% 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:255885,
author = {Jordan, Jakob and Petrovici, Mihai and Pfeil, Thomas and
Breitwieser, Oliver and Bytschok, Ilja and Bill, Johannes
and Gruebl, Andreas and Schemmel, Johannes and Meier,
Karlheinz and Diesmann, Markus and Tetzlaff, Tom},
title = {{D}eterministic neural networks as sources of uncorrelated
noise for probabilistic computations},
reportid = {FZJ-2015-05991},
year = {2015},
abstract = {Neural-network models of brain function often rely on the
presence ofnoise [1-5]. To date, the interplay of
microscopic noise sourcesand network function is only poorly
understood. In computersimulations and in neuromorphic
hardware [6-8], the number of noisesources (random-number
generators) is limited. In consequence, neuronsin large
functional network models have to share noise sources and
aretherefore correlated. In general, it is unclear how
shared-noisecorrelations affect the performance of
functional networkmodels. Further, there is so far no
solution to the problem of how alimited number of noise
sources can supply a large number offunctional units with
uncorrelated noise.Here, we investigate the performance of
neural Boltzmann machines[2-4]. We show that correlations in
the background activity aredetrimental to the sampling
performance and that the deviations fromthe target
distribution scale inversely with the number of
noisesources. Further, we show that this problem can be
overcome byreplacing the finite ensemble of independent
noise sources by arecurrent neural network with the same
number of units. As shownrecently, inhibitory feedback,
abundant in biological neural networks,serves as a powerful
decorrelation mechanism [9,10]: Shared-noisecorrelations are
actively suppressed by the network dynamics. Byexploiting
this effect, the network performance is
significantlyimproved. Hence, recurrent neural networks can
serve as naturalfinite-size noise sources for functional
neural networks, both inbiological and in synthetic
neuromorphic substrates. Finally weinvestigate the impact of
sampling network parameters on its abilityto faithfully
represent a given well-defined distribution. We showthat
sampling networks with sufficiently strong negative feedback
canintrinsically suppress correlations in the background
activity, andthereby improve their performance
substantially.Acknowledgments: Partially supported by the
Helmholtz Associationportfolio theme SMHB, the Jülich
Aachen Research Alliance (JARA), EUGrant 269921
(BrainScaleS), The Austrian Science Fund FWF
#I753-N23(PNEUMA), The Manfred Stärk Foundation, and EU
Grant 604102 (HumanBrain Project, HBP).[1] Rolls ET, Deco G:
The noisy brain. Oxford University Press, 2010[2] Hinton GE,
Sejnowski TJ, Ackley DH: Boltzmann machines:
constraintsatisfaction networks that learn. Technical
report, Carnegie-MellonUniversity, 1984[3] Buesing L, Bill
J, Nessler B, Maass W: Neural Dynamics asSampling: A Model
for Stochastic Computation in Recurrent Networks ofSpiking
Neurons. PloS CB, 2011, 7, e1002211.[4] Petrovici MA, Bill
J, Bytschok I, Schemmel J, Meier K: Stochasticinference with
deterministic spiking neurons. arXiv, 2013,
1311.3211v1[q-bio.NC][5] Probst D, Petrovici MA, Bytschok I,
Bill J, Pecevski D, SchemmelJ, Meier K: Probabilistic
inference in discrete spaces can beimplemented into networks
of LIF neurons. Front. Comput. Neurosci.,2015, 9:13.[6]
Schemmel J, Bruederle D, Gruebl A, Hock M, Meier K, Millner
S: AWafer-Scale Neuromorphic Hardware System for Large-Scale
NeuralModeling. Proceedings of the 2010 International
Symposium on Circuitsand Systems (ISCAS), IEEE Press, 2010,
1947-1950[7] Bruederle D, Petrovici M, Vogginger B, Ehrlich
M, Pfeil T, MillnerS, Gruebl A, Wendt K, Mueller E, Schwartz
MO et al.: A comprehensiveworkflow for general-purpose
neural modeling with highly configurableneuromorphic
hardware systems. Biological Cybernetics, 2011,
104,263-296[8] Petrovici MA, Vogginger B, Mueller P,
Breitwieser O, Lundqvist M,Muller L, Ehrlich M, Destexhe A,
Lansner A, Schueffny R et al.:Characterization and
Compensation of Network-Level Anomalies inMixed-Signal
Neuromorphic Modeling Platforms. PLoS ONE, 2014,
9(10):e108590.[9] Renart A, De La Rocha J, Bartho P,
Hollender L, Parga N, Reyes A,Harris KD: The asynchronous
State in Cortical Circuits. Science, 2010,327: 587-590[10]
Tetzlaff T, Helias M, Einevoll G, Diesmann M: Decorrelation
ofneural-network activity by inhibitory feedback. PloS CB,
2012, 8,e1002596},
month = {Jul},
date = {2015-07-18},
organization = {CNS, Prague (Chech Republic), 18 Jul
2015 - 23 Jul 2015},
cin = {INM-6 / IAS-6},
cid = {I:(DE-Juel1)INM-6-20090406 / I:(DE-Juel1)IAS-6-20130828},
pnm = {574 - Theory, modelling and simulation (POF3-574) / 899 -
ohne Topic (POF2-899) / SMHB - Supercomputing and Modelling
for the Human Brain (HGF-SMHB-2013-2017) / HBP - The Human
Brain Project (604102) / BRAINSCALES - Brain-inspired
multiscale computation in neuromorphic hybrid systems
(269921)},
pid = {G:(DE-HGF)POF3-574 / G:(DE-HGF)POF2-899 /
G:(DE-Juel1)HGF-SMHB-2013-2017 / G:(EU-Grant)604102 /
G:(EU-Grant)269921},
typ = {PUB:(DE-HGF)1},
url = {https://juser.fz-juelich.de/record/255885},
}