Interpreting Rate-Distortion of Variational Autoencoder and Using Model Uncertainty for Anomaly Detection

Building a scalable machine learning system for unsupervised anomaly\ndetection via representation learning is highly desirable. One of the prevalent\nmethods is using a reconstruction error from variational autoencoder (VAE) via\nmaximizing the evidence lower bound. We revisit VAE from the perspective of\ninformation theory to provide some theoretical foundations on using the\nreconstruction error, and finally arrive at a simpler and more effective model\nfor anomaly detection. In addition, to enhance the effectiveness of detecting\nanomalies, we incorporate a practical model uncertainty measure into the\nmetric. We show empirically the competitive performance of our approach on\nbenchmark datasets.\n