Reproduce “Deep Semi-Supervised Anomaly Detection” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Lukas Ruff, Robert A. Vandermeulen, Nico Görnitz, Alexander Binder, Emmanuel Müller, Klaus-Robert Müller, Marius Kloft

Deep SAD (Deep Semi-Supervised Anomaly Detection) introduces an end-to-end deep learning framework that effectively leverages both labeled normal and labeled anomalous samples alongside unlabeled data to enhance anomaly detection. The method consistently outperforms or matches competitors on image datasets like CIFAR-10 and demonstrates robustness to polluted training data, particularly excelling on complex, high-dimensional data.

Key insights

  • Leveraging even a small number of labeled normal and anomalous samples significantly improves anomaly detection performance and robustness compared to purely unsupervised methods.
  • An information-theoretic perspective, where normal data is compressed (low entropy) and anomalies are dispersed (high entropy) in the latent space, provides a theoretical foundation for the approach.
  • The method demonstrates strong generalization to diverse anomaly types, even when only a few known anomaly classes are available during training, by focusing on defining normality rather than categorizing all anomalies.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.