Simula@bi: Online inference from graph-connected multi-variate time series
Speaker: Baltasar Beferull-Lozano
In this talk, we present algorithm design for online learning of causal dependencies from multiple streaming data time-series, which in many real-world scenarios, are non-linear and non-stationary.
Moving from the linear case, we first cover online kernel-based algorithms for non-linear vector autoregressive models (NLVAR), where the course of dimensionality is tackled by using a Fourier-based random feature approximation.
Exploiting the fact that real-world time-series exhibit typically sparse connectivities, we propose a group-Lasso based optimization framework, which is solved using an iterative composite objective mirror descent method, yielding an online algorithm with fixed computational complexity per iteration.
We provide theoretical guarantees, proving that the algorithm can achieve sub-linear dynamic regret. Experimental performance results are also shown on prediction and topology inference for both real and synthetic data. Then, we consider the case where the time series are partially observed, that is, some of the samples are not available.
Finally, for completeness, we summarise several other related methods we have designed, such as online learning of non-linear structural equation models, simplicial-VAR models for higher order data structures, or based on invertible neural networks.
About the speaker
Prof. Baltasar Beferull-Lozano holds a MSc. in Physics (U. Valencia, Spain),
and MSc. and PhD. in Electrical Engineering (Univ. of Southern California).
He is an adjunct Chief Research Scientist at SimulaMet, where he is the Director of the Dept. of Signal and Information Processing for Intelligent Systems (SIGIPRO), created in Nov. 2021, and he is also the Director of the WISENET Center at UiA.
He has coordinated or participated as PI in more than 20 international Projects including EU Projects, a large number of NFR projects, such as a FRIPRO TOPPFORSK Grant Award, and two SFF-like Centers in Switzerland and Spain.
Prof. Beferull brings extensive research and technology transfer
experience from both the US. (Univ. of Southern California, AT&T Shannon/Bell Labs) and Europe (EPFL, Univ. of Valencia). He has co-authored over 200 technical papers in high-quality international journal and conferences (including several paper awards), 3 U.S. patents, supervised 12 PhD Theses and imparted seminars/courses at more than 80 different international institutions.
He is also an Evaluator for the EU Commission, Qatar QNRF, Research Council of Finland, US. NSF and Swiss NSF. Some of his publications have been used as part of the selected mandatory reading in various US and European universities.
His main areas of expertise are: Signal Processing, Data Science and Machine Learning for multi-sensor data and information systems, Distributed cooperative intelligence, Optimization and Control for cyber-physical systems and networks.