Laboratoire SIGMA - SIGnaux, Modèles, Apprentissage statistique
LIVRES, ACTES DE CONFERENCES : RESUMES ET TEXTES (format .ps et .ps.gz) / BOOKS, CONFERENCE PROCEEDINGS : ABSTRACTS AND TEXTS (.ps and .ps.gz formats)
[ Retour à la liste des publications / Back to the list of publications ]
G. Dreyfus, O. Macchi, S. Marcos, O. Nerrand, L. Personnaz, P. Roussel-Ragot, D. Urbani, C. Vignat
Adaptive Training of Feedback Neural Networks for Non-Linear Adaptive Filtering,
Neural Networks for Signal Processing II, 550 (IEEE, 1992).
Abstract
The paper proposes a general framework which encompasses the training of neural networks and the adaptation of filters. It is shown that neural networks can be considered as general non-linear filters which can be trained adaptively, i.e. which can undergo continual training. A unified view of gradient-based training algorithms for feedback networks is proposed, which gives rise to new algorithms. The use of some of these algorithms is illustrated by examples of non-linear adaptive filtering and process identification.
Cliquez ici pour obtenir le document / Click here to download the document postscript (or ps.gz) or .pdf.
[ Retour à la liste des publications / Back to the list of publications ]
O. Nerrand, L. Personnaz G. Dreyfus
Non-Linear Recursive Identification and Control by Neural Networks: a General Framework,
Proc. European Control Conference ECC'93, vol. 1, pp. 93-98 (Groningen, 1993).
Abstract
The development of engineering applications of neural networks makes it necessary to clarify the similarities and differences between the concepts and methods developed for neural networks and those used in more classical fields such as filtering and control. In previous papers [Nerrand et al. 1993], [Marcos et al. 1993], the relationships between non-linear adaptive filters and neural networks have been investigated, and a general framework has been introduced, which encompasses the recursive training of neural networks and the adaptation of non-linear filters. Out of this approach, three new families of training algorithms for feedback networks emerged; algorithms used routinely in adaptive filtering and in the training of neural networks were shown to be specific cases of this general approach.
The adaptive identification of non-linear processes is a natural field of application of these algorithms. The first part of the paper will be devoted to a short survey of the recursive training of feedback (also termed recurrent) discrete-time neural networks for non-linear identification; the algorithms presented in that section can be used either for adaptive or for non-adaptive systems.
Pursuing our effort along the same lines,we show that algorithms for the adaptive control of non-linear processes by neural networks can be derived from the above approach. However, process control has its own goals and constraints: therefore, the algorithms must be tuned to such specificities. The second part of the paper is devoted to the presentation of these algorithms, which are illustrated in detail in the third part.
Cliquez ici pour obtenir le document / Click here to download the document postscript (or ps.gz) or .pdf.
[ Retour à la liste des publications / Back to the list of publications ]
C. Linster, D. Marsan, C. Masson, M. Kerszberg, G. Dreyfus, L. Personnaz
A Formal Model of the Insect Olfactory Macroglomerulus: Simulations and Analytical Results,
Neural Information Processing Systems 5, S.J. Hanson, J.D.Cowan, C. Lee Giles, eds, pp 1022-1029 (Morgan Kaufmann Publishers, 1993).
Abstract
It is known from biological data that the response patterns of interneurons in the olfactory macroglomerulus (MGC) of insects are of central importance for the coding of the olfactory signal. We propose an analytically tractable model of the MGC which allows us to relate the distribution of response patterns to the architecture of the network
Cliquez ici pour obtenir le document / Click here to download the document postscript (or ps.gz) or .pdf.
I. Rivals, L. Personnaz, G. Dreyfus, D. Canas
Real-time Control of an Autonomous Vehicle : a Neural Network Approach to the Path Following Problem,
5th International Conference on Neural Networks and their Applications (Nîmes , 1993).
Abstract
A neural-network based approach to the control of non-linear dynamical systems such as wheeled mobile robots is presented. A general framework for the training of neural controllers is outlined, and applied to the lateral control of a vehicle for the path following and trajectory servoing problems. Simulation as well as experimental results on a four-wheel drive vehicle equipped with actuators and sensors are shown.
Cliquez ici pour obtenir le document / Click here to download the document postscript (or ps.gz) or .pdf.
D. Urbani, P. Roussel-Ragot, L. Personnaz, G. Dreyfus
The Selection of Neural Models of Non-linear Dynamical Systems by Statistical Tests,
Neural Networks for Signal Processing IV, 229-237 (IEEE, 1993).
Abstract
A procedure for the selection of neural models of dynamical processes is presented. It uses statistical tests at various levels of model reduction, in order to provide optimal tradeoffs between accuracy and parsimony. The efficiency of the method is illustrated by the modeling of a highly non-linear NARX process.
Cliquez ici pour obtenir le document / Click here to download the document pdf
I. Rivals, L. Personnaz, G. Dreyfus, D. Canas
Modeling and Control of a Wheeled Mobile Robot Using Recurrent Neural Networks : an Application to the Path Following Problem
International Workshop on Neural Networks for Computing (1994)
Abstract
We adress the problem of controlling an autonomous vehicle using recurrent neural networks for the path following problem. Our testbed is the full-scale outdoor robot REMI (Robot Evaluator for Mobile Investigations), a standard four-wheel-drive truck that has been equipped with actuators for the steering wheel, the throttle, the brakes and the gear, and with the sensors needed for navigation (an inertial dead-reckoning unit and an odometer).
Wheeled mobile robots are highly non-linear dynamical systems, their kinematics involving geometrical non-linearities and their actuators introducing dynamical non-linearities, such as saturations and the non-linear dynamics of the thermal motor. Thus, the identification of such processes requires non-linear modeling methods. We therefore start by presenting the identification of REMI's dynamical behaviour using recurrent neural networks.
We then outline a general framework for the training of neural controllers for non-linear dynamical systems, using a reference model, a model of the process to be controlled, and a neural controller implementing a state-feedback control law. We apply the principles presented in this framework to the control of the vehicle for the path following problem. The problem is to steer the vehicle along a known reference path with a predetermined velocity profile. We express the path following problem in terms of controlling the vehicle orientation, the velocity being imposed either by a human operator or by a neural speed-controller which is designed and trained independently. The aim is to derive a state-feedback control law for the steering command so as to (i) have the vehicle rally and follow the path and (ii) have the orientation of the vehicle tangent to the path when the vehicle follows the path.
Simulation results illustrating the training and the behaviour of the control system designed are shown. This control system has also been successfully tested with the robot REMI on various trajectories, on smooth and rough terrain, and we present the corresponding real-time experimental results, as well as comparisons with the classical approaches previously used.
I. Rivals, D. Canas, L. Personnaz, G. Dreyfus
Modeling and Control of Mobile Robots and Intelligent Vehicles by Neural Networks,
IEEE Conference on Intelligent Vehicules (Paris , 1994).
Abstract
This paper introduces the four-wheel-drive vehicle REMI, a testbed developed by SAGEM for research purposes in mobile robotics and intelligent car systems. The motion control architecture of the robot is presented, with an emphasis on the guidance and piloting modules. The latter relies on neural network techniques, and the principles underlying its design are outlined. A robust neural control scheme using an internal model of the process is developed. Experimental results are presented.
Cliquez ici pour obtenir le document / Click here to download the document postscript (or ps.gz) or .pdf.
C. Linster, D. Marsan, C. Masson, M. Kerszberg
Odor processing in the bee: a preliminary study of the role of central input to the antennal lobe
In Advances in Neural Information Processing Systems 6, Cowan, J.D., Tesauro, G., Alspector, J. (eds), Morgan Kaufmann Publishers: 527-534 (1994).
Abstract
Based on precise anatomical data of the bee's olfactory system, we propose an investigation of the possible mechanisms of modulation and control between the two levels of olfactory information processing: the antennal lobe glomeruli and the mushroom bodies. We use simplified neurons, but a realistic architecture. As a first conclusion, we postulate that the feature extraction performed by the antennal lobe (glomeruli and interneurons) necessitates central input from the mushroom bodies for fine tuning. The central input thus facilitates the evolution from fuzzy olfactory images in the glomerular layer towards more focussed images upon odor presentation.
J.L. Ploix, G. Dreyfus
Knowledge-Based Neural Networks for the Modeling of Complex Industrial Systems
International Workshop on Neural Networks for Computing (1995).
Abstract
The modeling of complex systems, such as encountered in chemical plants for instance, typically requires solving systems of hundreds of coupled non-linear differential equations ; when used for such purposes as early fault detection, this must be performed very accurately in real-time, a task which usually defeats PC-type computers or workstations. We present a general methodology which capitalizes on the black-box abilities of neural networks, while retaining the wealth of mathematical knowledge which is very often available. Essentially, the method consists in designing a recurrent neural network obeying the recurrent difference equations, derived from the physics and chemistry of the process, which describe approximately the dynamics of the process, and in adding black-box networks intended to account for unmodeled dynamics. The resulting network is then trained with appropriate algorithms. This permits very accurate process simulation in real time. Industrial applications of the method will be presented.
I. Rivals, L. Personnaz
Black-box Modeling with State-Space Neural Networks
in: Neural Adaptive Control Technology R. Zbikowski and K.J. Hunt, eds. (World Scientific, 1995).
