way to get into neural networks because they allow researchers to apply these frameworks to non-regular domains.

Graphs are an object that is composed by two elements: nodes and edges. Obviously, an image too could be represented as a graph: pixels are nodes and edges exist where two pixels are adjacent. [3] pointed out that there is a connection between the Graph Laplacian, compositionality, locality, and convolutions, that allow to use convolutions on graph structured data. [4] Demonstrate how to deal practically with convolutions on static graphs but its contribution since publication gave the opportunity to develop models that deal also with temporal aspect, (i.e.[5]), and to build different classes network architectures on graphs (Encoder or Attention based), that are helpful in a various range of tasks.

References

[1] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature 2015, vol. 521, no. 7553, p. 436, 2015.

[2] F. R. Chung and F. C. Graham, “Spectral graph theory”. American Mathematical Soc., 1997, no. 92.

[3] Joan Bruna, Wojciech Zaremba, Arthur Szlam, and Yann LeCun, “Spectral networks and deep locally connected networks on graphs”. In Proceedings of the 2nd International Conference on Learning Representations, 2013.

[4] Kipf, T. N., and Welling, M. 2017,”Semi-supervised classification with graph convolutional networks”. In ICLR.

[5] Pareja, Aldo, et al. "Evolvegcn: Evolving graph convolutional networks for dynamic graphs." arXiv:1902.10191