While reading the InfoGAN paper and implement it taking help from a previous implementation, I'm having some difficulty understanding how it learns the discrete categorical code when trained on MNIST.

The implementation we tried to follow uses a target as a randomly generated integer from 0 to 9. My doubt is this: how can it learn categorical information if from the start it's learning using a loss which takes in random values.

If this implementation is wrong, then what should the target be while training the Q network, when using the categorical-cross-entropy loss on the output logits?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.