Reproduce “Statistical Test for Anomaly Detections by Variational Auto-Encoders” on OpenResearch

About this paper

Daiki Miwa, Tomohiro Shiraishi, Vo Nguyen Le Duy, Teruyuki Katsuoka, Ichiro Takeuchi

A new framework, the VAE-AD Test, quantifies the statistical reliability of anomalies detected by Variational Autoencoders by integrating Conditional Selective Inference. This approach produces valid p-values that control the Type I error rate, preventing false positives, while demonstrating higher statistical power compared to other conservative methods, and successfully identifies anomalies in real brain imaging data.

Key insights

• Variational Autoencoders, constructed from piecewise-linear deep learning operations, can be rigorously characterized as piecewise-assignment functions, making them amenable to Conditional Selective Inference.
• Naive statistical testing of data-driven anomaly regions leads to severe selection bias, resulting in significantly inflated Type I error rates (false positives).
• Conditioning the inference on the specific anomaly region selected by the VAE, as done by CSI, is crucial for obtaining statistically valid p-values and controlling the Type I error.

Summary by alphaXiv.