论文阅读：图神经网络

会议：
年份：2018
题目：Relational inductive biases, deep learning, and graph networks
作者：DeepMind, Google Brain, MIT, University of Edinburgh

1. Introduction

A key signature of human intelligence is the ability to make “infinite use of finite means”.
这句话反映了的 combinatorial generalization，用已知的 building block 去创建新的推理，预测和行为。
Humans’ capacity for combinatorial generalization depends critically on our cognitive mechanisms for representing structure and reasoning about relations.
复杂系统的表示方法：实体和联系。
in previous eras, data and computing resouces were expensive, and the improved sample complexity afforded by structured approaches’ strong inductive biases was very valuable.
modern deep learning methods often follow an “end-to-end” design philosophy which emphasizes minimal a priori representational and compuational assumptions, and seek to avoid explicit structure and “hand-engineering”.
前者在采样上很好，后者因算力和数据的廉价而放弃了采样效率。
当代的深度学习方法不能很好处理复杂语言和情景的理解，对结构化数据的推理认知，迁移学习以及小样本学习。

最近已经出现了综合结构化方法和深度学习的工作，他们对于结构化数据，特别是图进行操作，考虑了实体以及它们间的关系，与过去的传统方法相比，它们的区别在于如何学习实体以及关系的表示以及如何进行相关计算。

Inductive biases can express assumptions about either the data-generating process or the space of solutions.
To explore the relational inductive biases expressed within various deep learning methods, we must identify several key ingredients, analogous to those in Box 1: what are the entities, what are the relations, and what are the rules for composing entities and relations, and computing their implications?