In GANs the two models compete with each other until the generative model comes out on top. Lets say we train a GAN to produce close-to-real face images yet which can be ascertained as fake by the human eye. My question is can we train a Discriminative model that performs well in distinguishing these fake images from the real ones? If yes what modifications are required to the conventional CNN to do so? Thank you for reading.
@siddwho819,
The goal is not for the generator to come out on top. We want both the generator and discriminator to do well. We need to improve the discriminator to challenge the generator to do better. As an extreme case, imagine a discriminator that was so bad that it considered everything to be real. Then, the generator would “trick” it every time, no matter how bad its images were, and it would never learn.
Stay tuned for the rest of course and specialization and you’ll see all kinds of interesting techniques and uses for GANs.
Happy learning!
2 Likes