Retour à la page Publications / Back to the Publications page
A. Johannet, L. Personnaz, G. Dreyfus, J.D. Gascuel, M. Weinfeld
Specification and Implementation of a Digital Hopfield-type
Neural Network with On-chip Training
IEEE Transactions on Neural Networks 3, 529 (1992).
Abstract
This paper addresses the definition of the requirements for
the design of a neural network associative memory, with on-chip
training, in standard digital CMOS technology. We investigate
various learning rules which are integrable in silicon, and we
study the associative memory properties of the resulting networks.
We also investigate the relationships between the architecture
of the circuit and the learning rule, in order to minimize the
extra circuitry required for the implementation of training. We
describe a 64-neuron associative memory with on-chip training,
which has been manufactured, and we outline its future extensions.
Beyond the application to the specific circuit described in the
paper, the general methodology for determining the accuracy requirements
can be applied to other circuits and to other auto-associative
memory architectures.
Cliquez ici pour obtenir le document / Click here to download the document .pdf.
S. Knerr, L. Personnaz, G. Dreyfus
Handwritten Digit Recognition by Neural Networks with Single-layer
Training
IEEE Transactions on Neural Networks 3, 962 (1992)
Abstract
We show that neural network classifiers with single-layer training
can be applied efficiently to complex real-world classification
problems such as the recognition of handwritten digits. We introduce
the STEPNET procedure, which decomposes the problem into simpler
subproblems which can be solved by linear separators. Provided
appropriate data representations and learning rules are used,
performances which are comparable to those obtained by more complex
networks can be achieved. We present results from two different
data bases: a European data base comprising 8,700 isolated digits,
and a zip code data base from the U.S. Postal Service comprising
9,000 segmented digits. A hardware implementation of the classifier
is briefly described.
Cliquez ici pour obtenir le document / Click here to download
the document postscript.
(or ps.gz) or .pdf.
[Retour
à la
page publication dans les articles de journaux]
S. Marcos, O. Macchi, C. Vignat, G. Dreyfus, L. Personnaz, and P. Roussel-Ragot
A unified framework for gradient algorithms used for filter adaptation and neural network training.
International Journal of Circuit Theory and Applications 20, 159--200 (1992). Best paper award.
Abstract
In this paper, we present in a unified framework the gradient algorithms employed in the adaptation of linear time filters (TF) and the supervised training of (non linear) neural networks (NN).
The optimality criteria used to optimize the parameters H of the filter or network are the least squares (LS) and least mean squares (LMS) in both contexts. They respectively minimize the total or the mean squares of the error e(k) between an (output) reference sequence d(k) and the actual system output y(k) corresponding to the input X(k). Minimization is performed iteratively by a gradient algorithm.
The index k, in (TF), is time; it runs indefinitely. Thus iterations start as soon as reception of X(k) begins. The recursive algorithm for the adaptation H(k - 1) - H(k) of the parameters is implemented each time a new input X(k) is observed. When training a (NN) with a finite number of examples, the index k denotes the example it is upperbounded. Iterative (block) algorithms wait until all the K examples are received to begin the network updating. However, K being frequently very large, recursive algorithms are also often preferred in (NN) training. But they raise the question of ordering the examples X(k).
Except in the specific case of a transversal filter, there is no general recursive technique for optimizing the LS criterion. However, X(k) is normally a random stationary sequence; thus LS and LMS are equivalent when k gets large. Moreover the LMS criterion can always be minimized recursively with the help of the stochastic LMS gradient algorithm, which has low computational complexity.
In (TF), X(k) is a sliding window of (time) samples, whereas in the supervised training of(NN) with arbitrarily ordered examples, X(k - 1) and X(k) have nothing to do with each other. When this (major) difference is rubbed out by plugging a time signal at the network input, the recursive algorithms recently developed for (NN) training become similar to those of adaptive filtering. In this context, the present paper displays the similarities between adaptive cascaded linear filters and trained multilayer networks. It is also shown that there is a close similarity between adaptive recursive filters and neural networks including feedback loops.
The classical filtering approach is to evaluate the gradient by "forward propagation" whereas the most popular (NN) training method uses a gradient backward propagation method. We show that when a linear (TF) problem is implemented by a (NN), the two approaches are equivalent. Yet, the backward method can be used for more general (non linear) filtering problems. Conversely, new insights can be drawn in the (NN) context by the use of a gradient forward computation. The advantage of the (NN) framework, and in particular of the gradient backward propagation approach, is evidently to have a much larger spectrum of applications than (TF), since (i) the inputs are arbitrary, and (ii) the (NN) can perform nonlinear (TF).
Cliquez ici pour obtenir le document / Click here to download the document .pdf.
S. Marcos, P. Roussel-Ragot, L. Personnaz, O. Nerrand,
G. Dreyfus, C. Vignat
Réseaux de Neurones pour le Filtrage Non-linéaire
Adaptatif
Traitement du Signal 8, 409-422 (1993)
Résumé
Nous introduisons une famille d'algorithmes adaptatifs permettant
l'utilisation de réseaux de neurones comme filtres adaptatifs
non linéaires, systèmes susceptibles de subir un
apprentissage permanent à partir d'un nombre éventuellement
infini d'exemples présentés dans un ordre déterminé.
Ces algorithmes, fondés sur des techniques d'évaluation
du gradient d'une fonction de coût, s'inscrivent dans un
cadre différent de celui de l'apprentissage "classique"
des réseaux de neurones, qui est habituellement non adaptatif.
Cliquez ici pour obtenir le document / Click here to download
the document postscript. (or
ps.gz) or .
C. Linster, C. Masson, M. Kerszberg, L. Personnaz,
G. Dreyfus
Computational Diversity in a Formal Model of the Insect Olfactory
Macroglomerulus
Neural Computation 5, 239-252 (1993)
Abstract
We present a model of the specialist olfactory system of selected
moth species and the cockroach. The model is built in a semi-random
fashion, constrained by biological (physiological and anatomical)
data. We propose a classification of the response patterns of
individual neurons, based on the temporal aspects of the observed
responses. Among the observations made in our simulations a number
relate to data about olfactory information processing reported
in the literature, others may serve as predictions and as guidelines
for further investigations. We discuss the effect of the stochastic
parameters of the model on the observed model behavior and on
the ability of the model to extract features of the input stimulation.
We conclude that a formal network, built with random connectivity,
can suffice to reproduce and to explain many aspects of olfactory
information processing at the first level of the specialist olfactory
system of insects.
Cliquez ici pour obtenir le document / Click here to download
the document postscript.
(or ps.gz)
or .pdf.
O. Nerrand, P. Roussel-Ragot, L. Personnaz, G. Dreyfus,
S. Marcos
Neural Networks and Non-linear Adaptive Filtering: Unifying
Concepts and New Algorithms
Neural Computation 5, 165-197 (1993)
Abstract
The paper proposes a general framework which encompasses the
training of neural networks and the adaptation of filters. We
show that neural networks can be considered as general non-linear
filters which can be trained adaptively, i. e. which can undergo
continual training with a possibly infinite number of time-ordered
examples. We introduce the canonical form of a neural network.
This canonical form permits a unified presentation of network
architectures and of gradient-based training algorithms for both
feedforward networks (transversal filters) and feedback networks
(recursive filters). We show that several algorithms used classically
in linear adaptive filtering, and some algorithms suggested by
other authors for training neural networks, are special cases
in a general classification of training algorithms for feedback
networks.
Cliquez ici pour obtenir le document / Click here to download
the document postscript.
(or ps.gz)
or .pdf.
O. Nerrand, D. Urbani, P. Roussel-Ragot, L. Personnaz,
G. Dreyfus
Training Recurrent Neural Networks : Why and How ? An Illustration
in Process Modeling
IEEE Transactions on Neural Networks 5, 178-184
(1994)
Abstract
The paper first summarizes a general approach to the training
of recurrent neural networks by gradient-based algorithms, which
leads to the introduction of four families of training algorithms.
Because of the variety of possibilities thus available to the
"neural network designer", the choice of the appropriate
algorithm to solve a given problem becomes critical. We show that,
in the case of process modeling, this choice depends on how noise
interferes with the process to be modeled; this is evidenced by
three examples of modeling of dynamical processes, where the detrimental
effect of inappropriate training algorithms on the prediction
error made by the network is clearly demonstrated.
Cliquez ici pour obtenir le document / Click here to download
the document postscript.
(or ps.gz) or
.pdf.
C. Linster, C. Masson
A neural model of olfactory sensory memory in the honeybee's
antennal lobe.
Neural Computation 8, 94-114 (1996).
Abstract
We present a neural model for olfactory sensory memory
in the honeybee's antennal lobe. To investigate the neural mechanisms
underlying odor discrimination and memorization, we exploit a
variety of morphological, physiological, and behavioral data.
The model allows us to study the computational capabilities of
the known neural circuitry, and to interpret under a new light
experimental data on the cellular as well as on the neuronal assembly
level. We propose a scheme for memorization of the neural activity
pattern after stimulus offset by changing the local balance between
excitation and inhibition. This modulation is achieved by changing
the intrinsic parameters of local inhibitory neurons or synapses
H. Gutowitz
The topological Skeleton of Cellular Automaton Dymamics
Physica D, accepted(1995).
Abstract
We have developed statistical techniques to study the structure
the state-transition graphs of cellular automata with periodic
boundary conditions, in the limit of large system size. We organize
our results around the concept of a topological skeleton. The
topological skeleton is the set of physically relevant states.
Covering this skeleton is a surface, typically thin and dense,
which contains the bulk of the set of states. States in the skeleton
have some long histories. States on the surface, by contrast,
have only short histories; they are reached only near the beginning
of cellular automaton evolution. We study in detail a sequence
of rules which exhibit mostly skeletal to mostly surface structure.
Cliquez ici pour lire l'article / Click here to read the paper
html.
Cliquez ici pour obtenir le document / Click here to download
the document postscript.
(or ps.gz) or .pdf.
C. Masson and C. Linster
Towards a cognitive understanding of odor discrimination: combining
experimental and theoretical approaches
Behavioural Processes, 35: 63-82 (1996).
Abstract
In response to changes in odorous environmental conditions,
most species (ranging from lower invertebrates to mammals) demonstrate
high adaptive behavioral performances. Complex natural chemical
signals (i.e. odorous blends involved in food search) are particularly
unstable and fluctuating, in quality, in space and time. Nevertheless,
adapted behavioral responses related to meaningful odor signals
can be observed even in complex natural odorous environments,
demonstrating that the underlying olfactory neural network is
a very performing pattern recognition device. In the honeybee,
a large number of experimental data have been collected at different
levels of observation of the olfactory system, from the signal
to the behaviour, including cellular and molecular properties.
However, no set of data considered by itself can give insight
into the mechanisms underlying odor discrimination and pattern
recognition. Here, concentrationg on the deciphering of the neural
mechanisms underlying encoding and decoding of the olfactory signal
in the first two layers of the neural network, we illustrate how
a theoretical approach helps us to integrate the different experimental
data and to extract relevant parameters (features) which might
be selected and used to store an odor representation in a behavioral
context.
