Chapter 9

1. No; the generator and discriminator are functionally different.
2. No, it can't, because the output of a discriminator must be a probability (that is, p_i ∈ (0, 1)).
3. Yes; it's correct. The discriminator can learn to output different probabilities very quickly, and the gradients of its loss function can become close to 0, reducing the magnitude of the correction feedback provided to the generator.
4. Yes; it's normally quite slower.
5. The critic is slower, because the variables are clipped after every update.
6. As the supports are disjointed, the Jensen-Shannon divergence is equal to log(2).
7. The goal is to develop highly selective units whose responses are only elicited by a specific feature set.
8. It's impossible to know the final organization during the early stages of the training process; therefore, it's not a good practice to force the premature...