Abstract
Neural network black-box modeling is usually performed using input-output models and the corresponding input-output neural predictors. The goal of this paper is to show that there is an advantage in using state-space models,which are more general, and their corresponding state-space neural predictors. We recall the fundamentals of both niput-output and state-space black-box modeling, and show the state-space neural networks to be potentially more efficient and more parsimonious than their conventional input-output counterparts. This is examplified on simulated processes as well as on a real one, the hydraulic actuator of a robot arm.
Cliquez ici pour obtenir le document / Click here to download the document postscript. (or ps.gz) or .pdf.
I. Rivals, L. Personnaz, G. Dreyfus
From Conventional Control to Robust Neural Control: a Methodology
International Workshop on Neural Networks for Computing (1995)
Abstract
Capitalizing on the universal approximation property, and on the "universality" of the training algorithms for recurrent networks that we introduced during the past years, we propose a general methodology for the design of robust "neural" control systems, which parallels the design methodology of conventional control systems. Based on two types of controllers, termed O-controllers (which impose a reference Output to the control system), and D-controllers (which impose a reference Dynamics to the control system), we introduce a very large family of control systems, relevant both to conventional and to "neural" control. We investigate in depth the "neural" implementation of the technique of Internal Model Control, which allows the design of control system with remarkable robustness with respect to modeling errors and to perturbations. Applications of this methodology to academic and to industrial problems illustrate the presentation.
D. Price, S. Knerr, L. Personnaz, G. Dreyfus
Pairwise Neural Network Classifiers with Probabilistic Outputs,
Neural Information Processing Systems 7 (1994).
Abstract
Multi-class classification problems can be efficiently solved by partitioning the original problem into sub-problems involving only two classes: for each pair of classes, a (potentially small) neural network is trained using only the data of these two classes. We show how to combine the outputs of the two-class neural networks in order to obtain posterior probabilities for the class decisions. The resulting probabilistic pairwise classifier is part of a handwriting recognition system which is currently applied to check reading. We present results on real world data bases and show that, from a practical point of view, these results compare favorably to other neural network approaches.
Cliquez ici pour obtenir le document / Click here to download the document postscript (or ps.gz) or .pdf.
I. Rivals, L. Personnaz, G. Dreyfus, J.L. Ploix
Modélisation, classification et commande par réseaux de neurones : principes fondamentaux, méthodologie de conception et illustrations industrielles,
Les réseaux de neurones pour la modélisation et la commande de procédés, J.P. Corriou, coordonnateur (Lavoisier Tec & Doc, 1995).
Résumé
Les réseaux de neurones connaissent depuis quelques années un succès croissant dans divers domaines des Sciences de l'Ingénieur ; celui du génie des procédés n'échappe pas à cette règle. Malheureusement, la littérature fourmille d'exemples où la mise en oeuvre des réseaux de neurones relève plus de la recette que d'une approche raisonnée. De plus, les connotations biologiques du terme "réseaux de neurones", et l'utilisation du terme d'"apprentissage", ont souvent introduit une grande confusion ; elles ont conduit à relier abusivement les réseaux de neurones à l'Intelligence Artificielle, alors qu'ils sont fondamentalement des outils statistiques. L'objectif de cet article est de montrer comment, à partir des notions fondamentales, il est possible d'aboutir à une véritable méthodologie de mise en oeuvre, notamment dans le cadre de la modélisation des processus. Nous montrons en particulier que, contrairement à une croyance répandue, les réseaux de neurones ne sont pas nécessairement des "boîtes noires" : bien au contraire, il est parfaitement possible, et même vivement recommandé, d'introduire dans le réseau de neurones, dès sa conception, toutes les connaissances mathématiques disponibles concernant le processus à modéliser ou à commander.
Cliquez ici pour obtenir le document / Click here to download the document postscript (or ps.gz) or .pdf.
B. Quenet, G. Dreyfus, C. Masson
Towards an Analytically Tractable Model of the Olfactory System
International Workshop on Neural Networks for Computing (1996). Invited presentation.
Abstract
We present an analytically tractable model for the formation of olfactory images in the glomerular system. The essential features of the model are the following:
- inhibition is dominant; this is expressed in the model by the fact that all inter- and intraglomerular connections are inhibitory;
- each family of receptors projects essentially onto a single glomerulus; this is expressed in the model by the fact that there is a one-to-one excitatory connection from receptors to glomeruli.
In addition, the activity of each glomerulus is modeled as that of a formal inhibitory neuron. For simplicity, all delays and all synaptic weights are assumed to be equal; synchronous dynamics is investigated.
This approach gives an understanding of the space of representation of the stimuli : we show that the system has the ability of extracting time-invariant key features of the input signal, thereby producing stable "images" at the glomerular level. This property results in an intrinsic robustness to the noise present in the signal. The existence of a Lyapunov function provides a tool for studying the dynamical behaviour of the system; the influence of synaptic noise is investigated.
J.L. Ploix, G. Dreyfus
Knowledge-based Neural Modeling: Principles and Industrial Applications,
ICANN'95 (Paris, 1995).
Abstract
A methodology for designing semi-physical neural models is presented. Starting from a mathematical model of the process, a recurrent neural network is constructed, and some of its weights are adjusted from the measurements by training. The application of this methodology to the modeling of an industrial distillation column, designed for early fault detection, is described.
Cliquez ici pour obtenir le document / Click here to download the document postscript (or ps.gz) or .pdf.
H. Stoppiglia, Y. Idan, G. Dreyfus
Neural-Network-Aided Portfolio Management,
ICANN'95 (Paris, 1995).
Abstract
The paper presents the design of an automated system assessing the risk of long-term investments. Although the problem is a relatively standard classification problem, it has specific features, especially as far as input selection is concerned. We show that the combination of "neural" and "standard" statistical methods allows us to obtain results similar to those obtained by a heuristic choice of descriptors, but in a more rigorous, principled and reliable fashion. The system is in actual routine use within a large French financial group.
Cliquez ici pour obtenir le document / Click here to download the document postscript (or ps.gz) or .pdf.
B. Quenet, G. Dreyfus, V. Cerny
The Effect of the Synaptic Noise on the Coding Properties of a Model of the Olfactory System: Analytical study and simulations
International Workshop on "Machines that Learn" (1997).
Abstract
The deterministic version of a model of the glomerular stage in the olfactory tract has revealed an interesting behaviour which consists in spontaneously extracting key features from fluctuating signals. In the model, binary neurons figure the glomeruli, each connected to another by inhibitory synapses. Receptor neurons send excitatory connections to the glomerular neurons according to a one-to-one correspondence. The model performs a mapping from the space of all possible activities of the receptors to the space of spatio-temporal patterns of the glomerular activities (glomerular images). The simplicity of the model, with all delays and all synaptic weights equal, allows an analytical approach which has already led to an in-depth understanding of its dynamics.
We have proved that the coding principles are conserved when synaptic noise is added; indeed, the limit probability vector of the glomerular activities tends to the cyclic steady states that minimize a Lyapunov function. We observe three noise regimes : (1) a low noise regime, where the only steady states are those that minimize the Lyapunov function, thereby enhancing the efficiency of the extraction of key features with respect to the noise-free regime; (2) a medium noise regime where the mean activity of the glomeruli comes close to the mean activity of the receptors themselves, and (3) a high noise regime, where all coding properties are blurred out. To summarize, a modulation of the synaptic noise may lead the model to perform a more or less sketchy representation of the receptor activities, the most efficient key feature extraction occurring in the low noise regime.
G. Dreyfus, Y. Idan
The Canonical Form of Nonlinear Discrete-time Models
International Workshop on "Machines that Learn" (1997).
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 any system 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.
Brigitte Quenet, Gérard Dreyfus, Pierre Roussel-Ragot et Vlado Cerny
Propriétés de codage d'un modèle de la voie olfactive : étude analytique, et simulations, de l'influence du bruit synaptique.
Colloque de la Société des Neurosciences, Bordeaux (1997).
Abstract
Un modèle des deux premiers étages de la voie olfactive, dont les propriétés de codage ont fait l'objet d'une analyse mathématique détaillée, a révélé, dans sa version déterministe, un comportement très intéressant : l'extraction spontanée de "facteurs clefs" (key features) dans un ensemble de signaux fluctuants.
Dans ce modèle, les unités glomérulaires sont figurées par des neurones binaires, interconnectés par des synapses inhibitrices. Les neurones récepteurs envoient leur information par des connexions excitatrices aux neurones glomérulaires. Le modèle réalise alors un codage des signaux d'entrée sous forme de cartes d'activité spatio-temporelles ('images glomérulaires'). La dynamique déterministe de ce modèle est maintenant très bien comprise.
Dans la présente communication, nous prouvons que les principes de codage sont conservés lorsque le bruit synaptique est pris en considération : en effet le vecteur limite des probabilités d'activité des glomérules définit une distribution des états de type Boltzmann, où les états les plus probables sont ceux qui minimisent une fonction de Lyapounov du système.
On observe l'émergence de trois régimes de bruit :
(1) un régime à bas bruit, où, en régime permanent, pratiquement tous les états d'activité glomérulaire sont ceux qui minimisent la fonction de Lyapounov ; ainsi, par l'élimination d'éventuels autres états stables, l'efficacité de l'extraction des facteurs clefs par rapport à la version déterministe du modèle est améliorée ;
(2) un régime à bruit moyen où l'activité glomérulaire moyenne se rapproche de l'activité des récepteurs ;
(3) un régime à bruit élevé, où toutes les propriétés de codage disparaissent.
