# Relational inductive biases, deep learning, and graph networks

**Authors:** Peter W. Battaglia, Jessica B. Hamrick, Victor Bapst, Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner, Caglar Gulcehre, Francis Song, Andrew Ballard, Justin Gilmer, George Dahl, Ashish Vaswani, Kelsey Allen, Charles Nash, Victoria Langston, Chris Dyer  
**arXiv:** [1806.01261](https://arxiv.org/abs/1806.01261)  
**Date:** 2018-06-04  
**Summary page:** https://www.summarizepaper.com/en/paper/1806.01261v3/

## Key points

- Significant progress in various domains of artificial intelligence (AI) such as vision, language processing, control systems, and decision-making
- Current AI approaches struggle to replicate key aspects of human intelligence, particularly the ability to generalize beyond one's experiences
- Authors advocate for prioritizing combinatorial generalization and leveraging structured representations and computations in AI
- Proposal to incorporate relational inductive biases into deep learning architectures through graph networks to enable machines to learn about entities, relations between them, and rules for composing these elements
- Graph networks extend neural network approaches operating on graphs and provide a simple interface for manipulating structured knowledge and generating structured behaviors
- The incorporation of structured representations into AI models can bridge the gap between current machine capabilities and human-like intelligence, enabling more sophisticated patterns of reasoning that are interpretable and flexible

## Summary

Artificial Intelligence: Advancements and Challenges



The field of artificial intelligence (AI) has seen significant progress in various domains such as vision, language processing, control systems, and decision-making. This is largely due to the availability of cheap data and computing resources that align well with the strengths of deep learning. However, despite these achievements, current AI approaches still struggle to replicate many key aspects of human intelligence that have evolved under different pressures. One major challenge is the ability to generalize beyond one's experiences - a skill that is fundamental to human intelligence from infancy. In their paper titled "Relational inductive biases, deep learning, and graph networks," authors Peter W. Battaglia and colleagues argue that achieving human-like abilities in AI requires prioritizing combinatorial generalization and leveraging structured representations and computations. They advocate for an approach that combines the strengths of "hand-engineering" and "end-to-end" learning rather than viewing them as mutually exclusive options. The authors propose incorporating relational inductive biases into deep learning architectures to enable machines to learn about entities, relations between them, and rules for composing these elements. To support this idea further,
they introduce a new building block for AI systems called graph networks which have a strong relational inductive bias. Graph networks extend existing neural network approaches operating on graphs and provide a simple interface for manipulating structured knowledge and generating structured behaviors. By utilizing graph networks,
the authors lay the groundwork for more sophisticated patterns of reasoning that are interpretable and flexible within AI systems. This not only bridges the gap between current machine capabilities and human-like intelligence but also allows for more practical applications. To facilitate further research in this area,
the authors have released an open-source software library for building graph networks along with practical demonstrations on how to implement them effectively. This will encourage other researchers to explore the potential of incorporating structured representations into AI models. Overall,
the work of Battaglia and colleagues highlights the importance of incorporating structured representations into AI models to achieve more sophisticated patterns of reasoning that are interpretable, flexible, and bridge the gap between current machine capabilities and human-like intelligence.