A Tale of Three Implicit Planners and the XLVIN agent

A video presentation by Petar Veličković about implicit planners, which could be seen as a middle-ground between model-based and model-free approaches for RL planning problems. The talk covers three popular implicit planners: VIN, ATreeC and XLVIN. All three focus on the recently popularised idea of algorithmically aligning to a planning algorithm, but with different realisations.

Graph Machine Learning research groups: Yaron Lipman

I do a series of posts on the groups in graph research, previous post is here. The 25th is Yaron Lipman, a professor in Israel, who has been co-authoring many papers on equivariances and the power of GNNs.

Yaron Lipman (~1980)
- Affiliation: Weizmann Institute of Science
- Education: Ph.D. at Tel Aviv University in 2008 (advisors: David Levin and Daniel Cohen-Or)
- h-index 41
- Interests: geometric deep learning, meshes, 3d point clouds, equivariant networks

GML Express: large-scale challenge, top papers in AI, and implicit planners.

Another issue of my newsletter. So I finally solved a struggle for me what to write this newsletter about: news or insights. GML express will cover news (which you mostly should get anyway from this channel) and GML In-Depth should cover my insights.

In this GML express you will find a bunch of learning materials, recent video presentations, blog posts, and announcements.

Fresh picks from ArXiv
This week on ArXiv: large-scale node prediction competition, GNNs for 3D objects, and improved performance for imbalanced classification 👯‍♀️

If I forgot to mention your paper, please shoot me a message and I will update the post.

GNNs
* OGB-LSC: A Large-Scale Challenge for Machine Learning on Graphs with Jure Leskovec
* Concentric Spherical GNN for 3D Representation Learning with Le Song
* GraphSMOTE: Imbalanced Node Classification on Graphs with Graph Neural Networks
* Catastrophic Forgetting in Deep Graph Networks: an Introductory Benchmark for Graph Classification

Applications
* A graph theoretical approach to the firebreak locating problem
* Exploiting Isomorphic Subgraphs in SAT (Long version)
* ChronoR: Rotation Based Temporal Knowledge Graph Embedding
* Code Completion by Modeling Flattened Abstract Syntax Trees as Graphs
* NetVec: A Scalable Hypergraph Embedding System

Software
* GNNerator: A Hardware/Software Framework for Accelerating Graph Neural Networks
* Characterizing the Communication Requirements of GNN Accelerators: A Model-Based Approach

DIG: Dive into Graphs library

A new python library DIG (and paper) in PyTorch for several graph tasks:
* Graph Generation
* Self-supervised Learning on Graphs
* Explainability of Graph Neural Networks
* Deep Learning on 3D Graphs

GNN User Group: meeting 3

Third meeting of GNN user group will include talks from Marinka Zitnik, Kexin Huang, and Xavier Bresson, talking about GNNs for therapeutics and combinatorial optimization. It will be tomorrow, 25th March.

Model distillation for GNNs

Model distillation is the approach to train a small neural network called student given a large pretrained neural network called teacher. Motivation for this is that you want to reduce the number of parameters of your production model as much as possible, while keeping the quality of your solution. One of the first approaches for this was by Geoffrey Hinton, Oriol Vinyals, Jeff Dean (what a combo) who proposed to train student network on the logits of the teacher network. Since then, a huge amount of losses has appeared that attempt to improve performance of student network, but the original approach by Hinton et al. still works reasonably well. A good survey is this recent one.

Surprisingly, there were not many papers on model distillation for GNNs. Here are a few examples:

* Reliable Data Distillation on Graph Convolutional Network SIGMOD 2020
* Distilling Knowledge from Graph Convolutional Networks CVPR 2020
* Extract the Knowledge of Graph Neural Networks and Go Beyond it: An Effective Knowledge Distillation Framework WWW 21

But these approaches were not convincing enough for me to say that knowledge distillation is solved for GNNs, so I'd say it's still an open question to research. I have also tried to train MLP model on GNN logits to see if we can replace GNN with MLP at inference time, and apparently you can get an uplift wrt vanilla MLP trained on targets; however, the performance is not as good as for GNNs. One of the good examples of significantly reducing the number of parameters of GNNs is the recent work on LP for node classification: LP has 0 parameters and with C&S it gets some MLP parameters but not as many as for GNNs.

Tensorflow GNN libraries

If I miss some libraries in tensorflow, please let me know, I will update the list.

* tf_geometric (paper)
* tf2-gnn (microsoft)
* tf-gnn-samples (microsoft)
* Spektral (documentation)
* graph_nets (deepmind)
* gnn (documentation)

GNN Explainer UI

Awesome tool that provides user interface for visualizing edge attributions on trained GNN models and compare different explanation methods. An explanation method takes as input a GNN model and a single sample graph and outputs attribution values for all the edges in the graph. Each explanation method uses a different approach for calculating how important each edge is and it is important to evaluate explanation methods as well.

