Semi-supervised learning with generative adversarial networks.
Semi-supervised refers to the training process where the model gets trained only on a few labeled images but the data set contains a lot more unlabeled images. This can be useful in situations where you have a humongous data set but only partially labeled.
In regular GAN the discriminator is trained in an unsupervised manner, where it predicts whether the image is real or fake (binary classification). In SGAN, in addition to unsupervised, the discriminator gets trained in a supervised manner on class labels for real images (multiclass classification).
In essence, the unsupervised mode trains the discriminator to learn features and the supervised mode trains on corresponding classes (labels). The GAN
can be trained using only a handful of labeled examples.
In a standard GAN our focus is on training a generator that we want to use to generate fake images. In SGAN, our goal is to train the discriminator to be an excellent classifier using only a few labeled images. We can still use the generator to generate fake images but our focus is on the discriminator.
Why do we want to follow this path is CNNs can easily classify images?
Apparently, this approach achieves better accuracy for limited labeled data compared to CNNs.
([ Ссылка ])
Another useful resource: [ Ссылка ]
