1 minute read

训练数据不包含标注信息的情况下完成目标任务;
表示学习资源

unsupervised learning

1 综述

  1. A survey on Semi-, Self- and Unsupervised Techniques in Image Classification
    2020-02-20

2 理论

3 生成式

  1. A fast learning algorithm for deep belief nets
    2006 paper

  2. Unsupervised representation learning by predicting image rotations
    ICLR 2018 2018-03-21 paper | pytorch-official | openreview
    通过图像旋转进行学习;

  3. AET vs. AED: Unsupervised Representation Learning by Auto-Encoding Transformations rather than Data
    CVPR 2019 2019-01-14 paper
    无监督通过数据还是自编码网络进行学习的讨论;

4 判别式

4.1 对比学习

  1. Dimensionality reduction by learning an invariant mapping
    CVPR 2006 2006 paper | blog
    $\bullet \bullet$
    首次提出对比学习;

  2. Discriminative unsupervised feature learning with convolutional neural networks
    NIPS 2014 2014-06-26 paper | caffe | blog
    Exemplar-CNN
    首次使用正负样本对抗;

  3. Data-Efficient Image Recognition with Contrastive Predictive Coding
    2019-05-22 paper | openreview | blog | blog
    使用了特殊的网络结构,用了 layer norm 和 数据增广;当时最好的方法;

  4. Learning representations by maximizing mutual information across views
    2019-06-03 paper | pytorch-official | blog
    $\bullet \bullet$ - AMDIM
    判别式对比学习,使用了特殊的网络结构,并且改动了数据增广;

  5. Momentum contrast for unsupervised visual representation learning
    2019-11-13 paper | blog | blog | blog | pytorch
    $\bullet \bullet$ - MoCo
    对数据做了混洗,用了 ResNet × n,耗费存储;
  6. A Simple Framework for Contrastive Learning of Visual Representations
    AAAI 2020 2020-02-13 paper | blog
    $\bullet \bullet$ SimCLR
    定位:探索基于对比学习的判别式半监督网络,包括数据增广,特征提取和相似度计算;
    总结:文章收录了较为全面的半监督学习相关信息,并对各环节做了实验,最终给除了效果最好的组合——SimCLR,达到 SOTA;
    核心:各环节做了丰富的对比实验,对工程有很大指导作用;
    • 各环节只给出实验结果,缺乏理论分析;
    • 只给了实验结果,未给出实验细节;
    • 文章的核心“对比学习”就是孪生网络,相关任务有图片检索(人脸验证,REID 等);不知道为什么非要换个名字;
    • 对比学习用于无监督分类准确度测试的时候,也没给出测试细节——怎么实现的半监督学习;
    • 结论说 SimCLR(基于 4×ResNet)在微调后达到监督学习(基于 ResNet)的水平是噱头,因为 backbone 不一样;

    看起来像是很努力但不聪明又的人写的文章;

4.2 其他

  1. Unsupervised Visual Representation Learning by Context Prediction
    ICCV 2015 2015-05-19 paper | caffe

  2. Unsupervised Feature Learning via Non-Parametric Instance-level Discrimination
    CVPR 2018 2018-05-05 paper | pytorch-official | pytorch | blog | blog
    $\bullet \bullet$ InsDis

5 应用

5.1 视觉异常检测

异常检测

  1. Manifolds for Unsupervised Visual Anomaly Detection
    2020-06-19 paper | [crack 分割数据集](https://github.com/khanhha/crack_segmentation)
    无监督 VAE 从医疗图像中识别出癌症图片;

6 其他

  1. Deep Clustering for Unsupervised Learning of Visual Features
    ECCV 28 28-07-15 paper | 解读

  2. Self-Supervised Representation Learning by Rotation Feature Decoupling
    CVPR 2019 2019 paper


TOP

附录

A 数据集

B 资源

  1. paper with code

Comments