Deep Anomaly Detection using Geometric Transformations (FAU Winter Semester 2024/2025)

This presentation occurred for the “Intro to Control and Machine Learning” master’s course at Frierich-Alexander Universität Erlangen-Nürnberg.

Repository:

• https://github.com/pedroblossbraga/DeepAnomDetecGeomTransf

Experiments

• New image transformations
  • quantile histogram equalization
  • color jitter
  • zoom / crop
• Alternative normality scoring, via Shannon Entropy
• Spatial Localization of Features
  • Averaging activation maps of relevant convolutional or activation layers
  • Grad-CAM: weighting 2D-activations with the average gradient
• Uncertainty Analysis
  • Monte Carlo Dropout

Presentation

View the presentation

Report - additional experiments

View the presentation

References

    [1] I. Golan and R. El-Yaniv, in Advances in Neural Information Processing Sys-
    tems, edited by S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-
    Bianchi, and R. Garnett (Curran Associates, Inc., 2018), vol. 31.