Ainsi, dans ce modèle, une modulation de l'intensité du bruit synaptique peut amener un même système à assurer deux fonctions : soit une fonction d'extraction des facteurs clés (à bas bruit), soit une fonction de recopie de l'activité moyenne des récepteurs (à bruit plus élevé).
Une variante du modèle, où les non-linéarités sont décentralisées du niveau somatique au niveau synaptique, est présentée.
I. Rivals, L. Personnaz
Internal Model Control using Neural Networks
IEEE International Symposium on Industrial Electronics ISIE'96 (Warsaw, 1996).
Abstract
We propose a design procedure of neural Internal Model control systems for processes with
delay. We assume that a stable discrete-time neural model of the process is available. We show
that the design of a Model Reference controller for Internal Model control necessitates only the
training of the inverse of the model deprived from its delay, provided this inverse exists and is
stable. As the robustness properties intrinsic to Internal Model control systems are only obtained
if the inverse model is exact, it is also shown how to limit the effects of a possible inaccuracy of
the inverse model due to its training. Computer simulations illustrate the proposed design
procedure.
Cliquez ici pour obtenir le document / Click here to download the document postcript (or ps.gz).
JL. Ploix, G. Dreyfus
Early fault detection in a distillation column: an industrial application of knowledge-based neural modelling,
Neural Networks: Best Practice in Europe, B. Kappen, S. Gielen, eds, pp. 21-31 (World Scientific, 1997).
Abstract
One of the most widespread misconceptions about neural networks is the fact that they are "black boxes" which (i) do not make any use of prior knowledge of the process to be modelled, and (ii) cannot be "understood" by the expert of the process. We show that, on the contrary, neural networks can be used as "grey box models", and that the designer can take full advantage of the mathematical knowledge which may exist on the process. Using a knowledge-based neural model, we have been able to design a real-time distillation column simulator, implemented on a PC, which allows the early detection of faults. The neural network is a dynamic model (recurrent neural net) with 102 state variables, presumably the largest recurrent neural network ever trained and implemented for industrial purposes.
Cliquez ici pour obtenir le document / Click here to download the document postscript (or ps.gz) or .pdf.
H. Stoppiglia, G. Dreyfus
Model Selection by Random Variable Generation
"Machines that Learn" (1998).
Abstract
A simple, efficient technique for model selection (selection of inputs and selection of the number of hidden neurons) is presented. The key idea is the following: a random input is appended to the pool of candidate inputs (or a neuron with random output is appended to the set of hidden neurons); all inputs (or neurons) are ranked in order of decreasing relevance by the Gram-Schmidt orthogonalization procedure ; all inputs (or neurons) which are less relevant than the random element are discarded. The risk of discarding an input, or a hidden neuron, while it is actually relevant, is estimated simply.
The method is validated, and compared to pruning techniques, on academic problems. Its efficiency for solving real problems is demonstrated on two financial applications - portfolio management (in routine operation since 1995), and the financial assessment of French townships - where the number of candidate descriptors is large, and where experts have conflicting opinions as to the actual relevance of many of them.
L. Constant, P. Lagarrigues, B. Dagues, I. Rivals, L. Personnaz
Neural modeling of an induction machine
Proceedings of the 9th International Symposium on System-Modeling-Control (SMC'98)
(Zakopane-Poland, April 27-May 1 1998).
Abstract
We present a new model of the induction machine based on neural network theory. After a brief presentation of the neural modeling methods used in this work, we introduce the neural model architecture, which is based in the Park model. We then describe the training procedure of the neural model, and give evaluations of its performances, for example on startups with a speed vector control.
Cliquez ici pour obtenir le document / Click here to download the document postscript (or ps.gz) or .pdf.
B. Quenet, G. Dreyfus, V. Cerny
The effects of internal noise, asynchrony and synaptic modifications on the coding properties of a simple model of the glomerular stage.
International Workshop on "Machines that Learn" (1998), présentation invitée.
Abstract
A simple model of the glomerular stage of the honeybee olfactory pathway has been designed, and its computational properties investigated. The main feature of this model is the fact that it provides a possible mechanism, through lateral inhibition, for the spontaneous emergence of a stable internal representation from a sequence of very different inputs, provided that all these inputs have some key features in common. Interestingly, this property is retained when some of the simplifying constraints of the model are relaxed, for instance when an internal synaptic noise is taken into account, when an external noise is added to the inputs, when the synaptic weights are changed, or when asynchronous dynamics is introduced.
When internal noise is taken into account, the dynamics of the system can be completely described as a Markov chain process whose asymptotic behavior can be derived analytically as a pseudo-Boltzmann distribution. In the case of synaptic modifications, the coding properties of the model are still tractable analytically, which is helpful for understanding the effects of learning.
I. Rivals, L. Personnaz
Construction of confidence intervals in neural modeling using a linear Taylor expansion
Proceedings of the International Workshop on Advanced Black-Box Techniques for Nonlinear
Modeling: Theory and Applications (Leuven-Belgium, July 8-10 1998).
Abstract
We introduce the theoretical results on the construction of confidence intervals for a nonlinear regression, based on the linear Taylor expansion of the corresponding nonlinear model output. The case of neural black-box modeling is then analyzed, and illustrated on an industrial application. We show that the linear Taylor expansion not only provides a confidence interval at any point of interest, but also gives a tool to detect overfitting.
Cliquez ici pour obtenir le document / Click here to download the document postscript (or ps.gz) or .pdf.
L. Oukhellou, P. Aknin, H. Stoppiglia, G. Dreyfus
A New decision Criterion for Feature Selection: Application to the Classification of Non Destructive testing Signature
European SIgnal Processing COnference (EUSIPCO'98), Rhodes (1998)
Abstract
This paper describes a new decision criterion for feature selection (or descriptor selection) and its application to a classification problem. The choice of representation space is essential in the framework of pattern recognition problems, especially when data is sparse, in which case the well-known curse of dimensionality appears inevitably. Our method associates a ranking procedure based on Orthogonal Forward Regression with a new stopping criterion based on the addition of a random descriptor. It is applied to a non destructive rail diagnosis problem that has to assign each measured rail defect to one class among several ones.
Cliquez ici pour obtenir le document / Click here to download the document postscript (or ps.gz) or .pdf.
G. Dreyfus
Reducing the Complexity of Neural Nets for Industrial Applications and Biological Models
Neuronal Information Processing - From Biological Data to Modelling and Applications, O. Parodi, ed. (World Scientific, 1998).
Abstract
The fundamental property of feedforward neural networks - parsimonious approximation - makes them excellent candidates for modeling static nonlinear processes from measured data. Similarly, feedback (or recurrent) neural networks have very attractive properties for the dynamic nonlinear modeling of artificial or natural processes; however, the design of such networks is more complex than that of feedforward neural nets, because the designer has additional degrees of freedom. In the present paper, we show that this complexity may be greatly reduced by (i) incorporating into the very structure of the network all the available mathematical knowledge about the process to be modeled, and by (ii) transforming the resulting network into a "universal" form, termed canonical form, which further reduces the complexity of analyzing and training dynamic neural models.
Cliquez ici pour obtenir le document / Click here to download the document postscript (or ps.gz) oor .pdf.
L. Constant, R. Ruelland, B. Dagues, I. Rivals, L. Personnaz
Identification and validation of a neural network estimating the fluxes of an induction machine
6th International Conference on Modelling and Simulation of Electric Machines, Converters and Systems ELECTRIMACS'99, Lisbonne (1999).
Abstract
In the frame of a study on real-time emulators of electromechanical systems, we have built a neural model of an induction machine. An original methodology is used to design the neural network architecture, as well as training sequences allowing a proper identification of the induction machine behavior in its whole range of operation. In the same spirit, exhaustive test sequences are built in order to obtain an accurate estimate of the neural model performance.
Cliquez ici pour obtenir le document / Click here to download the document postscript (or ps.gz) oor .pdf.
B. Quenet, G. Dreyfus, C. Masson
From Complex Signal to Adapted Behavior: a theoretical approach of the honeybee olfactory brain
Neuronal Information Processing - From Biological Data to Modelling and Applications, O. Parodi, ed. (World Scientific, 1998).
Abstract
The honeybee olfactory pathway is an attractive system for modeling: it is relatively simple, and it is well described functionally and morphologically. Moreover, due to the conservation of the olfactory structure through phylogeny, models may bring information of generic interest. From the point of view of behavior, this system has the ability of encoding the sensory messages into stable representations, and extracting key features from them. The neural bases of these mechanisms are still largely unknown; the purpose of the present paper is to present three different models of the same system, which make use of the same corpus of morphological and electrophysiological data, but which incorporate these data with different levels of details. We show the interrelations between these models and the specific contribution of each of them to the modeling of the olfactory pathway. We show that the design of the simplest model capitalized on the results of the previous ones, and that it suggests mechanisms for simultaneous generation of stable internal representations and key feature extraction.
Cliquez ici pour obtenir le document / Click here to download the document postscript (or ps.gz) or .pdf.
L. Constant, P. Lagarrigues, B. Dagues, I. Rivals, L. Personnaz
Modeling of Electromechanical systems using Neural Networks
Computational Intelligence and Applications, P. S. Szczepaniak (ed.), (Physica-Verlag, c/o Springer-Verlag, 1999).
Abstract
We present a new model of the induction machine based on neural network theory. After a brief presentation of the neural modeling methods used in this work, we introduce the neural model architecture, which is based in the Park model. We then describe the training procedure of the neural model, and give evaluations of its performances, for example on startups with a speed vector control.
