Seonglae ChoVariational Auto-Encoder with maximizing ELBO
Regularization to Standard Normal Distribution by Reparameterization trick
입력 데이터의 잠재 변수를 학습하여 새로운 데이터를 생성하는 데 사용 estimates implicitly by Variational Inference
보통 Reconstruction Loss와 KL Loss의 비율이 10:1 ~ 100:1 범위
The difference between AE and VAE is AE store the value of z in the latent space and VAE stores density (mean, variation) to generate parameter
That makes we can generate very plausible results.
Encoder can be useful for Semi-supervised Learning
Limitation is that VAE generate blurry and low-quality generations compared to state-of-the-art like GAN
VAE Notion
VAE Variations
