Connectionists: Two new papers on learning theory.

[Simone G.O. FIORI]

Dear colleagues,

I take the liberty to announce the availability of two new 
papers on learning theory. They are related to unsupervised 
adapting of inherently-stable IIR filters implemented in 
state-space form, with application to blind system 
deconvolution, and to variate generation by look-up-table
type adaptive-activation-function-neuron learning.

1] "Blind Adaptation of Stable Discrete-Time IIR Filters in 
State-Space Form", by S. Fiori, University of Perugia, Italy.
Accepted on the IEEE Transactions on Signal Processing

*Abstract: Blind deconvolution consists of extracting a source 
sequence and impulse response of a linear system from their 
convolution. In presence of system zeros close to the unit 
circle, which give rise to very long impulse responses, IIR 
adaptive structures are of use, whose adaptation should be 
carefully designed in order to guarantee stability. In this 
paper, we propose a blind-type discrete-time IIR adaptive 
filter structure realized in state-space form that, with a 
suitable parameterization of its coefficients, remains stable. 
The theory is first developed for a two-pole filter, whose 
numerical behavior is investigated via computer-based 
experiments. The proposed structure/adaptation theory is then 
extended to a multi-pole structure realized as a cascade of 
two-pole filters. Computer-based experiments are proposed and 
discussed, which aim at illustrating the behavior of the filter 
cascade on several cases of study. The numerical results 
obtained show the proposed filters remain stable during 
adaptation and provide satisfactory deconvolution results.

Draft available at: 
http://www.unipg.it/sfr/publications/tsp05.pdf

2] "Neural Systems with Numerically-Matched Input-Output 
Statistic: Variate Generation", by S. Fiori, University of 
Perugia, Italy. Accepted on Neural Processing Letters

*Abstract: The aim of this paper is to present a neural system 
trained to exhibit matched input-output statistic for random 
samples generation. The learning procedure is based on a 
cardinal equation from statistics that suggests how to warp 
an available samples set of known probability density function 
into a samples set with desired probability distribution. The 
warping structure is realized by a fully-tunable neural system 
implemented as a "look-up table". Learnability theorems are 
proven and discussed and the numerical features of the 
proposed methods are illustrated through computer-based 
experiments. 

Draft available at: 
http://www.unipg.it/sfr/publications/rng_nepl.pdf

Bets regards.

 
 =================================================
|         Simone FIORI (Elec.Eng., Ph.D.)         |
| * Faculty of Engineering - Perugia University * |
|  * Polo Didattico e Scientifico del Ternano *   |
| Loc. Pentima bassa, 21 - I-05100 TERNI (Italy)  |
|   eMail: fiori at unipg.it - Fax: +39 0744 492925  |
|        Web: http://www.unipg.it/sfr/            |
 =================================================