C. Linster, G. Dreyfus
A model for Pheromone Discrimination in the Insect Antennal
Lobe: Investigation of the Role of Neuronal Response Complexity
Chemical Senses 21, 19-27 (1996)
Abstract
Based on anatomical and physiological data pertaining to several
moth species and the cockroach, we propose a neural model for
pheromone discrimination in the insect antennal lobe. The model
exploits the variety of neuronal response patterns observed in
the macroglomerulus, and predicts how these complex patterns of
excitation and inhibition can participate in the discrimination
of the species-specific pheromone blend. The model also allows
us to investigate the relationship between the distribution of
observed response patterns and the neural organization from which
these patterns emerge.
Cliquez ici pour obtenir le document / Click here to download
the document postscript.
(or ps.gz)
or .pdf.
H. Gutowitz
Cellular automata and the science of complexity
Complexity, vol 1, nos 5, 6 (1996).
Abstract
This two-part article reviews selected problems in the theory
of cellular automata, aiming to locate this theory with respect
to the theory of complex systems in general.
Cliquez ici pour lire l'article / Click here to read the paper
html [1],
[2].
Cliquez ici pour obtenir le document / Click here to download
the document postscript [1],
[2]. (or ps.gz
[1], [2]).
P. Simon, H. Gutowitz
A cellular automaton model of bi-directionnel traffic
Phys. Rev. E, vol. 57, 2441-2444 (1998).
Abstract
We investigate a cellular automaton (CA) model of traffic on
a bi-directional two-lane road. Our model is an extension of the
one-lane CA model of [Nagel 92], modified to account for interactions
mediated by passing. Values for the various parameters were chosen
so as to approximate real traffic. A density-flow diagram is calculated
and compared to that of a one-lane model.
Cliquez ici pour lire l'article / Click here to read the paper
html.
Cliquez ici pour obtenir le document / Click here to download
the document postscript.
(or ps.gz) or .pdf.
G. Dreyfus, Y. Idan
The Canonical Form of Non-linear Discrete-Time Models
Neural Comptation, 10, 133-164 (1998).
Abstract
Discrete-time models of complex nonlinear processes, whether
physical, biological or economical, are usually under the form
of systems of coupled difference equations. In order to analyze
such systems, one of the first tasks is that of finding a state-space
description of the process, i.e. a set of state variables and
the associated state equations. We present a methodology for finding
a set of state variables and a canonical representation of a class
of systems described by a set of recurrent discrete-time, time-invariant
equations. In the field of neural networks, this is of special
importance since the application of standard training algorithms
requires the network to be in a canonical form. Several illustrative
examples are presented.
Cliquez ici pour obtenir le document / Click here to download
the document postscript.
(or ps.gz) or .pdf.
B. Quenet, A. Lutz, G. Dreyfus, V. Cerny, C. Masson
A Dynamic Model of Key Feature Extraction
: the Example of Olfaction, I - Biological
Backgroundand Overview of the Properties of
the Model
Research report (1998).
Abstract
It has been inferred from experimental data that the extraction
of key features, and the emergence of stable internal representations,
are essential steps in the processing of the odorant signal. This
report presents a model of the formation of glomerular activity
patterns which accounts for these properties: despite the fluctuations
of the activity of the sensory neurons, the dynamics of this model
exhibits stable attractors which code for the key features of
the input signal, leading to a stable internal representation
at the glomerular level. The model is simple enough to be fully
analytically tractable, yet it embodies the biological ingredients
which allow it to perform relevant functions. One of the salient
features of the model is the fact that three regimes of synaptic
noise appear: (i) at low noise, the extraction of key features
and the stabilization of glomerular patterns are enhanced with
respect to the noise-free operation, (ii) at medium noise the
glomerular pattern of activity is similar to the pattern of activity
of the receptors, and (iii) at high noise the glomerular pattern
is very weakly correlated to the input signal. The first part
of this two-part report describes the behavioral and biological
background on which the model is based, and gives an overview
of the properties of the model. The companion report presents
a full mathematical treatment of the model.
Cliquez ici pour obtenir le document / Click here to download
the document postscript. (or
ps.gz) or .pdf.
B. Quenet, V. Cerny, G. Dreyfus, A. Lutz
A Dynamic Model of Key Feature Extraction:
the Example of Olfaction: II - Theoretical
Analysis by a Boltzmann-type Distribution
of Attractors,
Research report (1998).
Abstract
This report presents theoretical results derived in the
analysis of the model proposed in part I for the olfactory pathway.
Some of these results are model-specific, others are of more generic
interest. The latter include the description of the dynamics in
the presence of noise as a two-step Markov process: this leads
to the derivation of a Boltzmann-type distribution of the steady-state
probabilities of attractors for a discrete-time dynamic systems
with cycles of maximum length two. This leads to a clear understanding
of the phenomena described from simulations in part I, including
the emergence of three different noise regimes. More specific
of the model is the description of the deterministic dynamics
and the mathematical justification of the coding properties emerging
from the prevalent lateral inhibition in the model.
Cliquez ici pour obtenir le document / Click here to download
the document postscript. (or
ps.gz) or .pdf.
A. Duprat, T. Huynh, G. Dreyfus
Towards a Principled Methodology for Neural Network Design
and Performance Evaluation in QSAR; application to the Prediction
of LogP
J. Chem. Inf. Comp. Sci., to be published.
Abstract
The prediction of properties of molecules from their structure
(QSAR) is basically a nonlinear regression problem. Neural networks
are proved to be parsimonious universal approximators of nonlinear
functions; therefore, they are excellent candidates for performing
the nonlinear regression tasks involved in QSAR. However, their
full potential can be exploited only in the framework of a rigorous
statistical approach. In the present paper, we describe a principled
methodology for designing neural networks for QSAR and estimating
their performances, and we apply this approach to the prediction
of logP. We compare our results to those obtained on the same
molecules by other methods.
Cliquez ici pour obtenir le document / Click here to download
the document pdf.
I. Rivals, L. Personnaz
A Recursive Algorithm Based on the Extended Kalman Filter for
the Training of Feedforward Neural Models,
Neurocomputing, 20, 173-188 (1998).
Abstract
The Extended Kalman Filter (EKF) is a well known tool for the
recursive parameter estimation of static and dynamic nonlinear
models. In particular, the EKF has been applied to the estimation
of the weights of feedforward and recurrent neural network models,
i.e. to their training, and shown to be more efficient than recursive
and non recursive first-order training algorithms; nevertheless,
these first applications to the training of neural networks did
not fully exploit the potentials of the EKF. In this paper, we
analyze the specific influence of the EKF parameters for modeling
problems, and propose a variant of this algorithm for the training
of feedforward neural models which proves to be very efficient
as compared to non recursive second-order algorithms. We test
the proposed EKF
algorithm on several static and dynamic modeling problems,
some of them being benchmark problems, and which bring out the
properties of the proposed algorithm.
Cliquez ici pour obtenir le document / Click here to download
the document postscript (or ps.gz) or .pdf.
L. Personnaz, G. Dreyfus
Comment on "Discrete-time Recurrent Neural Network Architectures:
a Unifying Review"
Neurocomputing, 20, 325-331 (1998)
Abstract
The paper [Neurocomputing (1997) vol. 15, pp. 183-223] by Tsoi
and Back aimed at providing a unified presentation of neural network
architectures. We show in the present comment (i) that the canonical
form of recurrent neural networks presented by Nerrand et al.
[Neural Computation (1993) vol. 5, pp. 165-199] many years ago
provides the desired unification, (ii) that what Tsoi and Back
call Nerrand's canonical form is not the canonical form introduced
by Nerrand et al., and that (iii) contrarily to the claim of Tsoi
and Back, all neural network architectures presented in their
paper can be tranformed into Nerrand's canonical form. We show
that the contents of Tsoi and Back's paper obscures the issues
involved in the choice of a recurrent neural network instead of
clarifying them: this choice is definitely much simpler than it
might seem from Tsoi and Back's paper.
Cliquez ici pour obtenir le document / Click here to download
the document postscript.
(or ps.gz) or .pdf.
L. Personnaz, G. Dreyfus
Comment on "Recurrent Neural Networks: a Constructive
Algorithm, and its Properties"
Neurocomputing, 20, 321-324 (1998).
Abstract
In their paper [Neurocomputing (1997) vol. 15, pp. 309-326],
Tsoi and Tan present what they call a "canonical form",
which they claim to be identical to that proposed in Nerrand et
al [Neural Computation (1993) vol. 5, pp. 165-199]. They also
claim that the algorithm which they present can be applied to
any recurrent neural network. In the present comment, we disprove
both claims.
Cliquez ici pour obtenir le document / Click here to download
the document postscript.
(or ps.gz) or .pdf.
Y. Oussar, I. Rivals, L. Personnaz, G. Dreyfus
Training Wavelet Networks for Nonlinear Dynamic Input-Output
Modeling
Neurocomputing, 20, 173-188 (1998).
Abstract
In the framework of nonlinear process modeling, we propose
training algorithms for feedback wavelet networks used as nonlinear
dynamic models. An original initialization procedure is presented,
that takes the locality of the wavelet functions into account.
Results obtained for the modeling of several processes are presented;
a comparison with networks of neurons with sigmoidal functions
is performed.
Cliquez ici pour obtenir le document / Click here to download
the document postscript.
(or ps.gz) or .pdf.
I. Rivals, L. Personnaz
On Cross-Validation for Model Selection
Neural Computation, 11 , 863-870 (1998) .
Abstract
In response to (Zhu and Rower, 1996), a recent communication
(Goutte, 1997) established that leave-one-out cross validation
is not subject to the "no-free-lunch" criticism. Despite
this optimistic conclusion, we show here that cross-validation
has very poor performance for the selection of linear models as
compared to classic statistical tests. We conclude that the statistical
tests are preferable to cross-validation for linear as well as
for non linear model selection.
Cliquez ici pour obtenir le document / Click here to download
the document postscript
(or ps.gz) or .pdf.
Y. Oussar, G. Dreyfus
Initialization by Selection for Wavelet Network Training
Neurocomputing, vol. 34, pp. 131-143 (2000).
Abstract
We present an original initialization procedure for the parameters
of feedforward wavelet networks, prior to training by gradient-based
techniques. It takes advantage of wavelet frames stemming from
the discrete wavelet transform, and uses a selection method to
determine a set of best wavelets whose centers and dilation parameters
are used as initial values for subsequent training. Results obtained
for the modeling of two simulated processes are compared to those
obtained with a heuristic initialization procedure, and demonstrate
the effectiveness of the proposed method.
Cliquez ici pour obtenir le document / Click here to download
the document postscript
(or ps.gz) or
.pdf.
I. Rivals, L. Personnaz
Construction of confidence intervals for neural networks based
on least squares estimation
Neural Networks, vol. 13, pp. 463-484 (2000).
