One training step of a GAN begins by showing the discriminator a batch: 8 real photographs and 8 samples fresh from the generator.
For each example the discriminator outputs a single number between and — how strongly it believes that example is real. It turns that number into a verdict the ordinary way: means it says real, means it says fake. No value in this batch sits exactly at .
| # | on the real photographs | on the generated samples |
|---|---|---|
| 1 | ||
| 2 | ||
| 3 | ||
| 4 | ||
| 5 | ||
| 6 | ||
| 7 | ||
| 8 |
Across the whole batch of 16, what percentage of the discriminator's verdicts are correct?
Give the answer as a percentage, rounded to 2 decimal places.