I. Opher, D. Horn, B. Quenet
Clustering with Spiking Neurons
International Conference on Artificial Neural Networks ICANN 99, IEE Conf. Publication 470, pp. 485-490 (1999)
Abstract
We present a novel neural method for data clustering using temporal segmentation of spiking neurons. Our clustering algorithm relies only on distances between data points. Each point is associated with a neuron, and the distances are used to tdetermine the synaptic weights between those neurons. The dynamical development of this system leads to synchronous firing of neurons that belong to the same cluster, while different clusters fire at different times. Such dynamic behavior is called temporal segmentation. It is achieved via two mechanisms - intra cluster synchrony, induced by excitatory connections within each cluster, and desynchronization between clusters induced by inhibitory competition. We test our clustering method on the Iris data set. For problems that do not have a unique clustering solution, we construct a pair-correlation matrix on the basis of multiple clustering solutions. By performing a second clustering algorithm, this time on a pair-correlation matrix, we are able to define second order clusters of the original distance matrix. The method is demonstrated on a biological data set.
Cliquez ici pour obtenir le document / Click here to download the document postscript (or ps.gz) or .pdf.
L. Constant, B. Dagues, I. Rivals, L. Personnaz
Undersampling for the training of feedback neural networks on large sequences; application to the modeling of an induction machine
6th IEEE International Conference on Electronics, Circuits and Systems ICECSS'99, Chypre (1999).
Abstract
This paper proposes an economic method for the nonlinear modeling of dynamic processes using feedback neural networks, by undersampling the training sequences. The undersampling (i) allows a better exploration of the operating range of the process for a given size of the training sequences, and (ii) it speeds up the training of the feedback networks. This method is successfully applied to the training of a neural model of the electromagnetic part of an induction machine, whose sampling period must be small enough to take fast variations of the input voltage into account, i.e. smaller than 1 microsecond.
Cliquez ici pour obtenir le document / Click here to download the document postscript (or ps.gz) or .pdf.
B. Quenet, G. Dreyfus, D. Horn
Codage spatio-temporel des odeurs par un modèle du niveau glomérulaire de la voie olfactive.
4ème colloque de la Société des Neurosciences, Marseille (1999).
Abstract
Un modèle neuronal simple de l'étage glomérulaire de la voie olfactive chez l'insecte a été élaboré, dont il est possible d'étudier analytiquement la dynamique collective, même en présence de bruit synaptique. La dynamique du modèle est synchrone, mais les poids et retards synaptiques, ainsi que les signaux d'entrée, peuvent être choisis arbitrairement. Le modèle présente la propriété de coder des signaux d'entrées au moyen d'attracteurs qui constituent des images spatio-temporelles. Ces attracteurs peuvent être des cycles de différentes longueurs, robustes vis à vis du bruit synaptique et de fluctuations des signaux d'entrée.
Les enregistrements intra et extracellulaires réalisés au niveau du lobe antennaire de la voie olfactive chez l'insecte mettent en évidence une oscillation globale des neurones de cet étage, qui pourrait constituer une "onde porteuse", assortie d'une activité oscillatoire complexe et très reproductible de certains neurones, en phase avec l'oscillation globale. Cette activité, spécifique d'un substance odorante, constitue un excellent candidat pour exprimer un "code" support de l'odeur (M. Wehr, G. Laurent, Nature (1996), vol. 384, 162 - 166). La description analytique du comportement dynamique de notre modèle nous a permis de développer une méthode de recherche de poids synaptiques et de signaux d'entrée capables d'induire, au niveau des neurones formels du modèle, un comportement identifiable au code observé expérimentalement.
G. Monari, G. Dreyfus
Local Linear Least-Squares: Performing Leave-One-Out Without Leaving Anything Out
"Learning" Conference, Snowbird (1999).
Abstract
Based on a local linear expansion, in parameter space, of the least squares solution, we show that the effect of removing an example from the training set can be accurately predicted, and that, in addition, the geometrical interpretation of this expansion leads to a method of detecting models which are likely to exhibit overfitting. As a consequence, the computationally expensive leave-one-out procedure can be replaced by a prediction thereof: instead of training N different models on the data sets with N-1 examples, one has to train a single model on the whole data set. Moreover, we demonstrate that selecting models on the basis of the performance evaluation after leave-one-out training may lead to overfitted model even if the jacobian matrix is shown to be non-singular.
The price to be paid for predicting the leave-one result is the computation of the diagonal elements of the orthogonal projection matrix onto the local solution subspace, in the neighborhood of the parameter vector obtained by training; this computational overhead is still very slight since this computation must be performed after training only. We have performed a detailed comparison between standard leave-one-out and our method, on a teacher-student problem, for performance evaluation and model selection. The method is by no means specific to neural networks: it can be applied to any nonlinear regression method, provided the jacobian matrix of the model is easily available.
Cliquez ici pour obtenir un document sur ce sujet/ Click here to download a relevant document postscript (or ps.gz) or .pdf.
P. Roussel-Ragot, C. Graff
La modélisation par des gaussiennes des décharges électriques de poissons Mormyridae permet d'étudier quels compromis ont conduit à la forme de ces décharges.
4ème Colloque de la Société des Neurosciences, Marseille (1999).
Abstract
Tous les poissons électriques Mormyridae produisent des décharges électriques brèves dues aux potentiels d'action d'éléments excitables : les électroplaques. Ces décharges présentent deux ou trois phases principales selon les espèces. La forme et la durée des décharges dépendent de la synchronisation des décharges de chaque face des électroplaques, ainsi que du temps de retard avec lesquelles les faces rostrales deviennent actives. Afin d'avoir accès à ces deux grandeurs, nous avons modélisé les décharges par une somme algébrique de deux gaussiennes à l'aide d'un algorithme d'optimisation du second ordre. Chaque gaussienne correspond à la distribution dans le temps de la dépolarisation de l'une des faces des électroplaques. Ces modélisations ont conduit à une bonne approximation de la forme des décharges électriques en fonction du temps, et nous ont donné accès aux grandeurs caractéristiques de l'activité de chaque face : l'amplitude maximale en tension, le centre sur l'abscisse des templs, et la dispersion (écart-type) dans le temps. Nous avons alors pu étudier l'effet d'une modification du temps de retard des faces rostrales par rapport aux faces caudales.
Il en résulte une modification de la forme des décharges dont l'analyse (amplitude maximale, composantes en fréquence, rendement) permet de constater que la décharge naturelle est beaucoup plus courte que celle qui fournit l'amplitude maximale (portée maximale du signal) ou la variation de polarité la plus rapide (stimulation optimale des électrorécepteurs). Ce résultat indiquerait que les décharges de Mormyres sont produites de façon à être aussi courtes que possible, dût-il en coûter une forte perte d'énergie, et que les durées courtes sont même favorisées par rapport à des transitions rapides. Ces caractéristiques semblent favoriser la fonction d'électrolocalisation active des objets présents dans le milieu par rapport à celle d'électrocommunication avec les congénères.
B. Quenet, G. Dreyfus, D. Horn
A Simple Neural Network with Adapted Synaptic Weights May Account for the Codes Observed in Oscillating Neural Assemblies of the Olfactory Pathway.
"Learning" Conference, Snowbird (1999), présentation invitée.
Abstract
A simple model of the glomerular stage of the insect olfactory pathway has been designed, which is analytically tractable, even when synaptic noise, random synaptic weights, inputs or delays are taken into account. This model exhibits the property of encoding its inputs through spatio-temporal patterns which are the attractors of its dynamics. These attractors can be long cycles, robust against synaptic noise and also to input fluctuations, provided that the latter occur within well defined limits.
The electrophysiological data recorded in the glomerular stage of the insect olfactory pathway show both a coherent global oscillating behaviour of the neurons of this stage - carrier waveform ? -, and a reproducible complex activity pattern - code ? - of some of these neurons, in phase with the global oscillation. Capitalizing on the analytical knowledge available on our formal model of this stage, we propose a possible interpretation of this type of biological activity patterns: we derive a method to build a set of adaptive weights and inputs leading to the attractor corrresponding to a given neuronal code. We also discuss the impact of these results on possible learning mechanisms.
G. Dreyfus, Y. Oussar, J.L. Ploix
Knowledge-based Dynamic Neural Modeling of Industrial Processes
Neural Computing in Science and Technology, NCST'99, Tel Aviv (1999), conférence invitée.
Abstract
Neural networks are traditionally designed as black-box models, that make little or no use of prior
knowledge. Knowledge-based neural modeling is a technique that allows the model designer to
take advantage of existing knowledge-based models (expressed as a set of algebraic and
differential equations), albeit approximate, to define the architecture of a recurrent neural model.
The resulting models combine the flexibility of black-box models with the legibility of
knowledge-based models. The principles of the method and two industrial applications are presented.
M. Stricker, F. Vichot, G. DReyfus, F. Wolinski
Two Steps Feature Selection and Neural Network Classification for the TREC-8 Routing
Eighth International Text REtrieval Conference (TREC-8), Gaithersburg (1999).
Abstract
At the Caisse des Dépôts et Consignations (CDC), the Agence France-Presse (AFP) news releases are filtered continuously according to the users' interests. Once a user has specified a topic of interest, a filter is customized to fit this user's profile. Until now, these filters would rely on rule-based methods, whose efficiency is proven [Vichot et al., 1999], but which require a large amount of work for each specific filter. This drawback can be avoided by using statistical methods which have the ability to learn from examples of relevant documents. Recently, we have developed a methodology for the AFP corpus. This paper presents its application to the TREC-8 corpus.
