Lifelong Anomaly Detection
This talk presents a task-agnostic unsupervised lifelong learning scheme for anomaly detection. The approach builds a long-term memory through hierarchical growth and uses change-point detection to set thresholds autonomously for new concepts and anomaly detection. A new changepoint detection method designed for high-dimensional settings, and for performing identification in addition to detection is introduced. Finally, a version of the system outfitted with memory consolidation, memory summarization, and experience replay is presented. The different versions of our system and its components are tested on cybersecurity, energy, weather prediction, and gravitational wave data as well as on the TCPD Benchmark for change-point detection.
Biography: Nathalie Japkowicz is the Chair of the Computer Science Department at American University. She joined the university in August 2016. Prior to that, she directed the Laboratory for Research on Machine Learning applied to Defense and Security at the University of Ottawa in Canada. She is an AI/Machine Learning researcher particularly interested in lifelong machine learning, anomaly detection, and the handling of uncharacteristic data including datasets plagued by class imbalances. She trained over 30 graduate students. Her research has been funded by the DARPA L2M program (US), NSERC (Canada), DRDC (Canada), Health Canada, and various private companies. Her publications include a co-authored book entitled Evaluating Learning Algorithms at Cambridge University Press (2011; newer version under contract), one edited book in the Springer Series on Big Data (2016), and over 100 book chapters, journal articles, and conference or workshop papers. She received five best paper awards, including the prestigious European Conference on Machine Learning 2014 Test of Time award, and was awarded the Canadian Artificial Intelligence Association Distinguished Service Award in 2021.