izhx / paper-reading

组内追更订会相关论文
7 stars 2 forks source link

Iterated learning for emergent systematicity in VQA #2

Open Ch4osMy7h opened 3 years ago

Ch4osMy7h commented 3 years ago

内容

本文基于VQA中的系统性泛化问题,探究如何使用迭代式学习来解决神经模块网络的结构自动生成。

信息

1 学习到的新东西:

这篇论文是一篇很硬核的任务,介绍了目前探究神经网络可解释性最前沿的一些东西。 系统性泛化任务:系统性地生成具有特定含义的神经网络(也被称为神经模块网络),从而让模型能够泛化到在训练集未出现的测试集上。但是目前系统性生成泛化任务大都基于很强的先验或者人工设计。相关工作发现如果采用自动生成神经模块网络的方法,会导致性能出现非常大的下降。 语言生成和组合性:来探究多个agents在环境中如何学会互相交流来解决特定任务。比如一个在论文中被称为multi-agent referential games,包含两种agent——speaker和listener,speaker说自己在图像中看到了什么,listener根据它的描述找到其在图像中的具体位置。

2 通过Related Work了解到了哪些知识

迭代式学习在哪些工作中作出了探究,以后想了解的话可以阅读论文中列出的文献。

3 实验验证任务,如果不太熟悉,需要简单描述

实验验证在作者自己提出的一个数验证框架,以及基本的VQA验证方法上进行了性能测试。

4 在你认知范围内,哪些其它任务可以尝试

其他图像和文本结合的任务。

5 好的词语、句子或段落

  1. We propose to address this problem using iterated learning, a cognitive science theory of the emergence of compositional languages in nature that has primarily been applied to simple referential games in machine learning.
  2. This poses a substantial challenge for the training of NMNs.
  3. IL explains how language evolves to become more compositional and easier-to-acquire in a repeated transmission process
  4. Our efforts highlight the potential of IL for broader machine learning applications beyond the previously-explored scope of language emergence and preservation (Lu et al., 2020).