Why DGL?

In the last few years, deep learning has enjoyed plenty of extraordinary successes. Many challenging tasks have been solved or close to being solved by Deep Learning, such as image recognition, rich-resource machine translation, game playing. These were made possible by a set of techniques that are composed of a number of representationally powerful building-blocks, such as convolution, attention and recurrence, applied to images, video, text, speech and beyond.

The development and deployment of these techniques often depend on the simple correlation of the given data; for example, CNN is based on the spatial correlation between nearby pixels while RNN family dwells on the assumption that its input is sequence-like.

More recently, there has been a steady flow of new deep learning research focusing on graph-structured data. Some of them are more conventional graph related problems, like social networks, chemical molecules and recommender systems, where how the entity interacts with its neighborhood is as informative as, if not more than, the features of the entity itself.

Some others nevertheless have applied graph neural networks to images, text or games. Very broadly speaking, any of the data structures we have covered so far can be formalized to graphs. For instance an image can be seen as grid of pixel, text a sequence of words… Together with matured recognition modules, graph can also be defined at higher abstraction level for these data: scene graphs of images or dependency trees of language.

To this end, we made DGL. We are keen to bringing graphs closer to deep learning researchers. We want to make it easy to implement graph neural networks model family. We also want to make the combination of graph based modules and tensor based modules (PyTorch or MXNet) as smooth as possible.

Committors

• Amazon Web Services AI Shanghai Lablet (ASAIL): Minjie Wang, Quan Gan, Mufei Li, Rui Ying, Tiajun Xiao, Tong He, Jian Zhang, Xiangkun Hu, Jinjing Zhou*, Zihao Ye*, Chao Ma*, Zheng Zhang
• Amazon Web Services Machine Learning: Da Zheng, Xiang Song, Israt Nisa, George Karypis
• NVIDIA: Dominique LaSalle, Xin Yao
• New York University: Lingfan Yu, Yu Gai, Jinyang Li

* Alumini

Contributors

Since released, DGL has received contributions from 200+ contributors. We are grateful to work with such an energetic community and looking forward to see more to join. See the list of contributors here.