% IMPORTANT: The following is UTF-8 encoded. This means that in the presence
% of non-ASCII characters, it will not work with BibTeX 0.99 or older.
% Instead, you should use an up-to-date BibTeX implementation like “bibtex8” or
% “biber”.
@INPROCEEDINGS{Jordan:155477,
author = {Jordan, Jakob and Petrovici, Mihai and Pfeil, Thomas and
Breitwieser, Oliver and Bytschok, Ilya and Bill, Johannes
and Gruebl, Andreas and Schemmel, Johannes and Meier,
Karlheinz and Diesmann, Markus and Tetzlaff, Tom},
title = {{N}eural {N}etworks as {S}ources of uncorrelated {N}oise
for functional neural {S}ystems},
reportid = {FZJ-2014-04643},
year = {2014},
abstract = {Neural-network models of brain function often rely on the
presence of noise [1,2]. Yet thebiological origin of this
noise remains unclear. In computer simulations and in
neuromorphichardware [3,4], the number of noise sources
(random-number generators) is limited. In conse-quence,
neurons in large functional network models have to share
noise sources and are thereforecorrelated. It is largely
unclear how shared-noise correlation affect the performance
of func-tional network models. Further, so far there is no
solution to the problem of how a limitednumber of noise
sources can supply a large number of functional units with
uncorrelated noise.Here, we first demonstrate that the
performance of two functional network models, attrac-tor
networks [5] and neural Boltzmann machines [2], is
substantially impaired by shared-noisecorrelations resulting
from a limited number of noise sources. Secondly, we show
that thisproblem can be overcome by replacing the finite
pool of independent noise sources by a (finite)recurrent
neural network. As shown recently, inhibitory feedback,
abundant in biological neu-ral networks, serves as a
powerful decorrelation mechanism [6,7]. Shared-noise
correlations areactively suppressed by the network dynamics.
By exploiting this effect, the network perfor-mance is
significantly improved. Finally, we demonstrate the
decorrelating effect of inhibitoryfeedback in a
heterogeneous network implemented in an analog neuromorphic
substrate [8]. Insummary, we show that recurrent neural
networks can serve as natural finite-size noise sourcesfor
functional neural networks, both in biological and in
synthetic neuromorphic substrates.},
month = {May},
date = {2014-05-07},
organization = {OCCAM 2014, Osnabrueck (Germany), 7
May 2014 - 9 May 2014},
cin = {INM-6 / IAS-6},
cid = {I:(DE-Juel1)INM-6-20090406 / I:(DE-Juel1)IAS-6-20130828},
pnm = {331 - Signalling Pathways and Mechanisms in the Nervous
System (POF2-331) / BRAINSCALES - Brain-inspired multiscale
computation in neuromorphic hybrid systems (269921) / SMHB -
Supercomputing and Modelling for the Human Brain
(HGF-SMHB-2013-2017) / HBP - The Human Brain Project
(604102) / 89574 - Theory, modelling and simulation
(POF2-89574)},
pid = {G:(DE-HGF)POF2-331 / G:(EU-Grant)269921 /
G:(DE-Juel1)HGF-SMHB-2013-2017 / G:(EU-Grant)604102 /
G:(DE-HGF)POF2-89574},
typ = {PUB:(DE-HGF)1},
url = {https://juser.fz-juelich.de/record/155477},
}