Gal Mishne (EE,Technion)
Tuesday, 8.11.2016, 11:30
In the analysis of high-dimensional data, manifold learning methods are used to reduce the dimensionality of the data, while preserving local neighborhoods and revealing meaningful structures. Out-of-sample function extension techniques are then used for analyzing new points, yet these techniques are inherently limited for handling outliers. I present an analysis of these limitations and propose a new Multiscale Anomaly Detection approach that overcomes them. We have applied our approach to challenging remote sensing datasets, demonstrating its robustness to outliers and its independence of the imaging sensor. As a more general solution, we propose Diffusion Nets, a new deep learning network for manifold learning that provides both out-of-sample extension and outlier detection. Our approach for out-of-sample extension is more efficient in both computational complexity and memory requirements than previous methods.
PhD research under the supervision of Prof. Israel Cohen.