Detecting out-of-distribution samples via variational auto-encoder with reliable uncertainty estimation

Area of Science:

• Deep Learning
• Bayesian Methods
• Generative Models

Background:

• Variational Autoencoders (VAEs) are powerful deep learning generative models.
• VAEs exhibit a critical weakness in assigning higher likelihoods to out-of-distribution (OOD) inputs than in-distribution (ID) inputs.
• Reliable uncertainty estimation is essential for understanding and mitigating OOD data issues in VAEs.

Purpose of the Study:

• To address the out-of-distribution (OOD) problem in Variational Autoencoders (VAEs).
• To propose a novel method for enhancing uncertainty estimation in VAEs.
• To improve the robustness of VAEs in anomaly detection tasks.

Main Methods:

• Integration of an improved noise contrastive prior (INCP) into the VAE encoder, creating the INCPVAE model.
• The proposed INCP is scalable, trainable, and compatible with existing VAE architectures.
• Leveraging the merits of INCP for accurate uncertainty estimation.

Main Results:

• The INCPVAE model demonstrates superior uncertainty estimation capabilities for OOD data compared to standard VAEs.
• Experiments show robust performance in anomaly detection tasks.
• The model successfully addresses the issue of VAEs assigning higher likelihoods to OOD inputs.

Conclusions:

• The INCPVAE offers a viable solution to the OOD problem in VAEs.
• This approach provides reliable uncertainty estimation for OOD inputs.
• The INCPVAE enhances the applicability of VAEs in real-world scenarios requiring OOD detection.