Publications
ワークショップ (国際) Disentangling Clustered Representations of Variational Autoencoders for Generating Diverse Samples
Tsubasa Takahashi, Tatsuya Komatsu, Koki Yamada (Tokyo University of Agriculture and Technology)
Learning Data Representation for Clustering (LDRC at IJCAI 2020)
2021.1.7
For learning representations of vast amounts of unlabeled data, variational autoencoders (VAEs) with Gaussian-mixture prior provide us generative models that have clustered representations. The Gaussian-mixture VAEs (GM-VAEs) give us good reconstructions and clustering, but generated samples from the model lack diversity. This is due to that VAEs prefer to reduce the reconstruction error and result in expanded and loose representations. How can we build a VAE with clustering the representations and having sampling diversity? To solve the above issues, we introduce an additional regularization term that encourages to fit learned representation space to the distributions of original data. Our proposed VAE with the regularization provides us to generate diverse samples from tight and disentangled representation space. In our experiments, the proposed method shows 95% clustering accuracy and outperforms the existing solution in terms of clustering accuracy.
Paper :
Disentangling Clustered Representations of Variational Autoencoders for Generating Diverse Samples
（外部サイト）