search

XFeiF / ComputerVision_PaperNotes

📚 Paper Notes (Computer vision)
1 stars 0 forks source link

19ICLR| Deep Anomaly Detection with Outlier Exposure #9

Closed XFeiF closed 3 years ago

XFeiF commented 4 years ago

paper && code

Briefly, the authors improve deep anomaly detection by training anomaly detectors against an auxiliary dataset of outliers, an approach they call Outlier Exposure (OE).

For example, they use CIFAR10 as in-distribution set, while use 80 Million Tiny Images (exclude examples in CIFAR10) as outlier exposure dataset.

This enables anomaly detectors to generalize and detect unseen anomalies. In extensive experiments on natural language processing and small- and large-scale vision tasks, they find that Outlier Exposure significantly improves detection performance.

XFeiF commented 4 years ago

Q: Will different outlier exposure dataset affect the performance of anomaly detection?
Suppose you use A set as OE dataset, and the performance may be good if you test with a dataset similar to A. But if you use dataset B which is totally different from A, will it generalize well?