Anomaly Detection by Deep Direct Density Ratio Estimation

Masahiro Abe, Masashi Sugiyama

Keywords: anomaly detection, density ratio estimation, neural networks

Abstract: Estimating the ratio of two probability densities without estimating each density separately has been shown to provide useful solutions to various machine learning tasks such as domain adaptation, anomaly detection, feature extraction, and conditional density estimation. However, density ratio estimation in the context of deep learning has not been extensively explored yet. In this paper, we apply a Bregman-divergence minimization method for density ratio estimation to deep neural networks and investigate its properties and practical performance in image anomaly detection. Our numerical experiments on the CIFAR-10, CIFAR-100 and Fashion-MNIST datasets demonstrate that deep direct density ratio estimation greatly improves the anomaly detection ability and reduces the computation time over state-of-the-art methods.