New method mitigates reconstruction-detection trade-off in VAE anomaly detection
A recent study highlights a compromise in β-VAE models used for unsupervised anomaly detection, balancing reconstruction quality against anomaly detection effectiveness. While models featuring restricted latent spaces yield better detection metrics, they tend to sacrifice reconstruction quality. The variability in performance across different random seeds is associated with the gap between the latent distributions of normal and abnormal data. To address this trade-off, two strategies are introduced: beta-scheduling and the Sparse VAE, with the latter demonstrating enhanced detection capabilities while preserving superior reconstruction quality.
Key facts
- Variational autoencoders are widely used for unsupervised anomaly detection.
- Hyperparameters are often chosen to minimize reconstruction error on normal samples.
- A trade-off exists between reconstruction quality and anomaly detection among β-VAE models.
- Models with constrained latent space reach higher detection metrics but lower reconstruction quality.
- Performance variability across random seeds is linked to distance between normal and abnormal latent distributions.
- Two methods mitigate the trade-off: beta-scheduling and Sparse VAE.
- Sparse VAE shows improvement in detection while maintaining high reconstruction quality.
- The paper is published on arXiv under Computer Science > Machine Learning.
Entities
Institutions
- arXiv