Self-Supervised Anomaly Detection With Neural Transformations.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Data augmentation is crucial for self-supervised learning, particularly in anomaly detection.
• Existing hand-crafted transformations are effective for image data but lack for non-image data.

Purpose of the Study:

• To develop effective transformations for end-to-end anomaly detection on arbitrary data.
• To investigate the suitability of contrastive loss for learning data transformations.

Main Methods:

• Employed a contrastive loss function to learn diverse yet semantically relevant data transformations.
• Applied neural transformation learning to various data modalities including time series, tabular, text, and graph data.
• Evaluated the method's performance and interpretability in anomaly detection tasks.

Main Results:

• Demonstrated that contrastive loss is superior to previous losses for transformation learning, both theoretically and empirically.
• Achieved state-of-the-art anomaly detection performance on time series, tabular, text, and graph datasets.
• Showcased improved interpretability in image anomaly detection through learned transformations at multiple abstraction levels.

Conclusions:

• Neural transformation learning with contrastive loss offers a powerful and versatile approach for anomaly detection across diverse data types.
• The method significantly advances the field by enabling effective anomaly detection where traditional methods fall short.
• Future work can explore further applications and refinements of learned transformations for enhanced interpretability and performance.