Abstract:
Graphs are a powerful data structure that is used to model objects and their interactions. Owing to their ability to capture rich information about interacting entities, they are used to model a wide range of data. For example, in robotics, an articulated robot's bodies can be represented with a graph's nodes while the joints linking the bodies together can be represented with a graph's edges. Recently, neural architectures that are able to process graph input data, called Graph Neural Networks(GNNs), haveemerged with promising results on many supervised learning tasks such as graph classification. Work has also been done to use these GNNs in many control-related applications such as for inference, system identification, model-predictive control and even for visual tasks like human action recognition. In this work, after an extensive review of the state-of-the-art literature on GNNs, including breakthrough applications in control, we combine two reference GNN architectures with a powerful attention mechanism, proposing two novel architectures and validating them on four benchmark graph classification datasets using rigorous methodology and cutting-edge software tools. Our proposed architectures achieve impressive gains in performance of up to 14% over baselines on certain datasets, opening up interesting perspectives for future work, especially in human action recognition and pose estimation problems.
