Anomaly Detection Requires Better Representations

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