Zero-Shot Visual Recognition using Semantics-Preserving Adversarial Embedding Networks

收获

• 多个网络如何组合? 相同目标collaborative, 不同目标adversarial

todo

• 查看groud truth class embedding
• 查看SAE

Conclusion

pro

1. using adversarial framework to combine two network: Semantic transfer can be achieved
2. an indepentdent visual-to-semantic mapping(tackling the semantic loss problem inherently in classification)

con

Overview

• problem: semantic loss
  
  • some semantics would be discarded during training
  • they are non-discriminative for traning classes
  • critical for recognizing test classes

• how

• 解决了分类网络不重要信息丢失的问题

1. Introduction

1.1 zero shot learning

• transferring knowledge from seen classes to unseen classes
• evolution
  
  • primitive attribute classifier
  • semantic embedding based framework
    

3. Model

- supervised Adeversial Autoencoder
- F encoder, G decoder
- F can be considered as the bottleneck layer, regularized to match supervised E

4. Implementation

4.1 Architecture

• E: resnet 101
• F: Alexnet + 2xfc
• leaky RELU: transform a vector into 3D feature map
• ground truth class embedding

4.2 Training details

• per-pixel mean subtraction
• fixed the Resnet-101 in E, initialized the AlexNet-like blocks in F with AlexNet and G with the pretrained generator
• learning rate started from 1e −4 and is multiplied by 0.1 when the error
  is plateaus.
• grid search to select parameter α and β.