research-article

Co-Exploration of Graph Neural Network and Network-on-Chip Design Using AutoML

Authors:

Daniel Manu,

Shaoyi Huang,

Caiwen Ding,

Lei YangAuthors Info & Claims

GLSVLSI '21: Proceedings of the 2021 Great Lakes Symposium on VLSI

Pages 175 - 180

https://doi.org/10.1145/3453688.3461741

Published: 22 June 2021 Publication History

Get Access

Abstract

Recently, Graph Neural Networks (GNNs) have exhibited high efficiency in several graph-based machine learning tasks. Compared with the neural networks for computer vision or speech tasks (e.g., Convolutional Neural Networks), GNNs have much higher requirements on communication due to the complicated graph structures; however, when applying GNNs for real-world applications, say in recommender systems (e.g. Uber Eats), it commonly has the real-time requirements. To deal with the tradeoff between the complicated architecture and the high-demand timing performance, both GNN architecture and hardware accelerator need to be optimized. Network-on-Chip (NoC), derived for efficiently managing the high-volume of communications, naturally becomes one of the top candidates to accelerate GNNs. However, there is a missing link between the optimize of GNN architecture and the NoC design.

In this work, we present an AutoML-based framework GN-NAS, aiming at searching for the optimum GNN architecture, which can be suitable for the NoC accelerator. We devise a robust reinforcement learning based controller to validate the retained best GNN architectures, coupled with a parameter sharing approach, namely ParamShare, to improve search efficiency. Experimental results on four graph-based benchmark datasets, Cora, Citeseer, Pubmed and Protein-Protein Interaction show that the GNN architectures obtained by our framework outperform that of the state-of-the-art and baseline models, whilst reducing model size which makes them easy to deploy onto the NoC platform.

Supplemental Material

MP4 File

This work entails the design of an Auto-ML based framework called GN-NAS to search for the optimal GNN architectures within the parameter search space such that they are also implementable on the NoC platform to analyze HW specifications. The designed GN-NAS framework consists of 3 parts: Parameter search space, GNAS evaluator and NoC platform. The designed GN-NAS controller has 3 phases: GNN architecture exploration, RL training and Parameter modification. We have also proposed a parameter sharing strategy called ?ParamShare? which ensures that already trained weights in the parent architecture are transferred to the child architecture. With the datasets, we consider the citation graph datasets (Cora, Citeseer and Pubmed) and PPI which we evaluate under transductive and inductive settings respectively. Based on our experimental results, GN-NAS models achieves state-of-the-art performance for both transductive and inductive learning. We also observe that GNN models with ParamShare have smaller model sizes as compared to models without ParamShare. This shows that our designed GN-NAS framework can achieve a better trade-of between performance and model sizes.

Download
487.35 MB

References

[1]

Sergi Abadal, Akshay Jain, Robert Guirado, Jorge Lopez-Alonso, and Eduard Alarcon. 2020. Computing Graph Neural Networks: A Survey from Algorithms to Accelerators. 1--12. arXiv:2010.00130 [cs.LG]

Abstract

Supplemental Material

References

Cited By

Index Terms

Recommendations

ALGNN: Auto-Designed Lightweight Graph Neural Network

Edge Based Graph Neural Network to Recognize Semigraph Representation of English Alphabets

AutoGraph: Automated Graph Neural Network

Comments

Information

Published In

Sponsors

Publisher

Publication History

Permissions

Check for updates

Author Tags

Qualifiers

Data Availability

Conference

Acceptance Rates

Upcoming Conference

Contributors

Other Metrics

Bibliometrics

Article Metrics

Other Metrics

Citations

Cited By

Login options

Full Access

View options

PDF

eReader

Share

Share this Publication link

Share on social media

Affiliations