Anomaly Detection Requires Better Representations

School of Computer Science and Engineering
The Hebrew University of Jerusalem, Israel
SSLWIN Workshop - ECCV 2022

Abstract

Anomaly detection seeks to identify unusual phenomena, a central task in science and industry. The task is inherently unsupervised as anomalies are unexpected and unknown during training. Recent advances in self-supervised representation learning have directly driven improvements in anomaly detection. In this position paper, we first explain how self-supervised representations can be easily used to achieve state-of-the-art performance in commonly reported anomaly detection benchmarks. We then argue that tackling the next-generation of anomaly detection tasks requires new technical and conceptual improvements in representation learning.

Anomaly Detection - A Two-stage Paradigm

Current, state-of-the-art anomaly detection methods follow this simple, two-stage paradigm:
(i) Each data point is transformed to a representation, often learned in a self-supervised manner.
(ii) A density estimation model, often as simple as a kNN estimator, is fitted to the normal data provided in a training set.
To classify a new sample as normal or anomalous, its estimated probability density is computed - low likelihood samples are denoted as anomalies.

Bottlenecks in Representations Learning

Altough the above two-stage paradigm is successful in solving many anomaly detection problems, the reality is more complex

MAE vs. DINO nearest neighbors the representations learnt by MAE underperform contrastive self-supervised methods on unsupervised tasks such as anomaly detection.

Multi-modal datasets anomaly detection accuracy is much lower on more complex datasets containing multiple small objects, complex backgrounds, and when anomalies consist of related object categories.

3D Point Clouds when dealing with 3D point-cloud, self-supervised representations are yet to outperform handcrafted features for anomaly detection.

Tabular Data representation learning for general datasets is an open research question. Some prior knowledge of the dataset must be used in order to learn non-trivial data representations, at least in the context of anomaly detection.

@article{reiss2022anomaly, title={Anomaly Detection Requires Better Representations}, author={Reiss, Tal and Cohen, Niv and Horwitz, Eliahu and Abutbul, Ron and Hoshen, Yedid}, journal={arXiv preprint arXiv:2210.10773}, year={2022} }