CPDGA: Change point driven growing auto-encoder for lifelong anomaly detection

Lifelong learning addresses the challenge of acquiring new knowledge and tackling tasks in a continually evolving environment. Although this thread research has recently received increased interest, most lifelong machine approaches proposed thus far focus on object recognition or classification tasks. In contrast, for anomaly detection are still unexplored. This paper presents method loosely based biological principles, which can adapt to environment efficiently recall old information from its memory bank. Inspired by interaction between cortex hippocampus biology, we combine deep with statistical change point detection. Our induces concepts organizes them semantically coherent forest structure an unsupervised manner. At runtime, analyze objects, one one, respect current concepts. If fits existing concept, it is added pool objects representing that concept. Otherwise, further analyzed determine whether represents sub-concept, anomaly. Experiments conducted over different applied settings show synergic yields higher performance than state-of-the-art methods.