总结
对比学习综述讲到这里也就基本结束了
来回顾一下对比学习的发展历程和一些要点吧!
发展历程
论文名称 | 贡献 | 网络类型 | 用的代理任务 | 延续了哪篇工作 |
---|---|---|---|---|
Inst Disc | 奠基作,引入 Memory Bank 大字典的概念,提出个体判别任务。 | 判别式网络 | 个体判别任务 | 无 |
定义正负样本的方式(因为不止一篇论文就不放名字了) | 引入生成式网络的概念,提出多视角作为不同正样本 | 前者是生成,后者是判别 | 时序预测任务 多视角判别任务 | 无 |
MoCo | 使用队列储存字典,提出动量编码器,提出 infoNCEloss | 判别式网络 | 个体判别任务 | Inst Disc |
SimCLR | 提出 projection head,提出把一张图片做不同的数据增强进行对比,并且分析了各种数据增强的贡献 | 判别式网络 | 个体判别任务 | Inst Disc |
SwAV | 使用聚类中心矩阵,提出新的聚类任务 | 聚类式网络(或许也算判别式) | 聚类 | SimCLR |
BYOL | 提出无负样本学习,提出预测头 predictor | 判别式网络 | 或许算预测? 但也有个体判别的感觉 | SimCLR |
SimSiam | 在 BYOL 基础上做了改进,总结了前面的工作,提出了对称式的孪生网络 | 判别式网络 | 同上 | BYOL |
MoCo v3 | 缝合了 MoCo 和 SimSiam,引入了 VIT,提出小 trick 解决训练不稳定 | 判别式网络 | 同上 | MoCo |
要点提问
1. 什么是 memory bank?
2. 个体判别任务的局限
3. 什么是 NCEloss,infoNCEloss 呢?
4. 能画出每个网络的结构图吗?
5. 能列出提到过的所有小 trick 吗?