Abstract
We present the theoretical results about the construction
of confidence intervals for a nonlinear regression based on least
squares estimation and using the linear Taylor expansion of the
nonlinear model output. We stress the assumptions on which these
results are based, in order to derive an appropriate methodology
for neural black-box modeling; the latter is then analyzed and
illustrated on simulated and real processes. We show that the
linear Taylor expansion of a nonlinear model output also gives
a tool to detect the possible ill-conditioning of neural network
candidates, and to estimate their performance. Finally, we show
that the least squares and linear Taylor expansion based approach
compares favourably with other analytic approaches, and that it
is an efficient and economic alternative to the non analytic and
computationally intensive bootstrap methods.
Cliquez ici pour obtenir le document / Click here to download
the document postscript
(or ps.gz) or .pdf.
Bruno Cauli, James T. Porter, Keisuke Tsuzuki, Bertrand
Lambolez, Jean
Rossier, Brigitte Quenet, and Etienne Audinat
Classification of fusiform neocortical interneurons based on
unsupervised clustering
Proceedings of the National Academy of Sciences, Biological Sciences
: Neurobiology, vol. 97, pp. 6144-6149 (2000)
Abstract
A classification of fusiform neocortical interneurons (n=60)
was performed using an unsupervised cluster analysis based on
the comparison of multiple electrophysiological and molecular
parameters studied by patch-clamp and single cell multiplex RT-PCR
(RT-mPCR) in rat neocortical acute slices. The RT-mPCR protocol
was designed to detect simultaneously the expression of GAD 65
and 67, calbindin (CB), parvalbumin (PV), calretinin (CR), neuropeptide
Y (NPY), vasoactive intestinal peptide (VIP), somatostatin (SS),
cholecystokinin (CCK), AMPA, kainate, NMDA, and metabotropic glutamate
receptor (mGluR) subtypes. Three groups of fusiform interneurons
with distinctive features were disclosed by the cluster analysis.
The first type of fusiform neurons (n=12), termed RSNP-SS cluster,
was characterized by a firing pattern of regular spiking nonpyramidal
(RSNP) cells and by a high occurrence of SS. The second type of
fusiform neurons (n=32), termed RSNP-VIP cluster, predominantly
expressed VIP and displayed also firing properties of RSNP neurons
with accommodation profile different from those of RSNP-SS cells.
Finally, the last cluster of fusiform neurons (n=16) contained
a majority of irregular spiking (IS) VIPergic neurons. In addition,
the analysis of glutamate receptors revealed cell type specific
expression profiles. This study shows that combinations of multiple
independent criteria define distinct neocortical populations of
interneurones potentially involved in specific functions.
Gerard Arnold, Brigitte Quenet, Christian Papin,
Claudine Masson, Wolfgang H. Kirchner
Intra-Colonial Variability in the Dance Communication in Honeybees
Ethology, vol. 108, pp. 1-12 (2002).
Abstract
Given the importance of food collection for the survival
of their colonies, honeybees have evolved numerous mechanisms
for increasing colony-level foraging efficiency, mainly the combined
system of scout-recruit division of labour and recruitment communication.
A successful forager performs waggle dances on the surface of
the comb where it interacts with nectar receivers and dance followers.
A forager uses tremble dance when it experiences difficulty finding
a receiver bee to unload food upon return to the hive. There is
presently no data concerning the intra-colonial variability in
the dance communication. In this study, we determined the subfamily
frequencies of waggle and tremble dancers in a natural colony
where the subfamilies (patrilines) can be identified by microsatellite
genetic markers.
Our results demonstrate that a genetic component is associated
with the dance communication in honeybees. Out of the seventeen
subfamilies of the colony, members of only four subfamilies performed
50% of the waggle dances and members of only five subfamilies
performed 60% of the tremble dances. In each group, one subfamily
was found to be particularly specialised, since its members performed
about 6 times as many waggle or tremble dances as would be expected
from their proportion in the colony.
Cliquez ici pour obtenir le document / Click here to download the document .pdf.
I. Rivals, L. Personnaz
Nonlinear Internal Model Control Using Neural Networks
IEEE Transactions on Neural Networks, vol. 11, pp.80-90 (2000).
Abstract
We propose a design procedure of neural internal model control
systems for stable processes with delay. We show that the design
of such non adaptive indirect control systems necessitates only
the training of the inverse of the model deprived from its delay,
and that the presence of the delay thus does not increase the
order of the inverse. The controller is then obtained by cascading
this inverse with a rallying model which imposes the regulation
dynamic behavior and ensures the robustness of the stability.
A change in the desired regulation dynamic behavior, or an improvement
of the stability, can be obtained by simply tuning the rallying
model, without retraining the whole model reference controller.
The robustness properties of internal model control systems being
obtained when the inverse is perfect, we detail the precautions
which must be taken for the training of the inverse so that it
is accurate in the whole space visited during operation with the
process. In the same spirit, we make an emphasis on neural models
affine in the control input, whose perfect inverse is derived
without training. The control of simulated processes illustrates
the proposed design procedure and the properties of the neural
internal model control system for processes without and with delay.
Cliquez ici pour obtenir le document / Click here to download
the document postscript
(or ps.gz) or .pdf.
G. Arnold, B. Quenet, C. Masson
Influence of the social environment on the genetically-based
subfamily signature in the honeybee
Journal of Chemical Ecology, vol. 26, pp. 2321-2333 (2000).
Abstract
In honeybee, the cuticular hydrocarbon profiles are partly
genetically-based and differ between the subfamilies, which suggests
that they could, eventually, be used by the workers as labels
for subfamily recognition. This ability could potentially form
the basis for nepotistic conflicts between subfamilies that would
be detrimental to the inclusive fitness of the colony. Here we
have compared the subfamily hydrocarbon profiles of 5-day old
workers maintained in isolation with those matured in their parental
colony. We demonstrate that the cuticular hydrocarbon profiles
tend to be less distant between most subfamilies in the hive,
under a social environment, compared with those in isolation.
The main consequence of this partial homogenisation of the majority
of subfamily signatures could result in a reduction of the number
of recognisable subfamilies in the colony. Nevertheless, only
a few subfamilies retain very distinct cuticular hydrocarbon profiles.
J.M. Devaud, B. Quenet, J. Gascuel, C. Masson
Statistical Analysis and Parsimonious Modelling of Dendrograms
of In Vitro Neurones
Bulletin of Mathematical Biology, vol. 62, pp. 657-674 (2000).
Abstract
The processes whereby developing neurones acquire morphological
features that are common to entire populations (thereby allowing
the definition of neuronal types), are still poorly understood.
A mathematical model of neuronal arborisations may be useful to
extract basic parameters or organisation rules, hence helping
to achieve a better understanding of the underlying growth processes.
We present a parsimonious statistical model, intended to describe
the topological organisation of neuritic arborisations with a
minimal number of parameters. It is based on a probability of
splitting which depends only on the centrifugal order of segments.
We compare the predictions made by the model of several topological
properties of neurones with the corresponding actual values measured
on a sample of honeybee (olfactory) antennal lobe neurones grown
in primary culture, described in a previous study. The comparison
is performed for three populations of segments corresponding to
three neuronal morphological types previously identified and described
in this sample. We show that simple assumptions together with
the knowledge of a very small number of parameters, allow the
topological reconstruction of representative (bi-dimensional)
biological neurones. We discuss the biological significance (in
terms of possible factors involved in the determinism of neuronal
types) of both common properties and cell-type specific features,
observed on the neurones and predicted by the model.
Cliquez ici pour obtenir le document / Click here to download
the document postscript
or .pdf.
G. Monari, G. Dreyfus
Withdrawing an Example from the Training Set : an Analytic
Estimation of its Effect on a Non-linear Parameterised Model
Neurocomputing, vol. 35, pp. 195-201 (2000).
Abstract
For a non-linear parameterised model, the effects of withdrawing
an example from the training set can be predicted. We focus on
the prediction of the error on the left-out example, and of the
confidence interval for the prediction of this example. We derive
a rigorous expression of the first-order expansion, in parameter
space, of the gradient of a quadratic cost function, and specify
its validity conditions. As a consequence, we derive approximate
expressions of the prediction error on a given example, and of
the confidence interval thereof, had this example been withdrawn
from the training set. We show that the influence of an example
on the model can be summarised by a single parameter. These results
are applicable to leave-one-out cross-validation, with a considerable
decrease in computation time with respect to conventional leave-one-out.
The paper focuses on the theoretical aspects of the question;
both academic illustrations and large-scale industrial examples
are described in Gaetan Monari, Sélection
de modèles non linéaires par leave-one-out;
Etude théorique et application
des réseaux de neurones au procédé de soudage
par points,Thèse de Doctorat de l'Université
Pierre et Marie Curie - Paris VI (Novembre 1999).
Cliquez ici pour obtenir le document / Click here to download
the document postscript
or .pdf.
B. Quenet, D.
Horn, G. Dreyfus, R. Dubois
Temporal Coding in an Olfactory Oscillatory Model
Neurocomputing, vol. 38-40, pp. 831-836 (2001).
Abstract
We propose a model of the glomerular stage of the insect olfactory pathway that exhibits coding of inputs through spatio-temporal patterns of the type observed experimentally in the locust. Making use of the temporal bins provided by the oscillatory field potential we find that it suffices to employ simple Little-Hopfield dynamics to account for a rich repertoire of patterns. In particular, we show that we are able to repoprduce complex activity patterns from electrophysiological recordingsin insects. Biologi cally plausible mechanisms of synaptic adaptation are discussed.
Cliquez ici pour obtenir le document / Click here to download the document .pdf.
Y. Oussar, G. Dreyfus
How to be a Gray Box: Dynamic Semi-physical Modeling
Neural Networks, invited paper, vol. 14, pp. 1161-1172
(2001).
Abstract
A general methodology for gray-box, or semi-physical, modeling
is presented. This technique is intended to combine the best of
two worlds: knowledge-based modeling, whereby mathematical equations
are derived in order to describe a process, based on a physical
(or chemical, biological, etc.) analysis, and black-box modeling,
whereby a parameterized model is designed, whose parameters are
estimated solely from measurements made on the process. The gray-box
modeling technique is very valuable whenever a knowledge-based
model exists, but is not fully satisfactory and cannot be improved
by further analysis (or can only be improved at a very large computational
cost). We describe the design methodology of a gray-box model,
and illustrate it on a didactic example. We emphasize the importance
of the choice of the discretization scheme used for transforming
the differential equations of the knowledge-based model into a
set of discrete-time recurrent equations. Finally, an application
to a real, complex industrial process is presented.
Cliquez ici pour obtenir le document / Click here to download
the document postscript
or .pdf.
R. Lestienne, B. Quenet, S. Bouret, O. Parodi
Functionality of divergence/convergence in a model of the early olfactory system of the insect
Biological Cybernetics, vol. 87, pp. 220-229 (2002).
