[1J5-GS-2] Machine learning: Fundamental theory (2)

[1J5-GS-2-03] Out-of-distribution detection for high dimensional data using low dimensional representation with deep generative models

〇Hirono Okamoto¹, Masahiro Suzuki¹, Yutaka Matsuo¹ (1. The University of Tokyo)

Keywords:Out-of-distribution detection

Out of distribution (OOD) detection is difficult for high dimensional data like images because density estimation is difficult due to the curse of dimension. Even if we use deep generative model and estimate the probability density, it is not guaranteed that the model estimates the probability density of OOD data is small in latent space because we cannot train the model using OOD data. Therefore, we hypothesize that the probability density of OOD data is small in latent space if the model can reconstruct the original image. We propose to use VQVAE for the model. In experiment, we confirm that VQVAE can reconstruct the original image with little reconstruction error even if we use the OOD data for the input of the model. Furthermore, we confirm that we can detect OOD data on same domain by using the data in the latent space compared with the real space.