...
/Anomaly Detection with Autoencoders
Anomaly Detection with Autoencoders
Discover how autoencoders excel in detecting rare events by analyzing reconstruction errors to differentiate between normal and anomalous states.
We'll cover the following...
Anomaly detection is unarguably one of the best approaches in rare event detection problems. Especially, if the event is so rare that there are insufficient samples to train a classifier. Fortunately, an (extremely) rare event often appears as an anomaly in a process. They can, therefore, be detected due to their abnormality.
Petsche et al. 1996 have one of the early works in deep learning which developed anomaly detectors for rare event detection. They developed an autoassociator to detect an imminent motor failure.
The autoassociator was essentially a reconstructor. Petsche et al. 1996 showed that the autoassociator has a small reconstruction error on measurements recorded from healthy motors but a larger error on those recorded from faulty motors.
This difference in the reconstruction errors: small for normal conditions and large for a rare anomaly forms the basis of anomaly detection models for rare events.
Anomaly detection approach
If an event is extremely rare, we can use an anomaly detection approach to predict the rare event. In this approach, the rare event is treated as an anomaly. The model is trained to detect the anomaly.
Given a dataset ...