Few-shot Visual Reasoning with Meta-analogical Contrastive Learning

[Ch4osMy7h]

这篇论文中，作者将abstract visual reasoning的任务规范为类比推理的范式，并且构建了一个基于元学习和对比学习结合的框架来解决abstract visual reasoning在few-shot下性能表现差的问题。

信息

• 主要作者：Youngsung Kim, Sung Ju Hwang
• 单位：Samsung Advanced Institute of Technology, Samsung Electronics
• 论文链接

1 学习到的新东西：

1. 可以将meta- learning融入到对比学习中，这样既能学习到有用的特征，同时也能增加模型的泛化性。

2. 作者通过对推理数据中的任务图片（给你7张有一定规律的图片，推断出第8张图片应该是什么）进行两种替换。第一种是对7张图片中的一种加入扰动改变它内容，另一种是改变图片中图形的形状。通过这两种方式来构造类似的正例来辅助对比学习。

2 通过Related Work了解到了哪些知识

1. 视觉推理的概念。
2. 类别集和类别推理的概念。

3 实验验证任务，如果不太熟悉，需要简单描述

在之前的模型上应用它们的框架，性能在few-shot下取得了非常大的提升（数十个点）。

4 在你认知范围内，哪些其它任务可以尝试

可以在跨模态相关的NLP任务里面应用这种元学习和对比学习结合的思想。

5 好的句子

1. In this work, we propose to solve such a few-shot (or low-shot) abstract visual reasoning problem by resorting to analogical reasoning, which is a unique human ability to identify structural or relational similarity between two sets.
2. This analogical contrastive learning allows to effectively learn the relational representations of given abstract reasoning tasks.
3. The visual reasoning task proposed in recent works [1, 2] often involves visual puzzles such as Raven Progressive Matrices (RPM), whose goal is to find an implicit rule among the given image panels, and predict the correct image panel that will complete the puzzle (see Figure 1b).
4. Analogical reasoning, which is a way to reason about a new problem by drawing an analogy between it and an existing problem, has been considered as a unique trait of human intelligence and a key component in building an AI system with general intelligence.
5. Perhaps the most relevant work to ours is Hill et al. [5], which proposes to perform analogical reasoning task on a visual reasoning task, but they focus on transferring a learned relation from one problem domain to another, while we focus on learning relational representations by explicitly enforcing relational similarities between pairs with known analogical relations.