Paper Link: Unsupervised Domain Adaptation by Backpropagation (mlr.press)

Overall Structure

1. Feature Extractor: Maps the source or target images to a feature vector
   • Minimizes the source classification error
   • Maximizes the loss of the domain discriminator (by making the two feature distributions as similar as possible so that the domain discriminator cannot tell whether the input is a source feature vector or a target feature vector)
2. Label Predictor: Maps the source feature vector to a vector of class probabilities
   • Minimizes the source classification error
3. Domain Discriminator: Classifies whether the input feature vector is a source or target feature vector
   • Minimizes the loss of the domain discriminator
4. Gradient Reversal Layer: Ensures that the feature distributions over the two domains are made similar (as indistinguishable as possible for the domain discriminator), thus resulting in the domain-invariant features

Overall Optimization Problem

• After gradient reverse layer, the optimization problem becomes:

Key Results

1. Minimizing the label prediction loss contributes to learning discriminative features
2. Maximizing the domain classification loss contributes to learning domain-invariant features
3. Training the domain discriminator is closely related to the estimation of HΔH divergence

Zenvi

One Just Vegetable Chick

Comments

  Write a comment ...