VAE-SIMCA — Data-driven method for building one class classifiers with variational autoencoders

The paper proposes a new method for building one class classifiers based on variational autoencoders (VAE). The classification decision is built on a linear combination of two squared distances: computed for the original and the reconstructed image as well as for the representation of the original image inside the latent space formed by VAE. Because both distances are well approximated by scaled chi-square distribution, the decision boundary is computed using the theoretical quantile function for this distribution and the predefined probability for Type I error, ⍺. Thereby the boundary does not require any specific optimization and is solely based on the model outcomes computed for the training set.
The original idea of the proposed method is inherited from another OCC approach, Data Driven Soft Independent Method for Class Analogies, where singular value decomposition is employed for building the latent space. In this paper we show how this idea can be adopted to be used with VAE for detection of anomalies on images. The paper describes the theoretical background, introduces the main outcomes as well as tools for visual exploration of the classification results, and shows how the method works on several simulated and real datasets.