Seminar: Change Detection in High Dimensional Datastreams

2 Apr 2019 11:00 AM - 12:00 PM
Presenter/Speaker: Giacomo Boracchi, Politecnico di Milano, DEIB - Dipartimento di Elettronica, Informazione e Bioingegneria
Location: G.1.15

Change detection problems are ubiquitous in science and engineering: promptly detecting changes is often key to understand the dynamics of a monitored process and for activating suitable countermeasures.

During this talk, I will address the problem of detecting distribution changes in high-dimensional datastreams and present QuantTree, a recursive binary splitting scheme that yields histograms for change-detection purposes. In fact, we theoretically prove that in
QuantTree the bin probabilities do not depend on the distribution of stationary data, and the same holds for any test statistics based on bin counts, like the Pearson's one. Therefore, when using QuantTree it is possible to numerically compute the detection thresholds on
univariate and synthetically generated data, yet guaranteeing a controlled false positive rate in any dimension and for any data distribution. Experiments show that QuantTree can effectively detect changes and control the false positive rate in high dimensional
datastreams, even when the number of training samples is relatively small.

Our extensive experiments also indicate that all the considered techniques suffer of the detectability loss problem, namely that detecting a change of a fixed magnitude becomes increasingly more difficult when the data dimension scales. This is an intrinsic difficulty of change-detection methods, which we further investigate and analytically demonstrate to occur when monitoring the log-likelihood of a Gaussian datastream.

BIO
Giacomo Boracchi is an Assistant Professor of Computer Engineering at Dipartimento di Elettronica, Informazione e Bioingegneria of the Politecnico di Milano, where he also received the Ph.D. degree in information technology (2008), after graduating in Mathematics
(Università Statale di Milano, 2004).  His reserch interests concern image processing and machine learning, and in particular image restoration and analysis, change/anomaly detection, domain adaptation.

Since 2016 he has been teaching PhD courses concerning image processing and classification in Politecnico di Milano and Tampere University of Technology (Finland).

Since 2015 he is leading industrial research projects concerning algorithms for X-ray inspection systems for airport security, and automatic quality inspection systems for monitoring silicon wafer production. He has published more than 70 papers in international
conferences and journals and he is currently associate editor in IEEE Transactions on Image Processing and IEEE Computational Intelligence Magazine. In 2015 he has received the IBM Faculty Award, in 2016 the IEEE Transactions on Neural Networks and Learning Systems Outstanding Paper Award and in 2017 the Nokia Visiting Professor Scholarship.