Abstract
Recent studies have shown that the insect olfactory system uses a spatio-temporal encoding of odours in the population of projection neurons in the antennal lobe, and suggest that the information thus coded is spread across a large population of Kenyon cells in the mushroom bodies. At this stage, the temporal part of the code might be transformed into a spatial code, especially via the temporally sensitive mechanisms of paired-pulse facilitation and feedback inhibition with its possible associated rebound. We explore here a simple model of the olfactory system using a three-layer network of formal neurons, comprising a fixed number (three) of projection and inhibitory neurons, but a variable number of Kenyon cells. We show how enlarging the divergence of the network (i.e. the ratio between the number of Kenyon cells to the number of input - projection - neurons) alters the number of different output spatial states in response to a fixed set of spatio-temporal inputs, and may therefore improve its effectiveness in discriminating between these inputs. Such enlarged divergence also reduces the variation of this effectiveness among random realisations of the network connectivity. Our model shows that the discriminative effectiveness first increases with the divergence, and then plateaus for a divergence factor of ~20. The maximal average number of different outputs was 470.2, which was computed from some simulations with random realisations of connectivity and with a set of 512 possible inputs. The discriminative effectiveness of the network is sensitive to paired-pulse facilitation, and especially to inhibition with rebound.
[ Retour à la liste des publications / Back to the list of publications ]
G. Monari, G. Dreyfus
Local Overfitting Control via Leverages
Neural Computation, vol. 14, pp. 1481-1506 (2002).
Abstract
We present a novel approach to dealing
with overfitting in black-box models. It is based on the leverages
of the samples, i.e. on the influence that each observation has
on the parameters of the model. Since overfitting is the consequence
of the model specializing on specific data points during training,
we present a selection method for nonlinear models, which is based
on the estimation of leverages and confidence intervals. It allows
both the selection among various models of equivalent complexities
corresponding to different minima of the cost function (e.g. neural
nets with the same number of hidden units), and the selection
among models having different complexities (e.g. neural nets with
different numbers of hidden units). A complete model selection
methodology is derived.
Cliquez ici pour obtenir le document / Click here to download the document pdf.
B. Quenet, S.
Sirapian, R. Dubois, G. Dreyfus, D. Horn
Modeling spatiotomporal olfactory data in two steps: from binary
to Hodgkin-Huxley neurons
Biosystems, vol. 67, pp. 203-211 (2002).
Abstract
Network models of synchronously updated
McCulloch-Pitts neurones exhibit complex spatiotemporal patterns
that are similar to activities of biological neurones in phase
with a periodic local field potential, such as those observed
experimentally by Wehr & Laurent (1996) in the locust olfactory
pathway. Modelling biological neural nets with networks of simple
formal units makes the dynamics of the model analytically tractable.
It is thus possible to determine the constraints that must be
satisfied by its connection matrix in order to make its neurones
exhibit a given sequence of activity (see, for instance, Quenet
et al, 2001). In the present paper, we address the following question:
how can one construct a formal network of Hodgkin-Huxley (HH)
type neurones that reproduces experimentally observed neuronal
codes? A two-step strategy is suggested in the present paper:
First, a simple network of binary units is designed, whose activity
reproduces the binary experimental codes,
Second, this model is used as a guide to design a network of more
realistic formal HH neurones.
We show that such a strategy is indeed fruitful: it allowed us
to design a model that reproduces the Wehr-Laurent olfactory codes,
and to investigate the robustness of these codes to synaptic noise.
Cliquez ici pour obtenir le document
/ Click here to download the document pdf.
C. Dreyfus, G.
Dreyfus
A Machine-learning Approach to the Estimation of the Liquidus
Temperature of Glass-forming Oxide Blends
Journal of Non-Crystalline Solids, vol. 318, pp. 63-78 (2003).
Abstract
Many properties of glasses and glass-forming liquids of oxide mixtures vary in a relatively simple and regular way with the oxide concentrations. In that respect, the liquidus temperature is an exception, which makes its prediction difficult: the surface to be estimated is fairly complex, so that usual regression methods involve a large number of adjustable parameters. Neural networks, viewed as parameterized nonlinear regression functions, were proved to be parsimonious: in order to reach the same prediction accuracy, a neural network requires a smaller number of adjustable parameters than conventional regression techniques such as polynomial regression. We demonstrate this very valuable property on some examples of oxide mixtures involving up to five components. In the latter case, we show that neural networks provide a sizeable improvement over polynomial methods.
Cliquez ici pour obtenir le document / Click here to download the document pdf.
B. Quenet, D.
Horn,
The Dynamic Neural Filter: a Binary Model o f Spatiotemporal
coding
Neural Computation, vol. 15, pp. 309-329 (2003).
Abstract
We describe and discuss the properties of a binary neural network that can serve as a dynamic neural filter (DNF), which maps regions of input sapce into spatiotemporal sequences of neuronal activity. Both deterministic and stochastic dynamics are studied, allowing the investigation of the stability of spatiotemporal sequences under noisy conditions. We define a measure of the coding capacity of a DNF and develop an algorithm for constructing a DNF that can serve as a source of given codes. On the basis of this algorithm, we suggest using a minimal DNF capable of generating observed sequences as a measure of the complexity of spatiotemporal data. This measure is applied to experimental observations in the locust olfactory system, whose reverberating local field potenntial provides a natural temporal scale allowing the use of a binary DNF. For random synaptic matrices, a DNF can generate very large cycles, thus becoming an efficient tool for producing spatiotemporal codes. The latter can be stabilized by applying to the parameters of the DNF a learning algorithm with suitable margins.
Cliquez ici pour obtenir le document / Click here to download the document pdf.
D. Horn, B. Quenet, G. Dror, O. Kliper
Modeling neural spatiotemporal behavior
Neurocomputing, vol. 52-54, pp. 799-804 (2003)
Abstract
We study some aspects of the dynamic neural filter (DNF), a recurrent network that produces spatiotemporal sequences in reaction to sets of constant inputs. The biological motivation for this study came from the observation of spatiotemporal patterns in the locust antennal lobe. Some of the aspects of these results can be reformulated and characterized by the DNF. Studying deterministic dynamics we find differences between low and high numbers of neurons. For low numbers there exists clear correlation between distances in input space and edit distances of spatiotemporal sequences. For large numbers of neurons we observe divergence between close-by spatiotemporal sequences. Nonetheless neuronal correlations survive for small changes in input space.
Cliquez ici pour obtenir le document / Click here to download the document pdf.
H. Stoppiglia,
G. Dreyfus, R. Dubois, Y. Oussar
Ranking a Random Feature for Variable and Feature Selection
Journal of Machine Learning Research, pp. 1399-1414 (2003).
Abstract
We describe a feature selection method, which can be applied directly to models that are linear with respect to their parameters, and indirectly to others. It is independent of the target machine. It is closely related to classical statistical hypothesis tests, but it is more intuitive, hence more suitable for use by engineers who are not statistics experts. Furthermore, some assumptions of classical tests are relaxed. The method has been used successfully in a number of applications that are briefly described.
Cliquez ici pour obtenir le document / Click here to download the document pdf.
D. Horn, G. Dror, B. Quenet Abstract Recurrent networks can generate spatio-temporal neural sequences of very large cycles, having an apparent random behavior. Nonetheless a proximity measure between these sequences may be defined through comparison of the synaptic weight matrices that generate them. Following the dynamic neural filter (DNF) formalism we demonstrate this concept by comparing teacher and student recurrent networks of binary neurons. We show that large sequences, providing a training set well exceeding the Cover limit, allow for good determination of the synaptic matrices. Alternatively, asssuming the matrices to be known, very fast determination of the biases can be achieved. Thus, a spatio-temporal sequence may be regarded as spatio-temporal encoding of the bias vector. We introduce a linear support vector machine (SVM) variant of the DNF in order to specify an optimal weight matrix. This approach allows us to deal with noise. Spatio-temporal sequences generated by different DNFs with the same number of neurons may be compared by calculating correlations of the synaptic matrices of the reconstructed DNFs. Other types of spatio-temporal sequences need the introduction of hidden neurons, and/or the use of a kernel variant of the SVM appraoch. The latter is being defined as a recurrent support vector network (RSVN).
Dynamic proximity of spatio-temporal sequences
IEEE Transactions on Neural Networks, vol. 15, pp. 1002-1008 (2004)
Y. Oussar, G.
Monari, G. Dreyfus
Reply to the Comments on "Local Overfitting Control via
Leverages" in "Jacobian Conditioning Analysis for Model
Validation" by I. Rivals and L. Personnaz.
Neural Computation , vol. 16, pp. 419 - 443 (2004).
Abstract
"Jacobian Conditioning Analysis for Model Validation" by Rivals and Personnaz is a comment on Monari and Dreyfus (2002). In the present reply, we disprove the claims of Rivals and Personnaz. We point to flawed reasoning in their theoretical comments, and to errors and inconsistencies in their numerical examples. Our replies are substantiated by seven counter-examples, inspired from industrial data, which show that (i) the comments on the accuracy of the computation of the leverages are unsupported, and that (ii) following the approach advocated by Rivals and Personnaz leads to discarding valid models, and to validating overfitted models.
Cliquez ici pour obtenir le document / Click here to download the document pdf.
O. Kliper, D. Horn,
B. Quenet, G. Dror
Analysis of spatiotemporal patterns in a model of olfaction
Neurocomputing, vol. 58-60, pp. 1027-1032 (2004)
Abstract
We model spatiotemporal patterns in locust olfaction with the dynamic neural filter, a recuurrent network that produces spatiotemporal patterns in reaction to sets of constant inputrs. We specify, wiithhin the model, inputs corresponding to diifferent oddors and different concentrations of the same odor. Then we proceed to analyze the resulting spatiotemporal patterns of the neurones of our model. Using SVD we investigate three kinds of data: global spatiotemporal data consisting of neuronal firing patterns over the period of odor presentation, spatial data, i.e. total spike counts during this period, and local spatiotemporal data which are neuronal spikes in single temporal bins.
Cliquez ici pour obtenir le document / Click here to download the document pdf.
O. Kliper, D. Horn, B. Quenet
The inertial-DNF Model: Spatiotemporal Patterns with Two Time-Scales
Neurocomputing vol. 65-66 pp. 543-548 (2005)
Abstract
We introduce the inertial-DNF (iDNF) model, an expansion of the dynamic neural filter (DNF) model, a model generating spatiotemporal patterns similar to those observed in the locust antennal lobes (ALs). The DNF model, which was described in previous works, includes one temporal scale defining the discrete dynamics inherent to the model. It lacks a second, slow, temporal scale that exists in the biological spatiotemporal data, where one finds slow temporal patterns of individual neurons in response to odor. Using the iDNF, we examine mechanisms that lead to temporal ordered spatiotemporal patterns, similar to those observed in the experimental data. We conclude that a second temporal scale is crucial for the creation of temporal order within the evolving spatiotemporal pattern.
Cliquez ici pour obtenir le document / Click here to download the document pdf.
B. Quenet, G. Horcholle-Bossavit, A. Wohrer, G.
Dreyfus
Formal modeling with multistate neurons and multidimensional
synapses
Biosystems, vol. 79, pp. 21-32 (2005).
