We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

stat.ML

Change to browse by:

References & Citations

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Machine Learning

Title: Rethinking Assumptions in Deep Anomaly Detection

Abstract: Though anomaly detection (AD) can be viewed as a classification problem (nominal vs. anomalous) it is usually treated in an unsupervised manner since one typically does not have access to, or it is infeasible to utilize, a dataset that sufficiently characterizes what it means to be "anomalous." In this paper we present results demonstrating that this intuition surprisingly seems not to extend to deep AD on images. For a recent AD benchmark on ImageNet, classifiers trained to discern between normal samples and just a few (64) random natural images are able to outperform the current state of the art in deep AD. Experimentally we discover that the multiscale structure of image data makes example anomalies exceptionally informative.
Comments: 17 pages; accepted at the ICML 2021 Workshop on Uncertainty & Robustness in Deep Learning
Subjects: Machine Learning (cs.LG); Machine Learning (stat.ML)
Cite as: arXiv:2006.00339 [cs.LG]
  (or arXiv:2006.00339v2 [cs.LG] for this version)

Submission history

From: Lukas Ruff [view email]
[v1] Sat, 30 May 2020 19:30:38 GMT (18226kb,D)
[v2] Sat, 10 Jul 2021 10:11:26 GMT (9753kb,D)

Link back to: arXiv, form interface, contact.