# Graph Structure of Neural Networks

**Authors:** Jiaxuan You, Jure Leskovec, Kaiming He, Saining Xie  
**arXiv:** [2007.06559](https://arxiv.org/abs/2007.06559)  
**Date:** 2020-07-13  
**Summary page:** https://www.summarizepaper.com/en/paper/2007.06559v2/

## Key points

- Authors: Jiaxuan You, Jure Leskovec, Kaiming He, Saining Xie
- Paper title: "Graph Structure of Neural Networks"
- Introduces relational graph as a novel representation
- Identifies a "sweet spot" in relational graphs for improved performance
- Network performance linked to clustering coefficient and average path length
- Efficient identification methods for optimal graph structures
- Top-performing networks resemble real biological networks
- Opens new avenues for designing effective architectures
- Deepens understanding of neural networks as a whole
- Presented at ICML 2020 with open-source code available

## Summary

In their paper titled "Graph Structure of Neural Networks," authors Jiaxuan You, Jure Leskovec, Kaiming He, and Saining Xie delve into the intricate relationship between the graph structure of neural networks and their predictive performance. They introduce a novel graph-based representation known as a relational graph to address the lack of comprehensive understanding regarding how this structure influences network effectiveness in making predictions. Through their systematic investigation, they make several key findings including a "sweet spot" within relational graphs for improved performance. They also observe that a network's performance is closely tied to its clustering coefficient and average path length. Efficient identification methods can pinpoint optimal relational graph structures for enhanced performance. Surprisingly, top-performing networks exhibit similar structures to those found in real biological networks. This research opens new avenues for designing effective architectures and deepens our understanding of neural networks as a whole. Presented at ICML 2020 with open-source code available for further exploration and implementation by researchers and practitioners in the field.