Hardware acceleration of Bayesian graph neural networks
File(s)
Author(s)
Wang, Ziwei
Type
Thesis
Abstract
Bayesian graph convolutional networks (Bayes-GCNs) have shown potential to quantify calibrated uncertainty along with their prediction for various graph-based learning applications. However, the inference of Monte Carlo (MC) Dropout-based Bayes-GCNs requires repeated MC sampling to perform uncertainty prediction, putting a heavy burden on hardware performance. To address this performance challenge, this work proposes an algorithm and hardware co-optimization approach to accelerate Bayes-GCNs. The first contribution of this thesis is to exploit three categories of sparsity to optimize the performance of Bayes-GCNs, including precision sparsity, layer sparsity, and sample sparsity. To automate this optimization, we propose a systematic framework to explore structured sparsity for given Bayes-GCNs while maintaining the same level of accuracy. The second contribution of this thesis is a reconfigurable FPGA-based hardware accelerator tailored for Bayes-GCNs. Several hardware optimization techniques are introduced to improve performance, including bit fusion to optimize the operation of single-bit data, compressed adjacency matrix storage to manage sparsity effectively, an optimized resource allocation strategy to improve resource utilization, and an optimized Bernoulli sampling strategy to improve data reuse. The third contribution of this thesis is the comprehensive evaluation of our approach based on widely-used datasets such as Cora, PubMed, and Flickr. The results demonstrate that our design can achieve up to 398 times faster performance than CPU implementations and up to 4.2 times faster than GPU implementations. Additionally, our design shows improvements in energy efficiency and latency compared to existing GCN accelerators.
Version
Open Access
Date Issued
2024-08-01
Date Awarded
01/04/2025
License URL
Advisor
Luk, Wayne
Publisher Department
Department of Computing
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Master of Philosophy (MPhil)