Abstract
Multistate neurons, a generalization of the popular McCulloch-Pitts binary neurons, are described; they are intended to model the fact that neurons may be in several different states of activity, while McCulloch-Pitts neurons model two states only: active or inactive. We show that, as a consequence, multidimensional synapses are necessary to describe the dynamics of the model. As an illustration, we show how to derive the parameters of formal multistate neurons and their associated multidimensional synapses from simulations involving Hodgkin-Huxley neurons. Our approach opens the way to solving, in a more biologically plausible way, two problems that were addressed before: (1) the resolution of 'inverse problems', i.e. the construction of formal networks whose dynamics follows a pre-defined spatio-temporal binary sequence, (2) the generation of spatio-temporal patterns that reproduce exactly the "code" extracted from experimental recordings (olfactory codes at the glomerular level).
Cliquez ici pour obtenir le document / Click here to download the document pdf.
R. Dubois, B. Quenet, Y. Faisandier, G. Dreyfus
Building meaningful representations for nonlinear modeling of 1D- and 2D-signals: applications to biomedical signals
Neurocomputing, vol. 69, pp. 2180 - 2192 (2006)
Abstract
The paper addresses two problems that are frequently encountered when modeling data by linear combinations of nonlinear parameterized functions. The first problem is feature selection, when features are sought as functions that are nonlinear in their parameters (e.g. Gaussians with adjustable centers and widths, wavelets with adjustable translations and dilations, etc.). The second problem is the design of an intelligible representation for 1D- and 2D- signals that have peaks and troughs that have a definite meaning for experts of the signal. To address the first problem, a generalization of the Orthogonal Forward Regression method is described. To address the second problem, a new family of nonlinear parameterized functions, termed Gaussian mesa functions, is defined. It allows the modeling of signals such that each significant peak or trough is modeled by a single, identifiable function. The resulting representation is sparse in terms of adjustable parameters, thereby lending itself easily to automatic analysis and classification, yet it is readily intelligible for the expert. An application of the methodology to the automatic analysis of electrocardiographic (Holter) recordings is described. Applications to the analysis of neurophysiological signals and EEG signals (early detection of Alzheimer’s disease) are outlined.
Cliquez ici pour obtenir le document / Click here to download the document pdf.
Y. Takahashi, P. Sanders, P. Jaïs, M. Hocini, R. Dubois, M. Rotter, T. Rostock, C. Nalliah, F. Sacher, J. Clementy, and M. Haïssaguerre,
Organisation of Frequency Spectra of Atrial Fibrillation: Relevance to Radiofrequency Catheter Ablation
Journal of Cardiovascular Electrophysiology, vol. 17, pp. 382-388, 2005.
Abstract
Introduction: We hypothesized that the frequency spectra of fibrillatory electrograms may reflect the complexity of activities perpetuating atrial fibrillation (AF). To test this hypothesis, we evaluated the frequency spectra in patients with paroxysmal AF in relation to catheter ablation. Methods and Results: This study comprised two protocols: 25 patients undergoing pulmonary vein (PV) isolation in protocol I, and 20 patients undergoing mitral isthmus linear ablation after PV isolation in protocol II. The mean of dominant frequency (DF) and organization index (the ratio of the area under the DF and its harmonics to the total power) were determined from 32-second recordings in the coronary sinus. In protocol I, a PV was considered “driver” of AF if isolation of the PV resulted in termination or slowing of AF (decrease in DF by ≥0.25 Hz). Twenty-one patients had AF termination during four PV isolation. Among these 21 patients, 13 patients with single driving PV showed significantly higher baseline organization index than eight patients with multiple driving PVs (0.45 ± 0.08 vs 0.35 ± 0.07, P = 0.009). Patients with multiple driving PVs showed a significant increase in the organization index to 0.45 ± 0.11 (P < 0.05) after isolation of the initial driving PVs. In protocol II, the baseline organization index was significantly higher in seven patients who had termination of AF during mitral isthmus ablation than 13 patients who did not (0.50 ± 0.10 vs 0.38 ± 0.07, P < 0.008). The baseline DF was not associated with
outcomes of ablation in both protocols.
Conclusions: A higher organization index of atrial electrograms is associated with termination of AF during limited ablation. This parameter may be useful to anticipate the extent of ablation.
Cliquez ici pour obtenir le document / Click here to download the document pdf.
P. Sanders, C. J. Nalliah, R. Dubois, Y. Takahashi, M. Hocini, M. Rotter, T. Rostock, F. Sacher, L.-F. Hsu, A. Jonsson, M. D. O'Neill, P. Jais, and M. Haissaguerre,
Frequency Mapping of the Pulmonary Veins in Paroxysmal Versus Permanent Atrial Fibrillation
Journal of Cardiovascular Electrophysiology, vol. 17, pp. 965-972, 2006.
Abstract
Aims: To evaluate the differences and contribution of pulmonary vein (PV) activity to the maintenance of paroxysmal and permanent atrial fibrillation (AF).
Methods and Results: Thirty-four patients with paroxysmal-(n=20) or permanent-AF (n=14) undergoing ablation were studied. Prior to ablation, using a 10-pole catheter, 32-seconds of electrograms were acquired from each PV and the coronary sinus (CS). The frequency of activity of each PV and the CS was defined as the highest amplitude frequency on spectral analysis. The effects of ablation on the AF cycle length (AFCL) and frequency within the CS, and on the termination of AF were determined.Significant differences were observed between paroxysmal- and permanent-AF. Paroxysmal-AF demonstrates a higher frequency PV activity (11.0±3.1 versus 8.8±3.0Hz;p=0.0003) but lower CS frequency (5.8±1.2 versus 6.9±1.4Hz;p=0.01) and longer AFCL (182±17 versus 158±21ms; p=0.002); resulting in greater PV-atrial frequency gradient (7.2±2.2 versus 4.2±2.9Hz; p=0.006). PV-isolation in paroxysmal-AF resulted in a greater decrease in atrial frequency (1.0±0.7 versus -0.05±0.4Hz;p<0.0001), prolongation of the AFCL (49±35 versus 5±6ms;p<0.0001) and AF termination (11/20 versus 0/14;p=0.0007) compared to permanent-AF.
Conclusion: Paroxysmal-AF is associated with higher frequency PV-activity and lesser atrial frequency compared to permanent-AF. Isolation of the PVs had a greater impact on the fibrillatory process in paroxysmal- compared to permanent-AF; suggesting that while the PVs have a role in maintaining paroxysmal-AF, these structures contribute less to the maintenance of permanent-AF.
Cliquez ici pour obtenir le document / Click here to download the document pdf.
A. Goulon-Sigwalt-Abram, A. Duprat, G. Dreyfus
From Hopfield Nets to recursive networks to graph machines: numerical machine learning for structured data (invited paper)
Theoretical Computer Science, vol. 344, pp. 298-334 (2005).
Abstract
The present paper is a short survey of the development of numerical learning from structured data, an old problem that was first addressed by the end of the years 1980, and has recently undergone exciting developments, both from a theoretical point of view and for applications. Traditionally, numerical machine learning deals with unstructured data, in the form of vectors: neural networks, graphical models, support vector machines, handle vectors of features that are assumed to be relevant for solving the problem at hand (classification or regression). It is often the case, however, that data is structured, i.e. is in the form of graphs; three examples will be described here: prediction of the properties of molecules, image analysis, and natural language processing. The traditional approach consists in handcrafting a vector representation of the structured data (features describing the molecules, “bag of words” for language processing), and subsequently training a machine to perform the task from that representation. By contrast, we describe here a family of approaches (RAAM’s, LRAAM’s, recursive or folding networks, graph machines) that are specifically designed to learn from structured data. We show that, despite the apparent diversity, two basic principles underlie the recent approaches: first, use structured machines to learn structured data; second, learn representations instead of handcrafting them; although neither principle is really new, they proved very successful for handling structured data, to the point of generating a novel branch of numerical machine learning.
I. Kulagina, S.M. Korogod, C. Batini, G. Horcholle-Bossavit and S.Tyc-Dumont
The electro-dynamics of the dendritic space in a Purkinje cell
Arch. Ital. Biology, in press.
Abstract
The functional geometry of the reconstructed dendritic arborization of Purkinje neurons is the object of this work. The combined effects of the local geometry of the dendritic branches and of the membrane mechanisms are computed in passive configuration to obtain the electrotonic structure of the arborization. Steady-currents applied to the soma and expressed as a function of the path distance from the soma form different clusters of profiles in which dendritic branches are similar in voltages and current transfer effectiveness. The locations of the different clusters are mapped on the dendrograms and 3D representations of the arborization. It reveals the presence of different spatial dendritic sectors clearly separated in 3D space that shape the arborization in ordered electrical domains, each with similar passive charge transfer effciencies. Further simulations are performed in active configuration with a realistic cocktail of conductances to find out whether similar spatial domains found in the passive model also characterize the active dendritic arborization. During tonic activation of excitatory synaptic inputs homogeneously distributed over the whole arborization, the Purkinje cell generates regular oscillatory potentials. The temporal patterns of the electrical oscillations induce similar spatial sectors in the arborization as those observed in the passive electrotonic structure. By taking a video of the dendritic maps of the membrane potentials during a single oscillation, we demonstrate that the functional dendritic field of a Purkinje neuron displays dynamic changes which occur in the spatial distribution of membrane potentials in the course of the oscillation. We conclude that the branching pattern of the arborization explains such continuous reconfiguration and discuss its functional implications.
Cliquez ici pour obtenir le document / Click here to download the document pdf.
G. Horcholle-Bossavit, B. Quenet, O. Foucart
Oscillation and coding in a formal neural network considered as a guide for plausible simulations of the insect olfactory system
Biosystems, in press
Abstract
For the analysis of coding mechanisms in the insect olfactory system, a fully connected network of synchronously updated McCulloch and Pitts neurons (MC-P type) was developed (Quenet and Horn, 2003). Using an internal clock, this "Dynamic Neural Filter" (DNF) produces spatiotemporal patterns identical to synchronized activities recorded from the Projection Neurons (PN) in the locust antennal lobe (AL) in response to different odors.
Here, in a first step, we separate the populations of PN and Local inhibitory Neurons (LN) and use the DNF as a guide for simulations based on biological plausible neurons (Hodgkin-Huxley: H-H type). We show that a parsimonious network of 10 H-H neurons generates action potentials corresponding exactly to the olfactory codes.
In a second step, we construct a new type of DNF in order to study the population dynamics when different delays are taken into account. We find synaptic matrices which lead to both the emergence of robust oscillations and spatio-temporal patterns, using a formal criterion, based on a Normalized Euclidian Distance (NED), in order to measure the use of the temporal dimension as a coding dimension by the DNF. Similarly to biological PN, the activity of excitatory neurons in the model can be both phase-locked to different cycles of the oscillations corresponding to the local field potential (LFP), and nevertheless exhibit dynamic behavior complex enough to be the basis of spatio-temporal codes.
Cliquez ici pour obtenir le document / Click here to download the document pdf.
F. Vialatte, C. Martin, R. Dubois, B. Quenet, R. Gervais, and G. Dreyfus,
A machine learning approach to the analysis of time-frequency maps, and its application to neural dynamics
Neural Networks, vol. 20, pp. 194 - 209 (2007).
