Direct Objective Function for Anomaly Detection

Currently, DeepLense supports the following models for unsupervised dark matter classification:

 - Adversarial Autoencoder
 - Convolutional Variational Autoencoder
 - Deep Convolutional Autoencoder
 - Restricted Bolzmann Machine

Most of the listed models are surrogate base approaches. While surrogate approaches have become the mainstream method for anomaly detection in recent years and have shown promising results, it is hard to ensure that the surrogate tasks share a consistent optimization direction with anomaly detection. The proposal offers to return to a direct objective function for anomaly detection, which maximized the distance between normal and anomalous data in terms of the joint distribution for image and feature representation. The above objective function is decomposed into the following fours components:

  - Mutual information between image space and latent space of normal data.
  - Entropy of normal data in latent space.
  - Expectations of cross-entropy between normal and anomalous samples in latent space.
  - Distribution distance between normal and anomalous samples in image space.