关系推理论文中总提到的inductivebias该如何理解

周志华《机器学习》对inductive bias定义如下&＃xff1a;

机器学习算法在学习过程中对某种假设&＃xff08;hypothesis&＃xff09;的偏好&＃xff0c;称为“归纳偏好”&＃xff08;inductive bias&＃xff09;&＃xff0c;或简称为“偏好”

书中举了例子&＃xff0c;例如对一组数据进行拟合的曲线有无数种&＃xff0c;其中有的比较“简单”&＃xff08;假设我们认为曲线更平滑意味着“更简单”&＃xff09;&＃xff0c;有的更复杂。例如一组可以用二次曲线来拟合的数据点&＃xff0c;用更复杂的更高阶的曲线也可以拟合&＃xff0c;那我们的模型应该选择哪条曲线/假设呢&＃xff1f;这就是模型对假设的偏好问题。

在知乎一篇文章中又看到一个解释&＃xff0c;也可以参考一下&＃xff1a;

所谓的inductive bias&＃xff0c;指的是人类对世界的先验知识&＃xff0c;对应在网络中就是网络结构。

再举个论文中的例子&＃xff0c;在《Recurrent Relational Networks》 一文中&＃xff0c;网络结构如下图所示&＃xff08;具体训练过程见原文解释&＃xff09;&＃xff1a;

论文将该模型的inductive bias总结为&＃xff1a;

This paper introduces a composite function, the recurrent relational network. It serves as a modular component for many-step relational reasoning in end-to-end differentiable learning systems. It encodes the inductive biases that

1. objects exists in the world
2. they can be sufficiently described by properties
3. properties can change over time
4. objects can affect each other and
5. given the properties, the effects object have on each other is invariant to time.