[CVPR 2017] Semantic Autoencoder for Zero-Shot Learning论文笔记

http://openaccess.thecvf.com/content_cvpr_2017/papers/Kodirov_Semantic_Autoencoder_for_CVPR_2017_paper.pdf

Semantic Autoencoder for Zero-Shot Learning,Elyor Kodirov Tao Xiang Shaogang Gong,Queen Mary University of London, UK,{e.kodirov, t.xiang, s.gong}@qmul.ac.uk

亮点

  • 通过对耦学习提升零次学习系统的性能(类似CycleGan)
  • 结构非常简洁,且可直接求解,速度非常快
  • 有效应用到其他相关任务(监督聚类)上,证明了范化性能

[CVPR 2017] Semantic Autoencoder for Zero-Shot Learning论文笔记

方法

Linear autoencoder

[CVPR 2017] Semantic Autoencoder for Zero-Shot Learning论文笔记

Model Formulation

[CVPR 2017] Semantic Autoencoder for Zero-Shot Learning论文笔记

which is a well-known Sylvester equation which can be solved efficiently by the Bartels-Stewart algorithm (matlab sylvester).

零次学习:基于以上算法有两种测试的方法:

  • 将一个未知的类别特征样本xi通过W映射到语义空间(属性)si,通过比较语义空间的距离找到离它最近的类别(无训练样本),即为它的标签
  • 将所有无训练数据类别的语义特征S通过WT映射到特征空间X,通过比较一个未知类别的样本xi和映射到特征空间的类别中心X的距离,找到离它最近的类别,即为它的标签
  • 以上两种算法得到结果的准确度基本相同。

监督聚类:在这个问题中,语义空间即为类别标签空间(one-hot class label)。所有测试数据被影射到训练类别标签空间,然后使用k-means聚合

与已有模型的关系:零度学习已有模型一般学习一个满足以下条件的影射:

[CVPR 2017] Semantic Autoencoder for Zero-Shot Learning论文笔记

或者,在[54]中将属性影射到特征空间,学习目标变为,

[CVPR 2017] Semantic Autoencoder for Zero-Shot Learning论文笔记

文中的算法结合了这两者,而且由于W*=WT,在对耦学习中W不可能太大(否则,x乘以两个范数很大的的矩阵无法恢复原来的初始值),正则化项可以被忽略。

[CVPR 2017] Semantic Autoencoder for Zero-Shot Learning论文笔记

实验

零次学习

数据集:Semantic word vector representation is used for large-scale datasets (ImNet-1 and ImNet-2). We train a skip-gram text model on a corpus of 4.6M Wikipedia documents to obtain the word2vec2 [38, 37] word vectors.

[CVPR 2017] Semantic Autoencoder for Zero-Shot Learning论文笔记

特征:除 ImNet-1用AlexNet提取外,其他均使用了GoogleNet

结果:

  • Our SAE model achieves the best results on all 6 datasets.
  • On the smallscale datasets, the gap between our model’s results to the strongest competitor ranges from 3.5% to 6.5%.
  • On the large-scale datasets, the gaps are even bigger: On the largest ImNet-2, our model improves over the state-of-the-art SS-Voc [22] by 8.8%.
  • Both the encoder and decoder projection functions in our SAE model (SAE (W) and SAE (WT) respectively) can be used for effective ZSL.
    • The encoder projection function seems to be slightly better overall.
  • Measures how well a zero-shot learning method can trade-off between recognising data from seen classes and that of unseen classes
    • Holding out 20% of the data samples from the seen classes and mixing them with the samples from the unseen classes.
    • On AwA, our model is slightly worse than the SynCstruct [13].
    • However, on the more challenging CUB dataset, our method significantly outperforms the competitors.

聚类

数据集: A synthetic dataset and Oxford Flowers-17 (848 images)

结果:

  • On computational cost, our model (93s) is more expensive than MLCA (39%) but much better than all others (hours~days).
  • Achieves the best clustering accuracy

p.p1 { margin: 0.0px 0.0px 0.0px 0.0px; font: 14.0px "Helvetica Neue"; color: #042eee }
p.p2 { margin: 0.0px 0.0px 0.0px 0.0px; font: 16.0px "Helvetica Neue"; color: #323333 }
p.p3 { margin: 0.0px 0.0px 0.0px 0.0px; font: 14.0px "Helvetica Neue"; color: #323333 }
p.p4 { margin: 0.0px 0.0px 0.0px 0.0px; font: 14.0px "Helvetica Neue"; color: #323333; min-height: 16.0px }
p.p5 { margin: 0.0px 0.0px 0.0px 0.0px; font: 17.0px STIXGeneral; color: #323333 }
p.p6 { margin: 0.0px 0.0px 0.0px 0.0px; font: 12.0px STIXGeneral; color: #323333 }
p.p7 { margin: 0.0px 0.0px 0.0px 0.0px; font: 9.0px STIXGeneral; color: #323333 }
p.p8 { margin: 0.0px 0.0px 0.0px 0.0px; text-align: center; font: 17.0px STIXGeneral; color: #323333 }
p.p9 { margin: 0.0px 0.0px 0.0px 0.0px; text-align: center; font: 17.0px "Helvetica Neue"; color: #323333; min-height: 20.0px }
p.p10 { margin: 0.0px 0.0px 0.0px 0.0px; text-align: center; font: 19.0px STIXSizeOneSym; color: #323333 }
p.p11 { margin: 0.0px 0.0px 0.0px 0.0px; font: 14.0px "Helvetica Neue"; color: #323333; min-height: 17.0px }
li.li3 { margin: 0.0px 0.0px 0.0px 0.0px; font: 14.0px "Helvetica Neue"; color: #323333 }
span.s1 { text-decoration: underline }
span.s2 { }
span.s3 { font: 19.0px STIXSizeOneSym }
ul.ul1 { list-style-type: disc }
ul.ul2 { list-style-type: circle }

上一篇:It is difficult to the point of impossiblity for sb to image a time when ...


下一篇:android 使用NinePatch图作Background,导致布局混乱