Paper

Introduction

Making High Quality images using 'Discrete Latent representation

Discrete Latent Variables

• Defining Latent Embedding space e(K*D)
• Decoder Input

Learning

• Send same Gradient from the decoder to the encoder
• Embedding space learning - Vector Quantisation
• Loss

                   sg: stop gradient 

• Encoder - 1,3/ Decoder - 1/ Embedding space - 2
• No KL-divergence Loss

Learn prior-Image Sampling(pixelCNN)

• PixelCNN
  • [ ] AutoRegressive Model : 자기 자신을 입력으로 하여 자기 자신을 예측