MetaDetector: Detecting Outliers by Learning to Learn from Self-supervision

Abstract

Using self-supervision in anomaly detection can increase sensitivity to subtle irregularities. However, increasing sensitivity to certain classes of outliers could result in decreased sensitivity to other types. While a single model may have limited coverage, an adaptive method could help detect a broader range of outliers. Our proposed method explores whether meta learning can increase the adaptability of self-supervised methods. Meta learning is often employed in few-shot settings with labelled examples. To use it for anomaly detection, where labelled support data is usually not available, we instead construct a self-supervised task using the test input itself and reference samples from the normal training data. Specifically, patches from the test image are introduced into normal reference images. This forms the basis of the few-shot task. During training, the same few-shot process is used, but the test/query image is substituted with a normal training image that contains a synthetic irregularity. Meta learning is then used to learn how to learn from the few-shot task by computing second order gradients. Given the importance of screening applications, e.g. in healthcare or security, any adaptability in the method must be counterbalanced with robustness. As such, we add strong regularization by i) restricting meta learning to only layers near the bottleneck of our encoder-decoder architecture and ii) computing the loss at multiple points during the few-shot process.