Meta Learning for Graph Neural Networks

Deep learning has enabled incredible advances in pattern recognition such as the fields of computer vision and natural language processing. One of the most successful areas of deep learning is Convolutional Neural Networks (CNNs). CNNs have helped improve performance on many difficult video and image understanding tasks but are restricted to dense gridded structures. Further, designing their architectures can be challenging, even for image classification problems. The recently introduced graph CNNs can work on both dense gridded structures as well as generic graphs. Graph CNNs have been performing at par with traditional CNNs on tasks such as point cloud classification and segmentation, protein classification and image classification, while reducing the complexity of the network. Graph CNNs provide an extra challenge in designing architectures due to more complex weight and filter visualization of generic graphs. Designing neural network architectures, yielding optimal performance, is a laborious and rigorous process. Hyperparameter tuning is essential for achieving state of the art results using specific architectures. Using a rich suite of predefined mutations, evolutionary algorithms have had success in delivering a high-quality population from a low-quality starter population. This thesis research formulates the graph CNN architecture design as an evolutionary search problem to generate a high-quality population of graph CNN model architectures for classification tasks on benchmark datasets.