[CVPR 2017] Semantic Autoencoder for Zero-Shot Learning論文筆記

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

亮點

• 經過對耦學習提高零次學習系統的性能（相似CycleGan）
• 結構很是簡潔，且可直接求解，速度很是快
• 有效應用到其餘相關任務（監督聚類）上，證實了範化性能

方法

 

Linear autoencoder

 

Model Formulation

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

零次學習：基於以上算法有兩種測試的方法：

• 將一個未知的類別特徵樣本xi經過W映射到語義空間（屬性）si，經過比較語義空間的距離找到離它最近的類別（無訓練樣本），即爲它的標籤
• 將全部無訓練數據類別的語義特徵S經過WT映射到特徵空間X，經過比較一個未知類別的樣本xi和映射到特徵空間的類別中心X的距離，找到離它最近的類別，即爲它的標籤
• 以上兩種算法獲得結果的準確度基本相同。

監督聚類：在這個問題中，語義空間即爲類別標籤空間（one-hot class label）。全部測試數據被影射到訓練類別標籤空間，而後使用k-means聚合

與已有模型的關係：零度學習已有模型通常學習一個知足如下條件的影射：

 

或者，在［54］中將屬性影射到特徵空間，學習目標變爲，

 

文中的算法結合了這二者，並且因爲W*=WT，在對耦學習中W不可能太大（不然，x乘以兩個範數很大的的矩陣沒法恢復原來的初始值），正則化項能夠被忽略。

 

 

實驗

零次學習

數據集：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.

 

特徵：除 ImNet-1用AlexNet提取外，其餘均使用了GoogleNet

結果：

• 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.

聚類

數據集： A synthetic dataset and Oxford Flowers-17 (848 images)

結果：

• 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