For the TREC-8 routing, one specific filter is built for each topic. Each filter is a classifier trained to recognize the documents that are relevant to the topic. When presented with a document, each classifier estimates the probability for the document to be relevant to the topic for which it has been trained. Since the procedure for building a filter is topic-independent, the system is fully automatic. Therefore, we describe it for one topic; the procedure is repeated 50 times.
By making use of a sample of documents that have previously been evaluated as relevant or not relevant to a particular topic, a term selection is performed, and a neural network is trained. Each document is represented by a vector of frequencies of a list of selected terms. This list depends on the topic to be filtered; it is constructed in two steps. The first step defines the characteristic words used in the relevant documents of the corpus; the second one chooses, among the previous list, the most discriminant ones. The length of the vector is optimized automatically for each topic. At the end of the term selection, a vector of typically 25 words is defined for the topic, so that each document which has to be processed is represented by a vector of term frequencies.
This vector is subsequently input to a classifier that is trained from the same sample. After training, the classifier estimates for each document of a test set its probability of being relevant; for submission to TREC, the top 1000 documents are ranked in order of decreasing relevance.
Cliquez ici pour obtenir le document / Click here to download the document postscript (or ps.gz) or .pdf
B. Quenet, G. Dreyfus, D. Horn
The Codes Observed in Oscillating Neural Assemblies of the Olfactory Pathway May Be Reproduced by a Simple Neural Network
Neural Computing in Science and Technology, NCST'99, Tel Aviv (1999).
Abstract
The electrophysiological data recorded in the glomerular stage of the insect olfactory pathway show both a coherent global oscillating behavior of the neurons of this stage - carrier waveform? -, and a reproducible complex activity pattern - code? - of some of these neurons, in phase with the global oscillation. We propose a possible interpretation of this type of biological activity patterns, using a simple model of the glomerular stage of the insect olfactory pathway that has been previously designed. This model is analytically tractable, even when synaptic noise, random synaptic weights, inputs or delays are taken into account. This model exhibits the property of coding its inputs through spatio-temporal patterns which are the attractors of its dynamics. These attractors can be long cycles, robust against synaptic noise and also to input fluctuations, provided that the latter occur within well-defined limits. We give an example of a set of adapted synaptic weights and inputs leading our model to the attractors corresponding to a given neuronal code, drawn from experimental data.
Y. Oussar, G. Dreyfus
Training Complex Dynamic Knowledge-Based Models
"Learning" Conference, Snowbird (2000), présentation invitée.
Abstract
Dynamic knowledge-based neural modeling is a technique that combines the best of two worlds: knowledge-based modeling and black-box modeling. A knowledge-based model is a mathematical description of the phenomena occurring in the process under investigation, based on the equations of physics and chemistry (or biology, sociology, etc.). Conversely, a black-box model is a mathematical description of the process that may have a large number of parameters, estimated from measurements performed on the process; it does not take into account any prior knowledge on the process (or a very limited one). A knowledge-based neural model may be regarded as a trade-off between a knowledge-based model and a black-box model. It makes it possible to take into account all phenomena that are not modeled, through prior knowledge, with the required accuracy. Since a larger amount of prior knowledge is used in the design of a knowledge-based neural model than in the design of a black-box model, a smaller amount of experimental data may be required to estimate its parameters reliably. The first step in the design of a dynamic knowledge-based neural network consists in (i) discretizing the knowledge-based model by an appropriate discretization scheme, and (ii) introducing black-box neural networks into the resulting model wherever necessary. If an explicit (forward-time) discretization scheme is used, the resulting model can be trained by standard backpropagation through time. However, explicit schemes are known to be prone to numerical instabilities; therefore, it may be mandatory to use implicit (backward-time) schemes. In that case, the training of the neural networks embedded into the model is far from trivial.
We address the problem of the training of such "hybrid" models, after discretization of the knowledge-based model by an implicit scheme. We propose a general training technique for dynamic parameterized models that can be used in the framework of implicit schemes. The integration of implicit recurrent equations involves either the use of a nonlinear algebraic equation or the use of a substitution scheme: we show that the latter method can be implemented in a form that allows the use of backpropagation for the training of the model.
Both a didactic example and an industrial application (the modeling of a drying process, involving coupled partial differential equations) are presented.
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 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 reproduce complex activity patterns from electrophysiological recordings in insects. Biologically plausible mechanisms of synaptic adaptation are discussed.
Cliquez ici pour obtenir le document / Click here to download the document postscript (or ps.gz) or .pdf.
M. Stricker, F. Vichot, G. Dreyfus, F. Wolinski
Towards the Automatic Design of Efficient Custom Filters
RFIA'2000, Paris (2000)
Abstract
The large amount of financial information released daily by press agencies makes the design of custom information filters, at a low development cost, an important issue. We present a comprehensive methodology for designing an information filter on a prescribed topic with a statistical approach. By using a neural network in conjunction with a search engine, we are able to automate to a large extent the construction of the training set. Since the performance of a filter depends strongly on the relevance of the selected inputs, we describe a method for performing feature selection. The filter is in routine use to filter news releases concerning company shareholding, with a precision and a recall of 90%.
Cliquez ici pour obtenir le document / Click here to download the document postscript (or ps.gz) or .pdf.
F. Wolinski, F. Vichot, M. Stricker
Using Learning-Based Filters to Detect Rule-based Filtering Obsolescence
RIAO'2000, Paris (2000).
Abstract
For years, Caisse des Dépôts et Consignations has produced information filtering applications. To be operational, these applications require high filtering performances which are achieved by using rule-based filters. With this technique, an administrator has to tune a set of rules for each topic. However, filters become obsolescent over time. The decrease of their performances is due to diachronic polysemy of terms that involves a loss of precision and to diachronic polymorphism of concepts that involves a loss of recall.
To help the administrator to maintain his filters, we have developed a method which automatically detects filtering obsolescence. It consists in making a learning-based control filter using a set of documents which have already been categorised as relevant or not relevant by the rule-based filter. The idea is to supervise this filter by processing a differential comparison of its outcomes with those of the control one.
This method has many advantages. It is simple to implement since the training set used by the learning is supplied by the rule-based filter. Thus, both the making and the use of the control filter are fully automatic. With automatic detection of obsolescence, learning-based filtering finds a rich application which offers interesting prospects.
Cliquez ici pour obtenir le document / Click here to download the document postscript (or ps.gz) or .pdf.
P. Roussel, F. Moncet, B. Barrieu, A. Viola
Modélisation d'un processus dynamique à l'aide de réseaux de neurones bouclés. Application à la modélisation de la relation pluie-hauteur d'eau dans un réseau d'assainissement et à la détection de défaillance de capteur.
Innovative technologies in urban drainage, 1, 919-926, G.R.A.I.E.
Abstract
Systematic measurements of the rainfall as well as of the resulting water height, flow and velocity in the pipes, have been performed for several years in the sewer system of a French department. Due to possible electrical failures or slow sensor drifts, a validation procedure must be performed in order to guarantee the validity of the stored measurements.
To this end, we have developed a neural dynamic black-box model of the rainfall-water height relationship on a simple urban catchment equipped with a single water gauge and a single rainfall gauge. The validity of the measurements is assessed from the comparison between the measured heights and the corresponding values predicted by the model.
Cliquez ici pour obtenir le document / Click here to download the .pdf. document
B. Quenet, S. Sirapian, R. Dubois, G. Dreyfus
Generation of Olfactory Neural Codes by a Network of Hodgkin-Huxley Neurones
4th International Workshop on Neural Coding, Plymouth, UK (2001)
Abstract
Networks of synchronously updated McCulloch & Pitts neurones exhibit spontaneously complex spatio-temporal patterns that can be compared to the activities of biological neurones in phase with a periodic LFP, as demonstrated experimentally by Wehr & Laurent in the locust olfactory pathway. Modelling biological neural nets with networks of very 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 and inputs in order to make its neurones exhibit a given sequence of activity. In the presentation, we address the following question: once a formal network has been built, that is able to reproduce quantitatively experimentally observed neuronal codes, can it serve as a guide to design a network of more realistic (Hodgkin-Huxley) formal neurones that exhibits the same dynamical behaviour? We demonstrate 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.
B. Quenet, D. Horn, R. Dubois, S. Sirapian, G. Dreyfus
Dynamic Neural Filter of Hodgkin-Huxley units
International Workshop on Learning, Snowbird (2002)
Abstract
The coding properties of a Dynamic Neural Filter (DNF) made of McCulloch and Pitts units can be completely described analytically, even in the presence of intrinsic noise. The main property of this type of network is its ability to map stable input vectors to spatiotemporal sequences of the neuronal activity. Two major problems can be solved with such a DNF: (1) the direct problem, which investigates the emerging spatiotemporal pattern, given the connection matrix and the input, and (2) the inverse problem, which investigates the family of connection matrices and inputs that can elicit a given spatiotemporal pattern. The existence of such a family is not guaranteed: the emergence of a given spatiotemporal pattern may require the existence of hidden neurons. Given an arbitrary temporal sequence of binary activities, we suggest here a method in order to (1) build an appropriate activity pattern for a parsimonious number of hidden neurons, when they are necessary, and (2) find the family of matrices and inputs compatible with the given sequence. Such a modeling approach may be useful when applied to experimental data recorded in biological neurons, once they are represented as binary temporal sequences of activities. This method provides a number of hidden neurons,which is an index of the complexity of the observed neural task, and also provides a numerical guide for the design of a model network of more biologically plausible formal neurons, such as Hodgkin-Huxley ones. We have therefore developed a tool that combines the advantages of an analytical approach and of a biologically plausible modeling, alllowing us to reproduce exactly experimentally observed sequences with spiking neurons. In this presentation, we illustrate our method with experimental data recorded by Wehr and Laurent in the locust antennal lobe.
