Signal Processing based on Unorganized Machines: Information-Theoretic Learning, Lp Norms and Deconvolution

• Grant number: 13/06322-0
• Support Opportunities: Scholarships in Brazil - Post-Doctoral
• Start date: August 01, 2013
• End date: December 31, 2014
• Field of knowledge: Engineering - Electrical Engineering - Telecommunications
• Principal Investigator: Romis Ribeiro de Faissol Attux
• Grantee: Levy Boccato
• Host Institution: Faculdade de Engenharia Elétrica e de Computação (FEEC). Universidade Estadual de Campinas (UNICAMP). Campinas , SP, Brazil

Abstract

The possibility of using information-theoretic learning (ITL) methods to supervised filtering can be regarded as attractive as it enables a more effective exploitation of the statistical information concerning the reference signal and the inputs of the filter.From a structural standpoint, an interesting example that aims at exploring the statistical properties of the signals in a more extensive manner without relinquishing a relatively simple training process is given by the class of modern unorganized machines, among which we highlight the echo state networks (ESNs) and the extreme learning machines (ELMs). The training process of these neural models is restricted to adjusting the parameters of the output layer (readout), being the standard adaptation criterion based on an estimate of the mean squared error (MSE).Motivated by the relevance of these features, we propose to study in this postdoctoral research project a new adaptive filtering paradigm which combines the structure of ESNs/ELMs with ITL-based criteria, more specifically, those of minimum error entropy and maximum correntropy, and also with criteria associated with different norms of the error signal, such as the L1 norm. This paradigm can be seen as a contribution to the field of artificial neural networks, since this possibility has not been thoroughly addressed in the corresponding literature, and will be also analized from the signal processing perspective, particularly in the context of the deconvolution task, both in the supervised and unsupervised scenarios.