敵対的生成ネットワークに関するクイズ

クイズに答えると、答えと解説が表示されます。

1. GAN 内で競合する 2 つのネットワークとは何ですか?
   • エンコーダとデコーダ
   • 読者と作家
   • 常に一致する 2 つの同一のコピー
   • 偽物を作るジェネレーターとそれを判断するディスクリミネーター
2. In the original unconditional GAN formulation, what input does the generator transform into a sample?
   • A guaranteed copy of a finished real image
   • Sampled latent noise
   • The discriminator's parameter file
   • A fixed human-written answer key
3. What adversarial learning signal can reach a GAN generator through its discriminator?
   • Gradients of the generator objective with respect to its parameters
   • A rule that every generated image is correct
   • A prohibition against all conditional image inputs
   • A copy of every discriminator weight as the output image
4. GAN における「モード崩壊」とは何ですか?
   • ディスクリミネーターはすべてのデータを削除します
   • ネットワークのメモリが不足しています
   • ジェネレーターは、さまざまな出力ではなく、少数の反復出力のみを生成します。
   • 画像の解像度が自動的に 2 倍になります
5. Which researchers introduced the GAN framework in the 2014 paper?
   • Alan Turing alone in 1950
   • イアン・グッドフェローと同僚
   • Geoffrey Hinton alone in 2006
   • Yann LeCun alone in 1998
6. How can a conventional feed-forward GAN generator produce one sample after training?
   • By requiring exactly a thousand discriminator updates at inference
   • By retrieving an unchanged target image in every case
   • Without evaluating any learned computation
   • With one forward evaluation of the generator
7. At the ideal original GAN equilibrium where generated and real distributions match, what probability does the optimal discriminator assign to an example being real?
   • Zero for every example
   • One for every example
   • One-half
   • A probability greater than one
8. In the conditional GAN formulation, where can a class label or other condition be supplied?
   • To both the generator and the discriminator
   • Only to an unrelated output filename
   • To neither network by definition
   • Only after all training and generation have finished