Anomaly Detection in Deep Nets

Anomaly Detection in Deep Nets

The Implicit Geometry of Failure: Using Manifold Hypnosis to Expose Out-of-Distribution Activations

The Hidden Geometry of Model FailureEvery neural network, no matter how well-trained, harbors a secret geometry. Activations don't just float in high-...

Anomaly Detection in Deep Nets

Hypnotic Attractors: Detecting Anomalous Representations in Deep Network Latent Spaces

The Hidden Danger of Hypnotic AttractorsIn our work with deep neural networks, we've encountered a subtle but pernicious failure mode: hypnotic attrac...