Dr. Jiaxin Zhang
Oak Ridge National Lab
Friday, August 27, 2021 3:30-4:30pm
Abstract: Anomaly detection that aims at identifying deviant instances and patterns from the normal data distribution has been one of the important problems in machine learning and artificial intelligence. One typical example is that the detection of manufacturing errors is critical in fabrication processes to ensure product quality. Unfortunately, many defects or imperfections occur very rarely, and their characteristics are mostly unknown, the detection is still a challenging research question in real-world applications. Many recent advances have been made in the field, including the unsupervised method, semi-supervised method etc. However, these existing methods mainly focus on image-based tasks and are not straightforward to be generalized for other types of data, such as high-dimensional time-series data, tabular data, and graph data. Another challenge is the detection score lacks explicit likelihood information such that the reliability and robustness concerns are rise up.
In this work, we present a unifying end-to-end view and propose a probabilistic method to relax current generalization assumptions with a flow-based model. We extend the applicability of transformation-based methods to general data by learning the transformation. The key idea is to embed the transformed data into a semantic space such that the transformed data still resemble their original form, while different transformations are distinguishable. To improve the robustness of the detection performance, we propose to extract the underlying features by a generative decoder leveraging the conditional normalizing flow which can explicitly estimate the likelihood of the encoded features. To demonstrate our capability, we will show several benchmarking problems, including time-series and tabular datasets, as well as real-world applications of manufacturing detections at ORNL MDF and beam accelerator failure detection at ORNL SNS.
Bio: Dr. Jiaxin Zhang is a Research Staff in Machine Learning and Data Analytics Group, Computational Science and Mathematics Division at Oak Ridge National Laboratory. He received his Ph.D. in Civil and System Engineering with a dual M.S. in Applied Mathematics & Statistics at Johns Hopkins University in 2018. His current research interest is on Artificial Intelligence for Science and Engineering (AISE). My broadly interests revolve around robust deep learning, AI for design, uncertainty quantification, inverse problems and optimization, with a wide range of scientific applications to computational mechanics, advanced materials design, and advanced manufacturing systems.