Graduate Student Seminar

Zois Boukouvalas

Title

An Introduction to Independent Component Analysis

Abstract

Independent component analysis (ICA) is a computational method for separating linearly mixed observations into the original sources by assuming that the sources are non-Gaussian signals and that they are all statistically independent from each other. ICA is a special case of blind source separation. One practical application is the "*cocktail party*". Therefore if we assume that we are in a room where two people are talking simultaneously and we also have located two microphones, then what we observe is a linear mixture of the voice signals that the two people produce. ICA algorithms attempt to separate those mixtures of signals and produce the original voice signals coming from the people. On this talk we will present the general framework of ICA method as well as a main algorithm named FastICA. Finally we present results of numerical experiments using natural sources (images) and examine the algorithm behavior if we violate the statistical independence.

Maria Barouti

Title

Adaptive Clustering for Monitoring Distributed Data Streams

Abstract

Clustering is a task of exploratory data mining that groups a set of objects. There are two types of clustering algorithms, partitional and hierarchical. Some desirable properties of a clustering algorithms are scalability as well as the ability to deal with different data types. Assuming that we have a set of vectors changing with time and we want to monitor their mean without computing it. Our goal is to reduce communication load by clustering. Therefore, clustering can be formulated as an optimization problem by partitioning this set into clusters, so that the norm of the cluster's mean is minimized. A correct choice of clusters yields a reduction in communication load. Unlikely many clustering algorithms that attempt to collect together similar data items, monitoring requires clusters with dissimilar vectors canceling each other as much as possible. Finally, even if clustering is no universal remedy, there are cases where clustering dissimilar vectors helps.