Efficient Representation Learning With Graph Neural Networks

Abstract

Graph neural networks (GNNs) have emerged as the dominant paradigm for graph representation learning, igniting widespread interest in utilizing sophisticated GNNs for diverse computer vision tasks in various domains, including visual SLAM, 3D object recognition and segmentation, as well as visual perception with event cameras. However, the applications of these GNNs often rely on cumbersome GNN architectures for favorable performance, posing challenges for real-time interaction, particularly in edge computing scenarios. This is particularly relevant in cases such as autonomous driving, where timely responses are crucial for handling complex traffic conditions. The objective of this thesis is to contribute to the advancement of learning efficient representations using lightweight GNNs, enabling their effective deployment in resource-constrained environments. To achieve this goal, the thesis explores various efficient learning schemes, focusing on four key aspects: the data side, the model side, the data-model side, and the application side. In terms of data-driven efficient learning, the thesis proposes an adaptive data modification scheme that allows a pre-trained model to be repurposed for multiple designated downstream tasks in a resource-efficient manner, without the need for re-training or fine-tuning. For model-centric efficiency, the thesis introduces a multi-talented and lightweight architecture, without accessing human annotations, that can integrate the expertise of the pre-trained complex GNNs specializing in different tasks. Furthermore, the thesis explores a dedicated binarization scheme on the data-model side that converts both input data and model parameters into 1-bit representations, resulting in lightweight 1-bit architectures. Finally, the thesis investigates an application-specific efficient learning scheme that models the style transfer process as message passing in GNNs, enabling efficient semi-parametric stylization.