Name
Title | ||
|---|---|---|
Analytical results on pseudo-polynomial functional-link neural units for blind density shaping |
Abstract | ||
|---|---|---|
In this paper we deal with the problem of approximat- ing the cumulative distribution of a quasi-stationary sig- nal by means of a parametric function. We present an al- gorithm based on a quasi-polynomial adaptive transforma- tion, and briefly recall an existing closely related technique 'mixture of densities' technique. Then through numerical simulations both on synthetic and real-world data we com- pare their performances in terms of convergence speed and computational complexity. pdf's are mono-modal, while the problem complicates when probability density functions are multi-modal. In this paper we propose to use a pseudo-polynomial non-linearity whose parameters are adjusted by means of a recursive algorithm based on the entropy-maximization principle; the proposed architecture is known as polynomial functional-link neuron. Functional-link neurons (12, 20) are well-known neural structures that have been introduced in order to strive to take into account the complex non-linear behavior of the biological neural units. Examples of re- cent applications in different research areas are given, for instance, in (8, 16). It is worth recalling here that biological evidence seems to suggest that density shaping is embedded in human neurons behavior (9). The proposed algorithm updates the parameters once for each step, there is no need for collecting samples, thus the algorithm needs no storage requirements. In order to asses the performances of the proposed ap- proach, we briefly recall a closely related method known as 'mixture of density' (18, 19) and compare through com- puter simulations the performances of the two techniques on both synthetic and real-world data. |
Year | DOI | Venue |
|---|---|---|
1999 | 10.1109/IJCNN.1999.831113 | Neural Networks, 1999. IJCNN '99. International Joint Conference |
Keywords | Field | DocType |
computational complexity,convergence,digital simulation,learning (artificial intelligence),maximum entropy methods,neural nets,polynomials,probability,signal processing,blind density shaping,convergence speed,cumulative distribution,mixture of densities technique,parametric function,pseudo-polynomial functional-link neural units,quasi-polynomial adaptive transformation,quasi-stationary signal,real-world data,synthetic data | Convergence (routing),Signal processing,Parametric equation,Pseudo-polynomial time,Polynomial,Computer science,Cumulative distribution function,Artificial intelligence,Artificial neural network,Machine learning,Computational complexity theory | Conference |
Volume | ISSN | ISBN |
2 | 1098-7576 | 0-7803-5529-6 |
Citations | PageRank | References |
0 | 0.34 | 5 |
Authors | ||
2 | ||
Name | Order | Citations | PageRank |
|---|---|---|---|
| Simone Fiori | 1 | 494 | 52.86 |
| P. Burrascano | 2 | 17 | 7.64 |