R. Dubois, B. Quenet, Y. Faisandier, G. Dreyfus
Modeling Beats with Bumps
International Workshop on Learning, Snowbird (2003), invited presentation.
Abstract
Long-time electrocardiographic records (Holter) are an important tool in non-invasive electrocardiology. Such a record features at least of 100,000 heart beats for a 24-hour record, but only a few of them may express a heart anomaly. Therefore, a fully automated analysis is desirable as a computer-aided diagnosis tool.
To that effect, we suggest a mathematical decomposition of each heart beat on a specific family of regressors ("bumps"). Each bump has five adjustable parameters. Unlike conventional regressors (wavelet, RBF,...), bumps are designed to fit the usual cardiac "waves" that are defined by cardiologists; the shape and position of the waves are the basis of the experts' diagnostics. Since each wave is fitted by a single regressor (or possibly two), the number of parameters needed to model the relevant information is parsimonious, and the decomposition meets the intelligibility requirements of automated medical diagnostics tools.
Modeling a complete heart beat with N bumps requires N iterations of the following algorithm:
- selection of a bump in a library of bumps that is generated once and for all; the selection is performed by a Gram-Schmidt orthogonalization-based selection method;
- estimation of the bump parameters by training with a second order optimization algorithm under constraints;
- orthogonalization of the rest of the library with respect to the bumps whose parameters were estimated in the previous iterations.
Once all the cardiac waves of a heart beat are modelled by N bumps, classifiers (linear, neural nets, SVM's, ) and hidden Markov models assign a "medical" label to each bump.
This approach was tested on several international databases, showing that the number of bumps needed to model accurately the medically significant waves is at most N = 6 .
G. Dreyfus, Y. Oussar
Non-linear black-box model selection (invited plenary presentation)
Neural Networks for Signal Processing (2003).
Abstract
Neural networks, and, more generally, nonlinear-in-their-parameters models, are recognized tools for engineers; the basic issues in the training of those models may be considered as essentially solved. However, there are more general issues, not specific to neural networks, which are still open. One of them has become all-important as new areas of applications for nonlinear modeling are opening up: the problem of model selection. That includes:
- variable selection: find, among a set of candidate variables, the variables that are really relevant to the task, i.e. whose influence on the quantity to be modeled is larger than the effect of noise or disturbances; present applications in bioinformatics or in natural language processing may involve hundreds of candidate variables;
- complexity selection: find the appropriate complexity (i.e. the number of hidden neurons for neural networks, of monomials for polynomial models, etc.)
- cost function minimum selection: for models that are nonlinear in their parameters, training by minimizing a cost function can generate different models corresponding to different minima of the cost function; therefore, a selection must be performed among models that have the same complexity but different sets of parameters; for linear-in-their-parameters models, that problem is irrelevant since the least-squares cost function has a single minimum.
The presentation will emphasize a model design methodology and recent developments in model selection, illustrated by academic examples and by industrial applications.
Y. Oussar, G. Dreyfus
Generalized leverages and generalization error
International Workshop on Learning (2004), invited presentation.
Abstract
A .Goulon-Sigwalt-Abram, A. Duprat, G. Dreyfus
Learning numbers from graphs
Applied Statistical Modeling and Data Analysis (2005)
Abstract
The recent developments of statistical learning focused mainly on vector machines, i.e. on machines that learn from examples described by a vector of features. There are many fields where structured data must be handled; therefore, it would be desirable to learn from examples described by graphs. The presentation describes graph machines, which learn real numbers from graphs. Applications in the field of Quantitative Structure-Activity Relations (QSAR), which aim at predicting properties of molecules from their (graph) structures, are described.
Cliquez ici pour obtenir le document / Click here to download the .pdf. document
G. Dreyfus
Model and variable selection for nonlinear model design: some developments and methodology (invited plenary presentation)
Engineering Application of Neural Networks (EANN 2005)
Abstract
The paper discusses the issues of model selection and variable (feature) selection for nonlinear modeling. Methods are described, which were designed to be simple and to involve little computational overhead, but are nevertheless generic, since they do not involve ad hoc heuristics. The principles of the methods are described, and pointers to the description of applications are provided.
Cliquez ici pour obtenir le document / Click here to download the .pdf. document
F. Vialatte, A. Cichocki, G. Dreyfus, T. Musha, S. L. Shishkin, R. Gervais
Early Detection of Alzheimer’s Disease by Blind Source Separation, Time Frequency Representation, and Bump Modeling of EEG Signals (invited presentation).
Lecture Notes in Computer Science, 3696 (2005) 683-692 (Springer).
Abstract
The early detection Alzheimer’s disease (AD) is an important challenge. In this paper, we propose a novel method for early detection of AD using electroencephalographic (EEG) recordings: first a blind source separation algorithm is applied to extract the most significant spatio-temporal components; these components are subsequently wavelet transformed; the resulting time-frequency representation is approximated by sparse “bump modeling”; finally, reliable and discriminant features are selected by orthogonal forward regression and the random probe method. These features are fed to a simple neural network classifier. The method was applied to EEG recorded in patients with Mild Cognitive Impairment (MCI) who later developed AD, and in age-matched controls. This method leads to a substantially improved performance (93% correctly classified, with improved sensitivity and specificity) over classification results previously published on the same set of data. The method is expected to be applicable to a wide variety of EEG classification problems.
Cliquez ici pour obtenir le document / Click here to download the .pdf. document
F. Vialatte, A. Cichocki, G. Dreyfus, T. Musha, T. M. Rutkowski, R. Gervais
Blind Source Separation and Sparse Bump Modelling of Time-Frequency Representation of EEG Signals: New Tools for Early Detection of Alzheimer's Disease
Machine Learning for Signal Processing (MLSP 2005)
Abstract
The early detection of Alzheimer’s disease (AD) is an important challenge. In this paper, we propose a novel method for early detection of AD using only electroencephalographic (EEG) recordings for patients with Mild Cognitive Impairment (MCI) without any clinical symptoms of the disease who later developed AD. In our method, first a blind source separation algorithm is applied to extract the most significant spatiotemporal uncorrelated components; afterward these components are wavelet transformed; subsequently the wavelets or more generally time-frequency representation(TFR) is approximated with sparse bump modeling approach. Finally, reliable and discriminant features are selected and reduced with orthogonal forward regression and the random probe methods. The proposed features were finally fed to a simple neural network classifier. The presented method leads to a substantially improved performance (93% correctly classified - improved sensitivity and specificity) over classification results previously published on the same set of data. We hope that the new computational and machine learning tools provide some new insights in a wide range of clinical settings, both diagnostic and predictive.
Cliquez ici pour obtenir le document / Click here to download the .pdf. document
O. Foucard , G. Horcholle-Bossavit, B. Quenet
Oscillation and coding in a formal neural network considered as a guide for plausible simulations of the insect olfactory system
6th International Neural Coding Workshop, August 2005, Marburg, Germany
Abstract
For the analysis of coding mechanisms in the insect olfactory system, a fully connected network of synchronously updated neurons (McCulloch-Pitts type or MCP) 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. As an additional step, we separate the populations of PN and local inhibitory neurons (LN) in the DNF which remains a guide, as done in a previous work (Quenet et al, 2005) for simulations based on biological plausible neurons (Hodgkin-Huxley type or H-H). We show that a parsimonious network of 10 H-H neurons generates actionpotentials with a spatio-temporal pattern corresponding to the olfactory codes at the antennal lobe level. Taking advantage of the possible analytical description of the dynamical behaviour of networks of MCP neurones, we study the effects of considering two different populations of neurones, excitatory and inhibitory, on both coding and on oscillatory properties.
In order to further improve the biological plausibility of the MCP neural network we construct a more complex recurrent network. Our aim is to characterize the population dynamics for different values of the delays, connectivity and inputs. We explore both the effects of the proportion of internal inhibition and excitation, and those of the proportion of external excitation to the internal one on the behaviours the network exhibits. These effects can be compared to the ones that have been demonstrated in networks of integrate and fire neurones (Brunel, 2000). In addition, different levels of noise are examined and lead to transitions from synchronous to asynchronous firing modes.
We find a set of parameters which leads to both the emergence of robust oscillations and spatio-temporal patterns showing that PNs' activity is phase-locked to different cycles of the oscillations similar to the local field potential (LFP), and change with inputs.
We show that it is possible to introduce in the formal model of discrete time neurons additional biological constraints provided by experimental data describing the functional relationships between neuron types in the olfactory system. The pertinent structure of the formal network (delays, connection matrix and input vector) can be used then for more realistic simulations with continuous time neurons.
References :
B. Quenet, D. Horn, The Dynamic Neural Filter: a Binary Model of Spatiotemporal Coding, Neural Computation, vol. 15, pp. 309-329 (2003).
B. Quenet, G. Horcholle-Bossavit, A. Wohrer, G. Dreyfus. Formal modeling with multistate neurones and multidimensional synapses. Biosystems, vol. 79, pp. 21-32, (2005)
N. Brunel, Dynamics of sparsely connected network of excitatory and inhibitory spiking neurons, J. Comput. Neurosci, vol. 8, pp 183-208 (2000).