Abstract
The statistical analysis of experimentally recorded brain activity patterns may require comparisons between large sets of complex signals in order to find meaningful similarities and differences between signals with large variability. High-level representations such as time-frequency maps convey a wealth of useful information, but they involve a large number of parameters that make statistical investigations of many signals difficult at present. In this paper, we describe a method that performs drastic reduction in the complexity of time-frequency representations through a modelling of the maps by elementary functions. The method is validated on artificial signals and subsequently applied to electrophysiological brain signals (local field potential) recorded from the olfactory bulb of rats while they are trained to recognize odours. From hundreds of experimental recordings, reproducible time-frequency events are detected, and relevant features are extracted, which allow further information processing, such as automatic classification.
Cliquez ici pour obtenir le document / Click here to download the document pdf.
B. Quenet, G. Horcholle-Bossavit
The locust olfactory system as a case study for modeling dynamics of neurobiological networks: from discrete time neurons to continuous time neurons
Arch. Ital.Biology, in press
Cliquez ici pour obtenir le document / Click here to download the document pdf.
A. Goulon, T. Picot, A. Duprat, and G. Dreyfus
Predicting activities without computing descriptors: graph machines for QSAR
SAR and QSAR in Environmental Resesarch, vol. 18, pp. 141 - 153 (2007)
Abstract
Keywords: Graph; Graph machine; Structured data; Machine learning; Phenol toxicity; Anti-HIV activity; Carcinogenicity
Cliquez ici pour obtenir le document / Click here to download the document pdf.
R. Dubois, P. Maison-Blanche, B. Quenet, G. Dreyfus
Automatic ECG wave extraction in long-term recordings using Gaussian mesa function models and nonlinear probability estimators
Computer Methods and Programs in Biomedicine, vol. 88, pp. 217-233 (2007)
Abstract
Keywords: machine-learning, neural network, orthogonal forward regression, adaptive signal processing, cardiac wave recognition, ECG.
Cliquez ici pour obtenir le document / Click here to download the document pdf.
[ Retour à la liste des publications / Back to the list of publications ]
F. Extramiana, A. Haggui, P. Maison-Blanche, R. Dubois, S. Takatsuki, P. Beaufils, A. Leenhardt
T-Wave Morphology Parameters Based on PrincipaL Component Analysis Reproducibility and Dependence
on T-Offset Position
Annals of Noninvasive Electrocardiology, vol. 12, pp. 354-363, 2007.
Abstract
Background: T-wave morphology parameters based on principal component analysis (PCA) are candidate to better understand the relation between QT prolongation and torsades de pointes.We aimed to assess the repeatability and to determine the influence of T-end position on PCA parameters.
Methods: Digital ECGs recorded from 30 subjects were used to assess short term (5 minutes), circadian and long-term (28 days) repeatability of PCA parameters. The T-end cursor position was moved backward and forward ( ± 8 ms) from its optimal position. We calculated QRS-T angle, PCA ratio, and T-wave residuum (TWR).
Results: At long-term evaluation, coefficients of variation were 11.3 ± 9.9%, 11.7 ± 7.1%, and 23.0 ± 22.0% for the QRS-T angle, PCA ratio, TWR, respectively. After moving the T-end cursor, repeatability was 0.42 ± 0.2%, 1.00 ± 1.04%, 4.0 ± 4.2% for the same PCA parameters.
Conclusions: T-wave morphology parameters based on PCA are reproducible with the exception of TWR and QRS-T angle. In addition, PCA is robust, showing only little dependence on T-end cursor position. These data should be taken into account for safety pharmacology trials.
Cliquez ici pour obtenir le document / Click here to download the document pdf.
Y. Takahashi, M. D. O'Neill, M. Hocini, R. Dubois, S. Matsuo, S. Knecht, S. Mahapatra, K.-T. Lim, P. Jais, A. Jonsson, F. Sacher, P. Sanders, T. Rostock, P. Bordachar, J. Clementy, G. J. Klein, and M. Haissaguerre,
Characterization of Electrograms Associated With Termination of Chronic Atrial Fibrillation by Catheter Ablation
Journal of the American College of Cardiology, vol. 51, pp. 1003-1010 (2008)
Abstract
Objectives: This study sought to determine the characteristics of atrial electrograms predictive of slowing or termination of atrial ?brillation (AF) during ablation of chronic AF.
Background: There is growing recognition of a role for electrogram-based ablation.
Methods: Forty consecutive patients (34 male, 59 ? 10 years) undergoing ablation for chronic AF persisting for a median of 12 months (range 1 to 84 months) were included. After pulmonary vein isolation and roof line ablation, electrogram-based ablation was performed in the left atrium and coronary sinus. Targeted electrograms were acquired in a 4-s window and characterized by: 1) percentage of continuous electrical activity; 2) bipolar voltage; 3) dominant frequency; 4) fractionation index; 5) mean absolute value of derivatives of electrograms; 6) local cycle length; and 7) presence of a temporal gradient of activation. Electrogram characteristics at favorable ablation regions, de?ned as those associated with slowing (a ?6-ms increase in AF cycle length) or termination of AF were compared with those at unfavorable regions.
Results: The AF was terminated by electrogram-based ablation in 29 patients (73%) after targeting a total of 171 regions. Ablation at 37 (22%) of these regions was followed by AF slowing, and at 29 (17%) by AF termination. The percentage of continuous electrical activity and the presence of a temporal gradient of activation were independent predictors of favorable ablation regions (p ? 0.016 and p ? 0.038, respectively). Other electrogram characteristics at favorable ablation regions were not signi?cantly different from those at unfavorable ablation regions.
Conclusions: Catheter ablation at sites displaying a greater percentage of continuous activity or a temporal activation gradient is associated with slowing or termination of chronic AF.
Cliquez ici pour obtenir le document / Click here to download the document pdf.
[ Retour à la liste des publications / Back to the list of publications ]
C. Faur, A. Cougnaud, G. Dreyfus and P. Le Cloirec
Modelling the breakthrough of activated carbon filters by pesticides in surface waters with static and recurrent neural networks
Chemical Engineering Journal, vol. 145, pp. 7–15 (2008).
Cliquez ici pour obtenir le document / Click here to download the document pdf.
S. Gazut, J. M. Martinez, G. Dreyfus, Y. Oussar
Towards The Optimal Design of Numerical Experiments
IEEE Trans. on Neural Networks, vol. 19, pp. 874 - 882 (2008).
Abstract
This paper addresses the problem of the optimal design of numerical experiments for the construction of nonlinear surrogate models. We describe a new method, called learner disagreement from experiment resampling (LDR), which borrows ideas from active learning and from resampling methods: the analysis of the divergence of the predictions provided by a pop-ulation of models,constructed by resampling,allows an iterative determination of the point of input space, where a numerical experiment should be performed in order to improve the accuracy of the predictor. The LDR method is illustrated on neural network models with bootstrap resampling, and on orthogonal polynomials with leave-one-out resampling. Other methods of experimental design such as random selection and optimal selection are investigated on the same benchmark problems.
Cliquez ici pour obtenir le document / Click here to download the document pdf.
[ Retour à la liste des publications / Back to the list of publications ]
Fabio Badilini, Martino Vaglio, Rémi Dubois, Pierre Roussel, Nenad Sarapa, Isabelle Denjoy, Fabrice Extramiana, Pierre Maison-Blanche
Automatic analysis of cardiac repolarization morphology using Gaussian mesa function modeling
Journal of Electrocardiology, vol. 41, pp. 588–594 (2008)
Abstract
A novel fully automated method for wave identification and extraction from electrocardiogram (ECG) waveforms is presented. This approach implements the combined use of a new machine-learning algorithm and of specified parameterized functions called Gaussian mesa functions (GMFs). Individual cardiac cycle waveforms are broken up into GMFs using a generalized orthogonal forward regression algorithm; each individual GMF is subsequently identified (wave labeling) and analyzed for feature and morphologic extraction. The GMF associated with the repolarization waveform of the main vector lead, based on principal components analysis, was analyzed, and a set of morphologic parameters were derived under 2 experimental settings: first, in 100 digital 12-lead ECG Holter recordings acquired during three 24-hour periods (baseline and after
160 and 320 mg of sotalol) from 38 healthy subjects; second, in drug-free 12-lead resting ECGs from 100 genotyped long QT syndrome (LQTS) patients (50 each with LQT1 and LQT2). QT- interval duration was measured using an on-screen method applied to the global representative beats and reviewed by a senior cardiologist. QTci (individual correction) was used for analysis. All parameters in the sotalol test showed highly significant differences between the time of peak plasma concentration (Tmax) and baseline ECGs; however, the dynamic pattern of individual parameters followed different patterns. The LQTS test confirmed the results of the sotalol test, showing that GMF-based repolarization parameters were strongly modified as compared with healthy controls. In particular, T-wave width and descending phase of repolarization were more prolonged in LQT2 compared to LQT1.
Cliquez ici pour obtenir le document / Click here to download the document pdf.
[ Retour à la liste des publications / Back to the list of publications ]
A. Magon de la Villehuchet, M. Brack, G. Dreyfus, Y. Oussar, D. Bonnefont-Rousselot, M.J. Chapman, A. Kontush
A machine learning approach to the prediction of oxidative stress in chronic inflammatory disease
Redox Reports, vol. 14, pp. 23-33 (2009).
Abstract
Oxidative stress is implicated in the development of a wide range of chronic human diseases, ranging from cardiovascular to neurodegenerative and inflammatory disorders. As oxidative stress results from a complex cascade of biochemical reactions, its quantitative prediction remains incomplete. Here we describe a machine-learning approach to the prediction of levels of oxidative stress in human subjects. From a database of biochemical analyses of oxidative stress biomarkers in blood, plasma and urine, nonlinear models have been designed, with a statistical methodology that includes variable selection, model training and model selection. Our data demonstrate that, despite a large inter- and intra-individual variability, levels of biomarkers of oxidative damage in biological fluids can be predicted quantitatively from measured concentrations of a limited number of exogenous and endogenous antioxidants.
Keywords: machine learning, neural networks, training, model selection, variable selection, oxidative stress, antioxidants, biological markers
Cliquez ici pour obtenir le document / Click here to download the document pdf.
[ Retour à la liste des publications / Back to the list of publications ]
G. Horcholle-Bossavit, B. Quenet,
Neural model of frog ventilatory rhythmogenesis,
Biosystems vol. 94, pp. 35-43 (2009).
