GAN → can be used for domain adaptation → we can learn target features → here we assume that we have the data for the target space → feature augmentation → we are creating a generator that generates features.
GAN → min-max game → creating realistic images → and using this idea → we can generate target domain images → for domain adaptation. (where there can be domain shift → we want a classifier that can act robustly from other domains as well). (the authors extend the general GAN approach → for data augmentation in the feature space). (feature augmentation is a new concept for me).
The related works in this section → are GAN as well as domain adaptation methods → more or less unsupervised learning. (the general approach → assumes that we have the other domain data → which is a reasonable assumption).
Wow, this is the whole architecture seems very complicated → first training a classifier → next training a generator → many of these methods are combined together. (training step involved → more than one step → for the whole model to be trained → wow, quite a lot of optimization involved). (feature generation → only need S model → to create more and more features → this is the method of generating new features). (quite a smart way of generating features from target domains). (a different set of images were used → MNIST to SVHN).
Would this method be good for generalizing into new domains?
Learned feature space can be seen above. (most of the network is composed of Convolution layer with batch norm).
And this type of method → seems to be a good method for domain adaptation. (seems promising).
In general, the author's method → gives a promising result. (comparing with other methods → the author's method → forces domain adaptation → coupled with data augmentation the author's method gives the best results. (in few data sets)).
Domain adaptation was achieved by a smart way of using GAN. (higher accuracy on target data → obtained feature extractor can be used for a different task as well).