Bruce Denby, Yacine Oussar, Gérard Dreyfus, Maureen Stone
Prospects for a silent speech interface using ultrasound imaging
International Conference on Acoustics, Speech and Signal Processing (CASSP 2006)
Abstract
The feasibility of a silent speech interface using ultrasound (US) imaging and lip profile video is investigated by examining the quality of line spectral frequencies (LSF) derived from the image sequences. It is found that the data do not at present allow reliable identification of silences and fricatives, but that LSF’s recovered from vocalized passages are compatible with the synthesis of intelligible speech.
Cliquez ici pour obtenir le document / Click here to download the .pdf. document
Abdeddayem Haggui, Fabrice Extramiana, Pierre Maison-Blanche, Rémi Dubois, Julien Seitz, Paul Milliez, Bruno Cauchemez, Philippe Beaufils, Antoine Leenhardt.
Stabilité dans le temps des paramètres de l’onde T évalués par analyse en composante principale.
XVIèmes Journées Françaises de Cardiologie (2006)
Abstract
On propose d'utiliser l’analyse en composantes principales (ACP) de l’onde T pour caractériser la complexité de la repolarisation ventriculaire. Une augmentation de la complexité de l’onde T semble avoir une valeur pronostique dans différentes pathologies. Les paramètres de l’onde T en analyse en composantes principales semblent stables sur une période de un mois. L’ACP peut donc être utilisée pour caractériser les effets de substances médicamenteuses sur la repolarisation ventriculaire.
Cliquez ici pour obtenir le document / Click here to download the .pdf. document
A .Goulon-Sigwalt-Abram, A. Duprat, G. Dreyfus
Graph Machines and Their Applications to Computer-Aided Drug Design: a New Approach to Learning from Structured Data
Unconventional Computing 2006, Lecture Notes in Computer Science, vol. 4135, pp. 1 – 19, Springer (2006)
Abstract
The recent developments of statistical learning focused on vector machines, which learn from examples that are described by vectors of features. However, there are many fields where structured data must be handled; therefore, it would be desirable to learn from examples described by graphs. Graph machines learn real numbers from graphs. Basically, for each input graph, a separate learning machine is built, whose algebraic structure contains the same information as the graph. We describe the training of such machines, and show that virtual leave-one-out, a powerful method for assessing the generalization capabilities of conventional vector machines, can be extended to graph machines. Academic examples are described, together with applications to the prediction of pharmaceutical activities of molecules and to the classification of properties; the potential of graph machines for computer-aided drug design is highlighted.
Cliquez ici pour obtenir le document / Click here to download the .pdf. document
[ Retour à la liste des publications / Back to the list of publications ]
Olivier Romain, Bruce Denby
Prototype of a software-defined broadcast media indexing engine
International Conference on Acoustics, Signal and Speech Processing (ICASSP), Honolulu (2007)
Abstract
The article compares two approaches to the description of ultrasound vocal tract images for application in a “silent speech interface,” one based on tongue contour modeling, and a second, global coding approach in which images are projected onto a feature space of Eigentongues. A curvature-based lip profile feature extraction method is also presented. Extracted visual features are input to a neural network which learns the relation between the vocal tract configuration and line spectrum frequencies (LSF) contained in a one-hour speech corpus. An examination of the quality of LSF’s derived from the two approaches demonstrates that the eigentongues approach has a more efficient implementation and provides superior results based on a normalized mean squared error criterion.
[ Retour à la liste des publications / Back to the list of publications ]
G. Dreyfus
Random probes for variable selection
Multiple Simultaneous Hypothesis Testing, Paris (2007)
Abstract
The random probe method for variable selection is a recently developed method, which provides direct control on Type I error and indirect control of Type II errors. We describe the principle of the method, discuss its limitations, and provide some typical applications.
[ Retour à la liste des publications / Back to the list of publications ]
T.Hueber, G.Chollet, B. Denby, M. Stone,L. Zouari
Ouisper: Corpus Based Synthesis Driven by Articulatory Data
International Conference on Phonetic Science (ICPhS), Saarbrücken, Germany (2007).
A. Goulon, A. Duprat, G. Dreyfus
Virtual Leave-One-Out Estimation of Generalization Error
International Learning Workshop, Puerto Rico (2007)
Virtual leave-one-out is an attractive alternative to cross-validation for estimating the prediction error of models, especially nonlinear ones: training is performed on the available data, and an estimate of the prediction error that would have been incurred on each example, if it had been withdrawn from the training set, is computed. For linear models, the virtual leave-one-out score reduces to the PRESS statistic. The computation of the leave-one-out score involves the computation of the leverage of each example, which indicates the influence of each example on the model.
There is a growing interest in graph machines and recursive networks, which learn from structured data, i.e. examples that are described by graphs instead of vectors. Graph machines encode the structure of the graphs and simultaneously provide a prediction of the properties of interest. Therefore the representation of the structured data is learnt together with the learning of the task, which exempts the model designer from finding a vector representation for the data.
We show how the computation of leverages, hence of the virtual leave-one-out score, can be extended to graph machines and recursive networks. We compare the real and virtual leave-one-out scores on several data sets. We describe examples of graph machine selection by virtual leave-one-out, and we show that, in addition, virtual leave-one-out can provide insight into the design of the predictors, i.e. the encoding of the input data into directed acyclic graphs. We illustrate these topics on several regression tasks, e.g. the estimation of the Gibbs free energy of solvation of molecules, the toxicity of halogenated aliphatic compounds, or the agonist activities of ecdysteroids.
[ Retour à la liste des publications / Back to the list of publications ]
B. Denby, Y. Oussar, I. Ahriz
Geolocalisation in Cellular Telephone Networks
Proceedings of NATO 2007 Advanced Study Institute on Mining Massive Data Sets for Security,
F. Fogelman-Soulié, D. Perrotta, J. Piskorski & R. Steinberger, Eds., IOS Press, pp. 357-365,
Amsterdam ( 2008).
Abstract
The paper gives an overview of GPS and radio interface based geolocalisation techniques for cellular telephone networks,
including the E911 and E112 initiatives, Location Based Services, and law enforcement/security applications. An example of
localisation using the Database Correlation Method is also presented.
T. Hueber, G. Aversano, G. Chollet, B. Denby, G. Dreyfus, Y. Oussar, P. Roussel, M. Stone
Eigentongue feature extraction for an ultrasound-based silent speech interface
International Conference on Acoustics, Signal and Speech Processing (ICASSP), Honolulu (2007)
Abstract
The article compares two approaches to the description of ultrasound vocal tract images for application in a “silent speech interface,” one based on tongue contour modeling, and a second, global coding approach in which images are projected onto a feature space of Eigentongues. A curvature-based lip profile feature extraction method is also presented. Extracted visual features are input to a neural network which learns the relation between the vocal tract configuration and line spectrum frequencies (LSF) contained in a one-hour speech corpus. An examination of the quality of LSF’s derived from the two approaches demonstrates that the eigentongues approach has a more efficient implementation and provides superior results based on a normalized mean squared error criterion.
Cliquez ici pour obtenir le document / Click here to download the .pdf. document
T. Hueber, G. Chollet, B. Denby, G. Dreyfus, M. Stone,
Continuous-Speech Phone Recognition from Ultrasound and Optical Images of the Tongue and Lips
Interspeech, Anvers (2007).
Abstract
The article describes a video-only speech recognition system for a “silent speech interface” application, using ultrasound and optical images of the voice organ. A one-hour audio-visual speech corpus was phonetically labeled using an automatic speech alignment procedure and robust visual feature extraction techniques. HMM-based stochastic models were estimated separately on the visual and acoustic corpus. The performance of the visual speech recognition system is compared to a traditional acoustic-based recognizer. Index Terms: speech recognition, audio-visual speech description, silent speech interface, machine learning
Cliquez ici pour obtenir le document / Click here to download the .pdf. document
[ Retour à la liste des publications / Back to the list of publications ]
T. Hueber, G. Chollet, B. Denby, G. Dreyfus, M. Stone,
Towards a Segmental Vocoder Driven by Ultrasound and Optical Images of the Tongue and Lips
Interspeech, pp. 2028-2031, Brisbane, Australia (2008).
Abstract
This article presents a framework for a phonetic vocoder driven by ultrasound and optical images of the tongue and lips for a “silent speech interface” application. The system is built around an HMM-based visual phone recognition step which provides target phonetic sequences from a continuous visual observation stream. The phonetic target constrains the search for the optimal sequence of diphones that maximizes similarity to the input test data in visual space subject to a unit concatenation cost in the acoustic domain. The final speech waveform is generated using “Harmonic plus Noise Model” synthesis techniques. Experimental results are based on a one-hour continuous speech audiovisual database comprising ultrasound images of the tongue and both frontal and lateral view of the speaker’s lips.
Index Terms: silent speech, corpus-based speech synthesis, visual speech recognition
Cliquez ici pour obtenir le document / Click here to download the .pdf. document
[ Retour à la liste des publications / Back to the list of publications ]
T. Hueber, G. Chollet, B. Denby, G. Dreyfus, M. Stone,
Phone Recognition from Ultrasound and Optical Video Sequences for a Silent Speech Interface
Interspeech, pp. 2032 - 2035, Brisbane, Australia (2008) .
