https://arxiv.org/abs/2002.06224
Top-k Training of GANs: Improving GAN Performance by Throwing Away Bad Samples (Samarth Sinha, Zhengli Zhao, Anirudh Goyal, Colin Raffel, Augustus Odena)
Top-K Training of GANs: Improving Generators by Making Critics Less Critical (Samarth Sinha, Anirudh Goyal, Colin Raffel, Augustus Odena)
GAN 트레이닝을 개선하는 간단한 방법: generator loss를 계산할 때 discriminator loss가 낮은 샘플 (즉 less realistic한)을 날려버리고 계산하기. 이전에도 이런 시도가 있었던 것 같은데.
#gan #regularization