 |
Center for Nonlinear Studies |
Motivated by rapidly changing dynamics of energy systems and poor performance of machine learning algorithms in non-stationary noisy environments, we present a novel memory-limited streaming algorithm to track time-varying subspaces for detecting anom-alies. Our main theoretical contribution includes a novel proof-technique for analyzing convergence of the algorithm. Our algorithm uses low storage memory and shows promis-ing results as compared to other existing algorithms. This talk is based on joint work with SeYoung Yun and Daniel Bienstock.
Host: Michael Chertkov