Abstract
In the adult frog respiratory system, periods of rhythmic movements of the buccal floor are interspersed by lung ventilation episodes. The ventilatory activity results from the interaction of two hypothesized oscillators in the brainstem. Here, we model these oscillators with two coupled neural networks, whose co-activation results in the emergence of new dynamics. One of the networks is built with “loop chains”
of excitatory and inhibitory neurones producing periodic activities. We define two groups of excitatory neurones whose oscillatory antiphasic sums of activities represent output signals as possible motor commands towards antagonist buccal muscles. The other oscillator is a small network with a self-modulated excitatory input to an excitatory neurone whose episodic firings synchronise some neurones of the first network chains. When this oscillator is silent, the output signals exhibit only regular oscillations, and, when active, the synchronisation process reconfigures the output signals whose new features are representative of lung ventilation motor patterns. The biological interest of this formal model is illustrated by the persistence of the relevant dynamical features when perturbations are introduced in the model, i.e. dynamic noises and architecture modifications. The implementation of the networks with clock-driven continuous time neurones provides simulations with physiological time scales.
Cliquez ici pour obtenir le document / Click here to download the document pdf.
[ Retour à la liste des publications / Back to the list of publications ]
F. Extramiana, R. Dubois, M. Vaglio, P. Roussel, G. Dreyfus, F. Badilini, A. Leenhardt, P Maison-Blanche
The time course of new T-wave ECG descriptors following single and double dose administration of Sotalol in healthy subjects
Annals of Noninvasive Electrocardiology, vol. 15, pp. 26 - 35 (2010).
Abstract
Introduction: The aim of the study was to assess the time course effect of IKr blockade on ECG biomarkers of ventricular repolarization and to evaluate the accuracy of a fully automatic approach for QT duration evaluation.
Methods: 12-lead digital ECG Holter were recorded in 38 healthy subjects (27 males, mean age=27.4±8.0 years) on baseline conditions (day 0) and after administration of 160 mg (day 1) and 320 mg (day 2) of d-l Sotalol. For each 24-hour period and each subject, ECGs were extracted every 10 minutes during the 4-hour period following drug dosage. Ventricular repolarization was characterized using 3 biomarker categories: conventional ECG time intervals, Principal Component Analysis (PCA) analysis on the T-wave, and fully automatic biomarkers computed from a mathematical model of the T-wave.
Results: QT interval was significantly prolonged starting 1h20 minutes after drug dosing with 160 mg and 1h 10 minutes after drug dosing with 320 mg. PCA ventricular repolarization parameters sotalol-induced changes were delayed (>3 hours). After sotalol dosing, the early phase of the T-wave changed earlier than the late phase prolongation. Globally, the modeled surrogate QT paralleled manual QT changes. The duration of manual QT and automatic surrogate QT were strongly correlated (R²=0.92, p<0.001). The Bland & Altman plot revealed a non-stationary systematic bias (bias =26.5 ms ±1.96*SD =16 ms).
Conclusions: Changes in different ECG biomarkers of ventricular repolarization display different kinetics after administration of a potent potassium channel blocker. These differences need to be taken into account when designing ventricular repolarization ECG studies.
Cliquez ici pour obtenir le document / Click here to download the document pdf.
X. Johnson, G. Vandystadt, S. Bujaldon, F.A. Wollman, R. Dubois, P. Roussel, J. Alric, D. Béal
A new setup for in vivo fluorescence imaging of photosynthetic activity
Potosynthesis Research, vol. 102, pp. 85 - 93 (2009)..
Abstract
We describe a new imaging setup able to assess in vivo photosynthetic activity. The system specifically measures time-resolved chlorophyll fluorescence in response to light. It is composed of a fast digital camera equipped with a wide-angle lens for the analysis of samples up to 10 9 10 cm, i.e. entire plants or petri dishes. In the choice of CCD, we have opted for a 12-bits high frame rate [150 fps (frames per second)] at the expense of definition (640 9 480 pixels). Although the choice of digital camera is always a compromise between these two related features, we have designed a flexible system allowing the fast sampling of images (down to 100 ls) with a maximum spatial resolution. This image readout system, synchronized with actinic light and saturating pulses, allows a precise determination of F0 and FM, which is required to monitor PSII activity. This new imaging system, together with image processing techniques, is useful to investigate the heterogeneity of photosynthetic activity within leaves or to screen large numbers of unicellular algal mutant colonies to identify those with subtle changes in photosynthetic electron flow.
Cliquez ici pour obtenir le document / Click here to download the document pdf.
B. Denby, T. Schultz, K. Honda, T. Hueber, J.M. Gilbert, J.S. Brumberg,
Silent speech interfaces,
Speech Communication, vol. 52, pp. 270 - 287 (2010).
Abstract
This article presents a segmental vocoder driven by ultrasound and optical images of the tongue and lips for a “silent speech interface” application, usable either by a laryngectomized patient or for silent communication. The system is built around an audio-visual dictionary which associates visual to acoustic observations for each phonetic class. Visual features are extracted from ultrasound and optical video sequences of the tongue and lips using a PCA-based image coding technique. Visual observations of each phonetic class are modeled by continuous HMMs. The system then combines a phone recognition stage with corpus-based synthesis. In the recognition stage, the visual HMMs are used to identify phonetic targets in a sequence of visual features. In the synthesis stage, these phonetic targets constrain the dictionary search for the sequence of diphones that maximizes similarity to the input test data in the visual space, subject to a concatenation cost in the acoustic domain. A prosody template is extracted from the training corpus, and the final speech waveform is generated using “Harmonic plus Noise Model” concatenative synthesis techniques. Experimental results are based on an audiovisual database containing one hour of continuous speech from each of two speakers.
Cliquez ici pour obtenir le document / Click here to download the document pdf.
T. Hueber, E. Benaroya, G. Chollet, B. Denby, G. Dreyfus, M. Stone
Development of a silent speech interface driven by ultrasound and optical images of the tongue and lips
Speech Communication, vol. 52, pp. 288 - 300 (2010).
Abstract
This article presents a segmental vocoder driven by ultrasound and optical images (standard CCD camera) of the tongue and lips for a “silent speech interface” application, usable either by a laryngectomized patient or for silent communication. The system is built around an audio-visual dictionary which associates visual to acoustic observations for each phonetic class. Visual features are extracted from ultrasound images of the tongue and from video images of the lips using a PCA-based image coding technique. Visual observations of each phonetic class are modeled by continuous HMMs. The system then combines a phone recognition stage with corpus-based synthesis. In the recognition stage, the visual HMMs are used to identify phonetic targets in a sequence of visual features. In the synthesis stage, these phonetic targets constrain the dictionary search for the sequence of diphones that maximizes similarity to the input test data in the visual space, subject to a concatenation cost in the acoustic domain. A prosody template is extracted from the training corpus, and the final speech waveform is generated using “Harmonic plus Noise Model” concatenative synthesis techniques. Experimental results are based on an audiovisual database containing one hour of continuous speech from each of two speakers.
Keywords: silent speech, ultrasound, corpus-based speech synthesis, visual phone recognition
Cliquez ici pour obtenir le document / Click here to download the document pdf.
M. Toukourou, A. Johannet, G. Dreyfus, P. A. Ayral
Rainfall-runoff modeling of flash floods in the absence of rainfall forecasts: the case of "Cévenol flash floods"
Applied Intelligence, published electronically 20. February 2010, doi 10.1007/s10489-010-0210-y.
Abstract
“Cévenol flash floods” are famous in the field of hydrology, because they are archetypical of flash floods that occur in populated areas, thereby causing heavy damages and casualties. As a consequence, their prediction has become a stimulating challenge to designers of mathematical models, whether physics based or machine learning based. Because current, state-of-the-art hydrological models have difficulty performing forecasts in the absence of rainfall previsions, new approaches are necessary. In the present paper, we show that an appropriate model selection methodology, applied to neural network models, provides reliable two-hour ahead flood forecasts.
Cliquez ici pour obtenir le document / Click here to download the document pdf
B. Denby, Y. Oussar, I. Ahriz, G. Dreyfus,Abstract
GSM trace mobile measurements are used to study indoor handset localization in an urban apartment setting. Nearest-neighbor, Support Vector Machine (SVM), and Gaussian Process classifiers are compared. A linear SVM is found to provide mean room-level classification efficiency near 100%, but only when the full set of GSM carriers is used. To our knowledge, this is the first study to use fingerprints containing all GSM carriers, and the first to suggest that GSM could be useful for very high-performance indoor localization.
Cliquez ici pour obtenir le document / Click here to download the document pdf
[ Retour à la liste des publications / Back to the list of publications ]
P. Milpied, R. Dubois, P. Roussel, C. Henry, G. Dreyfus
Arrhythmia Discrimination in Implantable Cardioverter Defibrillators using Support Vector Machines applied to a New Representation of Electrograms
IEEE Transactions on Biomedical Engineering, vol. 56, pp. 1797 - 1803 (2011).
Abstract
Arrhythmia classification remains a major challenge for appropriate therapy delivery in Implantable Cardioverter Defibrillators (ICDs). The purpose of this paper is to present a new algorithm for arrhythmia discrimination based on a statistical classification by Support Vector Machines of a novel 2-D representation of electrograms, named Spatial Projection Of Tachycardia electrograms (SPOT). SPOT-based discrimination algorithm provided sensitivity and specificity of 98.8% and 91.3% respectively, on a test database. A simplified version of the algorithm is also presented, which can be directly implemented in the ICD.
Index Terms—Implantable Cardioverter Defibrillators, arrhythmias, inappropriate therapy, electrogram morphology, Support Vector Machines, SPOTCliquez ici pour obtenir le document / Click here to download the document pdf
J. Dauwels, K. Srinivasan, M. Ramasubba Reddy, T. Musha, F. Vialatte, C. Latchoumane, J. Jeong, A. Cichocki
Slowing and Loss of Complexity in Alzheimer's EEG: Two Sides of the Same Coin?
International Journal of Alzheimer’s Disease, doi: 0.4061/2011/539621 (2011).
Abstract
Medical lstudies have shown that EEG of Alzheimer’s disease (AD) patients is “slower” (i.e., contains more low-frequency power) and is less complex compared to age-matched healthy subjects. The relation between those two phenomena has not yet been studied, and they are often silently assumed to be independent. In this paper, it is shown that both phenomena are strongly related. Strong correlation between slowing and loss of complexity is observed in two independent EEG datasets: (1) EEG of pre-dementia patients (a.k.a. Mild Cognitive Impairment; MCI) and control subjects; (2) EEG of mild AD patients and control subjects. The two datasets are from different patients, different hospitals, and obtained through different recording systems. The paper also investigates the potentia lof EEG slowing and loss of EEG complexity as indicators of AD onset. In particular, relative power and complexity measures are used as features to classify the MCI and MiAD patients vs. age-matched control subjects; linear and quadratic discriminant analysis and support vector machines are applied as classi?ers. When combined with two synchrony measures (Granger causality and stochastic event synchrony), classi?cation rates of 83% (MCI) and 98% (MiAD) are obtained. By including the compression ratios as features, slightly better classi?cation rates are obtained than with relative power and synchrony measures alone.
Index Terms — Alzheimer’s disease (AD), mild cognitive impairment (MCI), electroencephalogram (EEG), compression ratio, relative power, Granger causality, stochastic event synchrony
Cliquez ici pour obtenir le document / Click here to download the document pdf
F. Vialatte, J. Dauwels, M. Maurice, T. Musha, A. Cichocki
Improving the specificity of EEG for Diagnosing Alzheimer’s disease
International Journal of Alzheimer’s Disease, doi: 0.4061/2011/259069 (2011).