Fresh picks from ArXiv
This week on ArXiv: tricks to improve GNNs, unlearning problem on graphs, and cheating on TOEFL with GNNs ✍️

If I forgot to mention your paper, please shoot me a message and I will update the post.

GNNs
* A nonlinear diffusion method for semi-supervised learning on hypergraphs with Austin R. Benson
* Bag of Tricks of Semi-Supervised Classification with Graph Neural Networks
* Self-supervised Graph Neural Networks without explicit negative sampling
* Graph Unlearning
* InsertGNN: Can Graph Neural Networks Outperform Humans in TOEFL Sentence Insertion Problem?
* Beyond permutation equivariance in graph networks
* Knowledge-aware Contrastive Molecular Graph Learning
* Autism Spectrum Disorder Screening Using Discriminative Brain Sub-Networks: An Entropic Approach

Survey
* A Comprehensive Survey on Knowledge Graph Entity Alignment via Representation Learning

Video and slides: GNN User Group meeting 3

In the third meeting of GNN user group, there are two talks:
* Therapeutics Data Commons: Machine Learning Datasets and Tasks for Therapeutics by Marinka Zitnik and Kexin Huang (Harvard)
* The Transformer Network for TSP by Xavier Bresson (NTU)

Slides are available in their slack channel.

Pytorch Geometric tutorial

Awesome tutorials on how to program your GNNs with PyTorch Geometric. I often say that the best way to learn about GNNs is through coding, so if you are new I would definitely recommend checking it out. There are upcoming sessions soon, if you want to do it live.

Graph Machine Learning research groups: Mingyuan Zhou

I do a series of posts on the groups in graph research, previous post is here. The 26th is Mingyuan Zhou, a professor at the University of Texas, who has been working on statistical aspects of GNNs.

Mingyuan Zhou (~1985)
- Affiliation: The University of Texas at Austin
- Education: Ph.D. at Duke University in 2013 (advisors: Lawrence Carin)
- h-index 30
- Interests: hyperbolic graph embeddings, bayesian GNNs, graph auto-encoders

Insights from Physics on Graphs and Relational Bias

A great lecture with lots of insights by Kyle Cranmer on the inductive biases involved in physics. Applying GNNs to life science problems is one of the biggest trends for ML and it's exciting to see more and more cool results in this area.

Open Research Problems in Graph ML

I thought I would make my first subscriber-only post on the open research problems in graph ML. These are the problems that I have thought a lot and think can have a transformational impact not only on this field, but also on the applications of graph models to other areas.

Fresh picks from ArXiv
This week on ArXiv: new heterophily datasets, improved inference and expressive power for GNNs 🦹‍

If I forgot to mention your paper, please shoot me a message and I will update the post.

GNNs
* New Benchmarks for Learning on Non-Homophilous Graphs
* Adaptive Filters and Aggregator Fusion for Efficient Graph Convolutions with Pietro Liò
* Improving the Expressive Power of Graph Neural Network with Tinhofer Algorithm
* Sub-GMN: The Subgraph Matching Network Model
* SST-GNN: Simplified Spatio-temporal Traffic forecasting model using Graph Neural Network

Math
* Using Graph Theory to Derive Inequalities for the Bell Numbers

Software
* LAGraph: Linear Algebra, Network Analysis Libraries, and the Study of Graph Algorithms

Survey
* Scene Graphs: A Survey of Generations and Applications

Topological GNNs and graph models for video

Two articles recently were featured on Synced. One is Making GNNs ‘Topology-Aware’ to Advance their Expressive Power about using persistent homology for additional expressivity of GNNs. Another is SOTA GNN ‘Reasons’ Interactions over Time to Boost Video Understanding about modeling images as graphs to reason about their content.

Adaptive Filters and Aggregator Fusion for Efficient Graph Convolutions

A blog post by Shyam A. Tailor about a simple modification of GCN layer that is both more efficient and more effective than many standard message-passing algorithms.

The London Geometry and Machine Learning Summer School 2021

A very cool one week school on geometric deep learning, happening online this summer. Early career researchers such as Ph.D. students will work in small groups under the guidance of experienced mentors on a research project. Applications are open until 31 May 2021.

Bag of Tricks for Semi-Supervised classification

There is a nice short paper on tricks employed on improving performance of GNN. The author, Yangkun Wang, from DGL team has a lot of high scoring entries in the OGB leaderboard, so it's worth employing these tricks: they boost performance a bit but do it consistently. The tricks include:

* data augmentation
* using labels as node features
* renormalization of adjacency matrix
* novel loss functions
* residual connections from the input