Towards Efficient Anomaly Detection Using Memory Broad Learning System

Anomaly detection aims to identify data points in real-world data that do not conform to expected patterns. The lack of labels, large sample size, and numerous features make it challenging to develop an effective and straightforward algorithm. This paper proposes a new approach called Memory Broad Learning System for Anomaly Detection (MemBLSAD) that can be trained end-to-end. MemBLSAD seamlessly integrates a memory module and a Double BLS-decoder module to achieve hybrid anomaly detection scores, thereby enhancing its capability. Despite the added complexity, MemBLSAD maintains the fast training ability of classic BLS and ensures an end-to-end training process. Our experiments on MNIST and CIFAR10 datasets demonstrate that MemBLSAD outperforms the comparison algorithms on the AUC measure. We also provide reliable visualizations to analyze MemBLSAD's outstanding performance and explain why it works better.