Table of Contents
Why Signals GAN Success When Discriminator Sees Fakes as 50% Real?
When GAN discriminator hits ~0.5 probability on generated samples, it means generator success—indistinguishable fakes, equilibrium—vs. failure modes like collapse or perfect classification for AI exam prep.
Question
What happens during GAN training when the Discriminator’s probability of a generated sample being real is approximately 0.5?
A. The training process stops, as the GAN has failed.
B. The Discriminator has learned to perfectly classify all fake samples.
C. The Generator is creating samples that are so realistic the Discriminator can no longer distinguish them from real data.
D. The Generator begins to experience mode collapse.
Answer
C. The Generator is creating samples that are so realistic the Discriminator can no longer distinguish them from real data.
Explanation
During GAN training, a discriminator probability of ~0.5 for generated samples signifies the Nash equilibrium where the generator produces outputs indistinguishable from real data, as the discriminator—trained to output 1 for real and 0 for fake—achieves maximum uncertainty (random guessing at 50% accuracy), fulfilling the minimax objective where D(x) ≈ 0.5 for both real and generated inputs, indicating successful convergence rather than issues like option A (training continues or restarts), option B (perfect classification would yield ~0 for fakes), or option D (mode collapse shows discriminator accuracy near 100% or 0% as generator ignores data diversity).