1517 articles – 1652 references  [version française]
HAL: hal-00136322, version 1

Detailed view  Export this paper
Fluctuation and Noise Letters 7, 1 (2007) L39-L60
On pooling networks and fluctuation in suboptimal detection framework
Steeve Zozor 1, Pierre-Olivier Amblard 1, Cédric Duchêne 1
(2007)

Motivated by biological neural networks and distributed sensing networks, we study how pooling networks – or quantizers – with random thresholds can be used in detection tasks. We provide a brief overview of the use of deterministic quantizers in detection by presenting how quantizers can be optimally designed for detection purposes. We study the behavior of these networks when they are used in a problem for which they are not optimal (mismatching). We show that adding random fluctuations to the thresholds of the networks can then enhance the performance of the quantizers, thus helping in the recovery of “a kind of” optimality. We also show that (for a small number of thresholds) it suffices to use random uniform quantizers, for which we provide a study of the behavior as a function of several parameters (size, fluctuation nature, observation noise nature). The conclusion to these studies are the robustness of the uniform quantizer used as a detector with respect to fluctuations added on its thresholds.
1:  Grenoble Images Parole Signal Automatique (GIPSA-lab)
CNRS : UMR5216 – Université Joseph Fourier - Grenoble I – Université Pierre-Mendès-France - Grenoble II – Université Stendhal - Grenoble III – Institut Polytechnique de Grenoble - Grenoble Institute of Technology
Computer Science/Signal and Image Processing

Engineering Sciences/Signal and Image processing
pooling networks – distributed sensing – noise-enhanced processing – detection – quantization