Transfer Learning&GAN

Unsupervised Transfer

Generate To Adapt: Aligning Domains using Generative Adversarial Networks

1.Main

using unlabeled target data help transfer(target image&class seen)

2.structure

3. Loss Func

• Source Data:
  D:
  G:
  

C&F:

• Target data
  
  

4.DataSet

• Digit classification (MNIST, SVHN and USPS
  datasets)
• Object recognition using OFFICE datasets
• Domain adaptation from synthetic to real data;CAD synthetic dataset (source) and a subset of PASCAL
  VOC dataset(target)
• VISDA dataset:Trasfer competation

5.metric

classification accurancy

Disentangled Classification and Reconstruction for Zero-shot learning

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

1.Main

prevents the semantic loss while target image&class unseen

2. Structure

3. Loss Func

• Class loss
  
• Reconstruction Loss
  
  3.Adversarial Loss
  

4.Dataset

CUB, AWA, SUN and aPY, SP-AEN

5.Metric

harmonic mean (H) on generalized ZSL
The Seen-Unseen accuracy Curve (SUC)

Conditional GAN on feature space

Adversarial Feature Augmentation for Unsupervised Domain Adaptation

1.Main

Gener new feature vec for augmentation

2.Structure

3.Loss Func

• S1 Train reference feature encoder E_S Classfier C in Source data
• S2 Conditional Gan for gener new feature vec in Source domain,get Encoder S
• S3 Train encoder in T&S advertising with S

4.Datasets

• mnist ,usps :white digit on black background
• svhn:real images of street view house numbers
• syn digits:syn on svhn
• nyud:object RGB->D

5.Metric

• t-SNE
• APs:feature augmentation
• Accuracy:compare with C trained on S,T,Other method