Designing Out-of-distribution Data Detection using Anomaly Detectors: Single Model vs. Ensemble

Abstract

Image classification neural networks tend to give high probabilities to images they in fact do not recognize. This paper compares three approaches to detect such out-of-distribution data: One-Class Support Vector Machine, Isolation Forest and Local Outlier Factor. The experiments show that Isolation Forest outperforms the other two approaches. However, when combining the three algorithms using a majority voter, the results show that this ensemble is better at detecting out-of-distribution data than using the Isolation Forest algorithm solely.