KCC 2021
Current Result Document :
| 한글제목(Korean Title) |
대조 손실이 VAE의 잠재 변수에 미치는 영향 |
| 영문제목(English Title) |
Effect of Contrastive Loss on Latent Space of Variational Autoencoder |
| 저자(Author) |
이현섭
최희열
Hyunsub Lee
Heeyoul Choi
|
| 원문수록처(Citation) |
VOL 48 NO. 01 PP. 0633 ~ 0635 (2021. 06) |
한글내용
(Korean Abstract) |
|
영문내용
(English Abstract) |
Variational autoencoder (VAE) is a generative model that produces samples from latent space. VAE is known to be easy to be trained and can learn data representation compared to other generative models such as generative adversarial networks (GAN) and autoregressive models. However, when VAE trained on MNIST generates samples from latent space, we sometimes get ambiguous images that seem to lie between different classes. We think that it is because the distance between the clusters on the latent space is too close to draw samples from a specific class. In this paper, we break through this problem using contrastive learning. We can train separable clusters on latent space each corresponding to one class and find that separated clusters result in high performance in classification tasks. Moreover, we can capture that contrastive loss helps VAE learn more disentangled representation. |
| 키워드(Keyword) |
|
| 파일첨부 |
PDF 다운로드
|
