Abstract
关系推理是智能行为中的核心,但是实验证明关系推理很难用神经网络来学习。
本文描述了如何使用Relation Networks 作为一种即插即用的Model从根本上解决关系推理问题。
Introduction
之前提到推理实体及实体见属性间的关系是智能行为中很重要的部分。
1514466203466_2.png
- Relational question
- requires explicit reasoning about the relations between the four objects in the image
- Non-relational question
- requires explicit reasoning about the relations between the four objects in the image
文章提出因为深度学习是以稀疏而复杂的关系映射作为基础,所以训练需要海量的数据。(However,a number of these approaches, such as deep learning, often struggle in data-poor problems where the underlying structure is characterized by sparse but complex relations)。
看似简单的推理问题对神经网络(CNNs MLPs)来说是非常困难的。
提出了一种专门做关系推理的Relation Network (RN)来解决神经网络中的关系推理问题。
之前提出的关系推理模型:
- Graph Neural Networks
- Gated Graph Sequence Neural Networks
- Interaction Networks
RNs 更简单,即插即用,专注于灵活的关系推理(exclusively focused on flexible relational reasoning)
通过联合训练,RNs可以影响和修改CNNs的上游表征(upstream representations) 来产生可以做关系推理的隐式对象
Relation Networks
就像CNNs具有空间平移不变性一样,RN天生具有关系推理的能力。
设计思想:
通过约束神经网络的功能形式(constrain the functional form of a neural network)来使RN具有捕捉关系推理的核心共同属性(captures the core common properties of relational reasoning)
在最简单的形式中RN是一个复合函数: