Targeted Collapse Regularized Autoencoder for Anomaly Detection: Black Hole at the Center.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Autoencoders are widely used for anomaly detection, assuming anomalies have high reconstruction errors.
• However, autoencoders can generalize, leading to low reconstruction errors for some anomalies.
• Existing methods often require complex components and training procedures.

Purpose of the Study:

• To propose a straightforward and effective alternative for enhancing autoencoder-based anomaly detection.
• To improve the differentiation between normal and anomalous samples without added complexity.
• To provide theoretical insights into the training process for anomaly detection.

Main Methods:

• Complementing the standard autoencoder reconstruction loss with a computationally light regularization term on latent space representations.
• Testing the proposed method on diverse visual and tabular datasets.
• Analyzing the training dynamics and theoretical underpinnings of the approach.

Main Results:

• The proposed method achieves performance comparable to or better than more complex alternatives.
• It demonstrates effectiveness across various data modalities (visual and tabular).
• Integration with state-of-the-art methods further boosts their performance.

Conclusions:

• A simple latent space norm regularization significantly improves autoencoder anomaly detection.
• The method is computationally efficient, requires minimal tuning, and is broadly applicable.
• This work demystifies autoencoder anomaly detection and opens avenues for future research.