A conditional GAN is trained on handwritten digits with ten classes. Both networks see the label, exactly as the lesson describes: the generator takes a noise vector together with a class label, and the discriminator takes an image together with a class label.
At every step the discriminator is shown exactly two kinds of example:
| batch | what it contains | what the discriminator is told to answer |
|---|---|---|
| A | a real image, paired with its own true label | real |
| B | a generated image, paired with the label it was generated from | fake |
Training goes well by every visible measure — the samples are sharp and look like real handwriting. But the dial does not work. Ask the generator for a and you get a convincing digit that is about as likely to be a .
The team decides to add one extra kind of batch to the discriminator's diet. Which addition forces the generator to start obeying the label, and why?
Select all that apply.