Keizo Kato, Jing Zhou, Tomotake Sasaki, Akira Nakagawa
To analyze high-dimensional and complex data in the real world, generative model approach of machine learning aims to reduce the dimension and acquire a probabilistic model of the data. For this purpose, deep-autoencoder based generative models such as variational autoencoder (VAE) have been proposed. However, in previous works, the scale of metrics between the real and the reduced-dimensional space (latent space) is not well-controlled. Therefore, the quantitative impact of the latent variable on real data is unclear. In the end, the probability distribution function (PDF) in the real space cannot be estimated from that of the latent space accurately. To overcome this problem, we propose Rate-Distortion Optimization guided autoencoder. We show our method has the following properties theoretically and experimentally: (i) the columns of Jacobian matrix between two spaces is constantly-scaled orthonormal system and data can be embedded in a Euclidean space isometrically; (ii) the PDF of the latent space is proportional to that of the real space. Furthermore, to verify the usefulness in the practical application, we evaluate its performance in unsupervised anomaly detection and it outperforms current state-of-the-art methods.