Stop Here - Finding the Best Epoch for Autoencoders in Unsupervised Anomaly Detection

Autoencoders (AE) have emerged as a powerful unsupervised anomaly detection method with widespread applications in industrial settings. However, the presence of anomalies in the training set often leads to performance degradation during the training stages. To optimize AE's effectiveness, identifying the best epoch during training becomes essential. In this paper, we introduce a novel method for detecting the best epoch. Our approach centers on a new metric, Density-based Score Entropy (DSE), which leverages the density of original data and the distribution of estimated anomaly scores without relying on label information. This metric effectively monitors AE's performance changes throughout the training process. Experimental results show that the proposed method can accurately estimate the best epoch and outperforms existing methods. Additionally, it provides valuable insights into revealing AE's dynamic characteristics during the training process.