Abstract
Objective: EEG has great potential as a cost-effective screening tool for Alzheimer’s disease (AD). However, the specificity of EEG is not yet sufficient to be used in clinical practice. In an earlier study, we presented preliminary results suggesting improved specificity of EEG to early stages of Alzheimer’s disease. The key to this improvement is a new method for extracting sparse oscillatory events from EEG signals in the time-frequency domain. Here we provide a more detailed analysis, demonstrating improved EEG specificity for clinical screening of MCI (mild cognitive impairment) patients. Methods: EEG data was recorded of MCI patients and age-matched control subjects, in rest condition with eyes closed. EEG frequency bands of interest were ? (3.5-7.5 Hz), ?1 (7.5-9.5 Hz), ?2 (9.5-12.5 Hz), and ? (12.5-25 Hz). The EEG signals were transformed in the time-frequency domain using complex Morlet wavelets; the resulting time-frequency maps are represented by sparse bump models.
Results: Enhanced EEG power in the ? range is more easily detected through sparse bump modeling; this phenomenon explains the improved EEG specificity obtained in our previous studies.
Conclusions: Sparse bump modeling yields informative features in EEG signal. These features increase the specificity of EEG for diagnosing AD.
Keywords: Alzheimer’s disease, EEG, screening, time-frequency, sparse bump modeling
Cliquez ici pour obtenir le document / Click here to download the document pdf
J. Dauwels, F. Vialatte, A. Cichocki
Quantifying Statistical Interdependence, Part III: N>2 Point Processes
Neural Computation vol. 24, pp. 408 – 454 (2012).
Abstract
Stochastic event synchrony (SES) is a recently proposed family of similarity measures. First, “events” are extracted from the given signals; next, one tries to align events across the different time series. The better the alignment, the more similar the N time series are considered to be. The similarity measures quantify the reliability of the events (the fraction of “nonaligned” events) and the timing precision. Sofar, SES has been developed for pairs of one-dimensional (Part I) and multidimensional (Part II) point processes. In this letter (PartI II), SES is extended from pairs of signals to N > 2 signals.The alignment and SES parametersare again determined through statistical inference, more speci?cally, by alternating two steps: (1) estimating the SES parameters from a given alignment and (2), with the resulting estimates, re?ning the alignment. The SES parameters are computed by maximum a posteriori (MAP) estimation (step1), in analogy to the pairwise case. The alignment (step2) is solved by linear integer programming. In order to test the robustness and reliability of the proposed N-variate SES method, it is ?rst applied to synthetic data. We show that N-variate SES results in more reliable estimates than bivariate SES. Next N-variate SES is applied to two problems in neuroscience: to quantify the ?ring reliability of Morris-Lecar neurons and to detect anomalies in EEG synchrony of patients with mild cognitive impairment. Those problems were also considered in Parts I and II, respectively. In both cases, the N-variate SES approach yields a more detailed analysis.
Cliquez ici pour obtenir le document / Click here to download the document pdf
Abstract
Oxidative stress is involved in chronic and acute pathologies: cardiovascular, neurodegenerative, neoplas-tic, inflammatory and infectious diseases. Clinical trials focused on prevention of cardiovascular and neoplastic diseases involving antioxidant supplemen- tation have however provided predominantly nega- tive obserations in large-scale studies. Screening of patient cohorts to assess baseline oxidative stress on the basis of a biomarker profile is decisive but lacking. For the first time, we evaluated the level of oxidative stress, testing more than 10 established biomarkers, in a comprehensive initial survey of 617 patients dis- playing chronic human pathologies. Multiple disease- specific abnormalities were identified in plasma, whole blood and/or urine. This is the case for vitamins and oligo elements, vitamin C, vitamin E, ?-carotene, se- lenium, zinc and copper; endogenous antioxidants such as reduced and oxidised glutathione, thiols, urate, and glutathione peroxidase activity, and a biomarker of oxidative DNA damage (8-hydroxy-2’-deoxyguanosine). The distinct biomarker profiles suggest the involvment of multiple forms of oxidative insults which are n some way partially specific to each pathological condition. This finding is in favor of the determina- tion of an integrated score to combine contributions of distinct biomarkers, in order to screen patients presenting elevated levels of oxidative stress.
Keywords: Vitamin C; Vitamin E; bêta-Carotene; Glutathione; Thiols; Urate; 8-Hydroxy-2’-Deoxyguanosine
Cliquez ici pour obtenir le document / Click here to download the document pdf
Y. Tomita, F. Vialatte, G. Dreyfus, Y. Mitsukura, H. Bakardjian, A. Cichocki
Bimodal BCI using simultaneously NIRS and EEG
IEEE Transactions on Biomedical Engineering, vol. 61, pp. 1274 - 1284 (2014).
Abstract
Although non invasive brain–computer interfaces (BCI) based on electroencephalographic (EEG) signals have been studied increasingly over the recent decades, their performance is still limited in twoi mportant aspects. First, the dif?culty of performing a reliable detection of BCI commands increases when EEG epoch length decreases, which makes high information transfer rates dif?cult to achieve. Second, the BCI system often misclassi?es the EEG signals as commands, although the subject is not performing any task. In order to circumvent thesel imitations, the hemodynamic ?uctuations in the brain during stimulation with steady-state visual evoked potentials (SSVEP) were measured using near-infrared spectroscopy (NIRS) simultaneously with EEG. BCI commands were estimated based on responses to a ?ickering checkerboard (ON-period). Furthermore, an “idle” command was generated from the signal recorded by the NIRS system when the checkerboard was not ?ickering (OFF-period). The joint use of EEG and NIRS was shown to improve the SSVEP classi?cation. For 13 subjects, the relative improvement in error rates obtained by using the NIRS signal, for nine classes including the “idle” mode, ranged from 85% to 53%, when the epoch length increase from 3 to 12s. These results were obtained from only one EEG and one NIRS channel. The proposed bimodal NIRS–EEG approach, including detection of the idle mode, may make current BCI systems faster and more reliable.
IndexTerms—Brain–computerinterface(BCI),bimodal,simul-taneouselectroencephalographic(EEG)andnear-infraredspec-troscopy(NIRS),steady-statevisualevokedpotentials(SSVEP).
Cliquez ici pour obtenir le document / Click here to download the document pdf
F. Dioury, A. Duprat, G. Dreyfus, C. Ferroud, J. Cossy
QSPR Prediction of the Stability Constants of Gadolinium(III) Complexes for Magnetic Resonance Imaging
Journal of Chemical Information and Modeling, vol. 54, pp. 2718?2731 (2014).
Abstract
Gadolinium(III) complexes constitute the largest class of compounds used as contrast agents for Magnetic Resonance Imaging (MRI). A quantitative structure-property relationship (QSPR) machine-learning based method is applied to predict the thermodynamic stability constants of these complexes (log KGdL), a property commonly associated with the toxicity of such organometallic pharmaceuticals. In this approach, the log KGdL value of each complex is predicted by a graph machine, a combination of parameterized functions that encodes the 2D structure of the ligand. The efficiency of the predictive model is estimated on an independent test set; in addition, the method is shown to be effective (i) for estimating the stability constants of uncharacterized, newly synthesized polyamino-poly carboxylic compounds, and (ii) for providing independent log KGdL estimations for complexants for which conflicting or questionable experimental data were reported. The exhaustive database of log KGdL values for 158 complexants, reported for potential application as contrast agents for MRI and used in the present study, is available in the Supporting Information (122 primary literature sources).
Cliquez ici pour obtenir le document / Click here to download the document pdf
K. Hiyoshi-Taniguchi, M. Kawasaki, Y. Yokota, H. Bakardjian, H. Fukuyama, A. Cichocki, F.B. Vialatte
EEG Correlates of Voice and Face Emotional Judgments in the Human Brain
Cognitive Computation vol. 7, p. 11 - 19 (2015).
Abstract
The purpose of this study is to clarify the neural correlates of human emotional judgment. This study aimed to induce a controlled perturbation in the emotional system of the brain by multimodal stimuli, and to investigate whether such emotional stimuli could induce reproducible and consistent changes in electroencephalography (EEG) signals. We exposed 12 subjects to auditory, visual, or combined audio–visual stimuli. Audio stimuli consisted of voice recordings of the Japanese word “arigato” (thank you) pronounced with three different intonations (angry—A, happy—H or neutral—N). Visual stimuli consisted of faces of women expressing the same emotional valences (A, H or N). Audio–visual stimuli were composed using either congruent combinations of faces and voices (e.g., H × H) or noncongruent combinations (e.g., A × H). The data were collected using an EEG system, and analysis was performed by computing the topographic distributions of EEG signals in the theta, alpha, and beta frequency ranges. We compared the conditions stimuli (A, H or N), and congruent versus noncongruent. Topographic maps of EEG power differed between those conditions. The obtained results demonstrate that EEG could be used as a tool to investigate emotional valence and discriminate various emotions.
X. Zhang, F.B. Vialatte, C. Chen, A. Rathi, G. Dreyfus
Embedded Implementation of Second-Order Blind Identification (SOBI) for Real-Time Applications in Neuroscience
Cognitive Computation vol. 7, p. 56 - 63 (2015).
Abstract
Blind source separation (BSS) is an effective and powerful tool for signal processing and artifact removal in electroencephalographic signals. For real-time applications such as brain–computer interfaces, cognitive neuroscience or clinical neuromonitoring, it is of prime importance that BSS is effectively performed in real time. In order to improve in terms of speed considering the optimal parallelism environment that hardware provides, we build a high-level hardware/software co-simulation based on MATLAB/Simulink for BSS application. To illustrate our approach, we implement the most critical parts of the second-order blind identification algorithm with a fixed-point algorithm on a commercial field-programmable gate array development kit. The results obtained show that co-simulation environment reduces the computation time from 1.9 s to 12.8 ns and thus has great potential for real-time applications.
J. Solé-Casals, F.B. Vialatte, J. Dauwels
Alternative Techniques of Neural Signal Processing in Neuroengineering
Cognitive Computation vol. 7, p. 1 - 2 (2015).
Abstract
Blind source separation (BSS) is an effective and powerful tool for signal processing and artifact removal in electroencephalographic signals. For real-time applications such as brain–computer interfaces, cognitive neuroscience or clinical neuromonitoring, it is of prime importance that BSS is effectively performed in real time. In order to improve in terms of speed considering the optimal parallelism environment that hardware provides, we build a high-level hardware/software co-simulation based on MATLAB/Simulink for BSS application. To illustrate our approach, we implement the most critical parts of the second-order blind identification algorithm with a fixed-point algorithm on a commercial field-programmable gate array development kit. The results obtained show that co-simulation environment reduces the computation time from 1.9 s to 12.8 ns and thus has great potential for real-time applications.
Cliquez ici pour obtenir le document / Click here to download the document pdf
[ SIGMA-ESPCI , SIGMA-ESPCI (English Version) | ESPCI ]