[4Rin1] Interactive 2

[4Rin1-86] Unsupervised anomaly detection using Exact Rate Distorion Auto Encoder

〇Hitoshi Nakanishi¹, Masahiro Suzuki¹, Yutaka Matsuo¹ (1.The university of Tokyo)

Keywords:anomaly detection, generative model, deep learning

The anomaly detection is widely applied in the medial diagnosis and defect detection in the industry. With the variation and scarce of the anomaly defect dataset, the unsupervised model should be essential in this field.
Recent deep model provides the anomaly score by the reconstructing error under model trained only normal dataset because the model couldn't reconstruct unseen anomaly area. The deep generative model additionally predicts the upper bound of log-likelihood (ELBO) whereas the detection was deteriorated against the auto-encoder. It was also reported that the deep generative model fails to predict higher likelihood on unseen dataset than the trained dataset. The author showed this phenomenon on rate-distortion plane. The proposed technique on unsupervised anomaly detection leverages the generative model with exact mutual information restriction, not VAE with the upper bound of mutual information. This technique outperformed the auto-encoder in different anomaly detection datasets.