When Signal Processing Meets Machine Learning: Two case Studies

Ημερομηνία Διεξαγωγής: 
17/05/2017 - 12:00 - 13:00

ΟΜΙΛΗΤΗΣ:  Σέργιος Θεοδωρίδης, Καθηγητής Τμήμα Πληροφορικής και Τηλεπικοινωνιών, ΕΚΠΑ

ΤΙΤΛΟΣ:   When Signal Processing Meets Machine Learning: Two case Studies

ΧΡΟΝΟΣ:   Τετάρτη, 17/5/2017, 12:00 μμ

ΑΙΘΟΥΣΑ:  Αίθουσα Συνεδριάσεων Α56 (1ος όροφος)



In this talk, we will discuss new results concerning some of the current trends in SP, in the framework of two case studies. In the first one, a new method is presented for learning nonlinear regression/classification models, in kernel spaces, via online learning techniques, when data are collected in a distributed fashion. Online learning, and in particular gradient descent type algorithms for learning from data, enjoy a strong come back. When dealing with big data, batch algorithms become extremely demanding; also, expensive online algorithms lose the advantages that they enjoy in cases of small size data. Given a fixed computational budget, a “cheap” algorithm can do better than an “expensive” one, since it can use more data, which is available, anyway. Furthermore, data can be distributed in various places or they can be obtained in a distributed environment, such as sensor networks.

In the second part of the talk, we focus on latent variables and blind matrix factorization, one of the hottest areas in learning, since compact representation of the available data is of paramount importance. Our focus is on dictionary leaning techniques. Signals are represented in terms of learnable, from the data, “bases” vectors (dictionaries); this is in direct contrast with expansions in terms of fixed bases, such as in Fourier or wavelet transforms. Moreover, these dictionaries can comprise a large number of possible vectors, and the notion of sparsity is mobilized to select the “good” ones. The use of DL is discussed in the context of functional Magnetic Resonance Imaging (fMRI). New techniques and optimization algorithms are discussed. Finally, and given the time, some hints on using multiway arrays (tensors) are discussed. Multiway arrays are becoming more and more popular, since they can grasp better the underlying structure (correlations) hidden in the data and can help in extracting and learn "knowledge" that resides in them.