讲座题目：Anomaly detection via robust autoencoding and variational autoencoding 主讲人：Dongmian Zou, Assistant Professor of Data Science at Duke Kunshan University 主持人：魏同权 副教授 开始时间：2020-11-13 11:00:00 讲座地址：中北校区理科大楼B504 主办单位：计算机科学与技术学院
报告人简介： Dongmian Zou is an Assistant Professor of Data Science at Duke Kunshan University. He has a B.Sc. in mathematics from the Chinese University of Hong Kong and a Ph.D. in applied mathematics and scientific computation from the University of Maryland, College Park. Before joining Duke Kunshan, he served as a post-doctorate researcher at the University of Minnesota, Twin Cities. His primary research is the intersection among applied harmonic analysis, machine learning and signal processing. He is especially interested in problems and methods in robust representations and structures in geometric and graph deep learning.
报告内容： Anomaly detection aims to identify data points that “do not conform to expected behavior”. It can be done either unsupervised (outlier detection) or semi-supervised (novelty detection). In this talk, we will discuss using robust reconstruction methods for both outlier detection and novelty detection. For outlier detection, we propose an autoencoder with a robust subspace recovery layer (RSR layer). This layer seeks to extract the underlying subspace from a latent representation of the given data and removes outliers that lie away from the subspace. Specifically, the encoder maps data into a latent space, from which the RSR layer extracts the subspace. The decoder then smoothly maps back the underlying subspace to a manifold close to the original inliers. Inliers and outliers are distinguished according to the distances between the original and mapped positions. For novelty detection, we propose a robust VAE with the following components: 1. Extracting crucial features of the latent code by a carefully designed dimension reduction component for distributions; 2. Modeling the latent distribution as a mixture of Gaussian low-rank inliers and full-rank outliers, where the testing only uses the inlier model; 3. Applying the Wasserstein-1 metric for regularization, instead of the KL-divergence; and 4. Using a least absolute deviation error for reconstruction. We illustrate state-of-the-art results for anomaly detection tasks on standard benchmarks. |