Abstract
Latest results on continuous speech phone recognition from video observations of the tongue and lips are described in the context of an ultrasound-based silent speech interface. The study is based on a new 61-minute audiovisual database containing ultrasound sequences of the tongue as well as both frontal and lateral view of the speaker’s lips. Phonetically balanced and exhibiting good diphone coverage, this database is designed both for recognition and corpus-based synthesis purposes. Acoustic waveforms are phonetically labeled, and visual sequences coded using PCA-based robust feature extraction techniques. Visual and acoustic observations of each phonetic class are modeled by continuous HMMs, allowing the performance of the visual phone recognizer to be compared to a traditional acoustic-based phone recognition experiment. The phone recognition confusion matrix is also discussed in detail.
Cliquez ici pour obtenir le document / Click here to download the .pdf. document
[ Retour à la liste des publications / Back to the list of publications ]
B. Denby, Y. Oussar, I. Ahriz, G. Dreyfus,
High-Performance Indoor Localization with Full-Band GSM Fingerprints,
International Conference on Communications ICC 2009, Dresden (2009).
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 .pdf. document
[ Retour à la liste des publications / Back to the list of publications ]
Haini Qu, Y.Oussar, G. Dreyfus,
Dynamic modeling by support vector machines,
International Learning Workshop, Tampa (2009).
Abstract
Kernel methods are very popular for static modeling, i.e. classification or nonlinear regression. Only few attempts have been made at extending the concepts to dynamic modeling, i.e. models where some or all variables of the model are the predictions provided by the model at previous sampling times. The feedback thus introduced results in an increased complexity; it is well known, however, that, in the presence of output noise, the optimal model is a recurrent model. Therefore, for modeling processes with output noise, which is a very frequent situation, taking feedback into account during training is mandatory.
To the best of our knowledge, the first attempt at training recurrent SVMs was performed by Suykens et al. in the framework of LS-SVMs. The authors made a drastic simplification by neglecting the regularization terms in the cost function, thereby losing one of the salient features of SVMs, i.e. their built-in regularization mechanism. In the present work, we show that such a simplification is not necessary, at the expense of an increased, but still manageable, complexity in the equations.
The approach is validated on various examples, academic and industrial. On academic problems, it is shown that the method does find optimal predictors, i.e. models for which the variance of the modeling error, estimated on a test set, is equal to the variance of the noise present in the training sequences.
On the negative side, it appears that the computational cost of training these models is much larger than that of recurrent neural networks, without any gain in accuracy in the examples investigated so far.
P. Bouchet, R. Dubois, C. Henry, P. Roussel, G. Dreyfus,
Machine learning for shock decision in implanted defibrillators,
International Learning Workshop, Tampa (2009).
Abstract
The discrimination of Ventricular Tachycardia (VT) from Supra-Ventricular Tachycardia (SVT) remains a major challenge for appropriate therapy delivery in Implantable Cardioverter Defibrillators (ICDs). Unlike SVT, VT is a life-threatening arrhythmia that may lead to sudden death unless an appropriate shock is delivered. The discrimination in ICDs is performed from endocardial measurements of the electrical activity of the heart (EGM). Historically, only time intervals extracted from EGMs were used for the diagnosis. In the last decade, an additional analysis of features extracted directly from the shape of a single EGM channel led to improved performances, especially in order to avoid inappropriate shocks, which are very painful and stressful for patients. A recent study shows that inappropriate shocks occurred in 11.5% of the prophylactic ICD patients and accounted for 31.2% of the total shock episodes [1].
The discrimination method proposed here relies on the simultaneous analysis of two different ventricular EGM channels, available in most common implanted defibrillators. Therefore, we have designed a two-dimensional representation of both a far-field (RVp-Can) and a near-field (RVp-RVd) EGM signals (Figure 1), named “Spatial Projection Of Tachycardia electrograms” (SPOT) (Figure 2). The SPOT curve of a cardiac cycle is the plot of the amplitude of the far-field sensing signal versus the amplitude of the near-field sensing signal, with time as a parameter. Features extracted in this space representation allow curve comparison. The underlying assumption is that the morphology of an SVT SPOT curve is similar to that of the reference curve constructed from the patient’s normal EGMs, while the SPOT curve for a VT is different (Figure 2): this is justified by the fact that the electrical signals pertaining to normal heartbeats and to SVT heartbeats originate from the atria and follow the same electrical conduction pathway to the ventricles, while VT electrical signals, originating from the ventricles, have a different activation pattern, leading to a change in the morphology of the signals received by the electrodes.
Morphological features are extracted from the curves, and candidate features for statistical classification, based on physiological prior knowledge, are computed to compare arrhythmia and reference SPOT curves: the average angle of the relative velocity vectors, the correlation coefficient between the norms of the velocity and the correlation coefficient between their curvatures. These three features, and two additional timing descriptors, form a set of candidate features, on which statistical feature selection was performed by the random probe method [2]. Classifiers of various types and complexities (Linear, Polynomial, Neural networks, Support Vector Machines) were subsequently trained. Model selection was carried out by leave-one-out.
SVM classification on a data base of 93 VT and 26 SVT from 73 patients provided 95.7% sensitivity and 92.3% specificity. Therefore, a substantial improvement in sensitivity and specificity is expected of SPOT-based discrimination algorithms for VT/SVT discrimination, which should result in a more comfortable therapy and an improved quality of life for defibrillator-implanted patients.
[ Retour à la liste des publications / Back to the list of publications ]
Hai-Ni Qu, Y. Oussar, G. Dreyfus, Weng Xu
Regularized Recurrent Least Squares Support Vector Machines
International Joint Conference on Bioinformatics, Systems Biology and Intelligent Computing, Shanghai, 2009
Abstract
Support vector machines are widely used for classification and regression tasks. They provide reliable static models, but their extension to the training of dynamic models is still an open problem. In the present paper, we describe Regularized Recurrent Support Vector Machines, which, in contrast to previous Recurrent Support Vector Machine, models, allow the design of dynamical models while retaining the built-in regularization mechanism present in Support Vector Machines. The principle is validated on academic examples; it is shown that the results compare favorably to those obtained by unregularized Recurrent Support Vector Machines and to regularized, partially recurrent Support Vector Machines.
Cliquez ici pour obtenir le document / Click here to download the .pdf. document
[ Retour à la liste des publications / Back to the list of publications ]
M. Pernot, E. Macé, R. Dubois, M. Couade, M. Fink, M. Tanter,
Mapping myocardial elasticity changes after RF-ablation using Supersonic Shear Imaging,
Computers in Cardiology, Park City (2009).
Abstract
Supersonic Shear Imaging (SSI) is a new ultrasound-based technique for imaging non-invasively and quantitatively the elastic modulus of soft tissues. Monitoring tissue stiffness changes during Radio-Frequency Ablation (RFA) may quantify the size and shape of the ablation necrosis and therefore assesses if the RFA is complete. We propose to apply SSI for monitoring the myocardial elasticity and evaluate the correlation with the RF Ablation necrosis size in both in vitro and in vivo experiments.
R. Dubois, P. Roussel, M. Hocini, F. Sacher, M. Haïssaguerre, G. Dreyfus,
A Wavelet Transform for Atrial Fibrillation Cycle Length Measurements,
Computers in Cardiology, Park City (2009).
Abstract
We describe a new algorithm for the estimation of Cycle Lengths (CL) in the atria. In the spirit of wavelet transforms, the algorithm correlates the EMG signal to a set of functions that are specifically designed to extract the cycle length present in the signal. This provides a CL vs time map, which is a highly informative representation of the electrical activation of the tissue. Subsequently, the information from this map is compressed into a histogram that unravels the distribution of the dominant CLs on a given time window. Finally, a sliding window tracks automatically the changes in CLs over a large time scale. Results on both synthetic and real data are presented. The correlation with known cycle lengths in the synthetic cases is strong, and the CL distributions on real data are similar to those obtained from manually annotated EGMs.
R. Dubois, P. Roussel, M. Vaglio, F. Extramiana, F. Badilini, P. Maison-Blanche, G. Dreyfus,
Efficient modeling of ECG waves for morphology tracking,
Computers in Cardiology, Park City (2009).
Abstract
We propose a new approach to fully automatic ECG wave extraction and morphology tracking. It is based on Generalized Orthogonal Forward Regression (GOFR), which allows decomposing a one-dimensional signal into a set of appropriate parameterized functions. Two applications of GOFR to ECG modeling are presented. First, in order to delineate ECG characteristic waves, we make use of a specific function, called the Gaussian Mesa function (GMF). Secondly, we track the evolution of the T-wave morphology by introducing a Bi-Gaussian function (BGF).
The approach was validated on three experimental settings; the results confirm that the combination of GOFR and of an appropriate parametric function is remarkably efficient for ECG wave modeling.
P. Bouchet, R. Dubois, C. Henry, P. Roussel, G. Dreyfus,
Spatial Projection of Tachycardia Electrograms for Morphology Discrimination in Implantable Cardioverter Defibrillator,
Computers in Cardiology, Park City (2009).
Abstract
Discrimination of Ventricular Tachycardia (VT) from Supra-Ventricular Tachycardia (SVT) remains a major
challenge for appropriate therapy delivery in Implantable Cardioverter Defibrillators (ICDs), especially in single
chamber devices. We propose here a new discrimination algorithm that analyzes, with a machine learning approach, the morphology of a two-dimensional representation
of both a far-field and a near-field ventricular sensing channel. Features extracted in this representation allow
comparisons between curves. Thus, arrhythmia discrimination is performed by comparing an arrhythmia curve to a reference curve. A statistical classifier was trained on a private database and tested on the standard Ann Arbor Electrogram Libraries.
Our discrimination algorithm demonstrated high sensitivity and specificity for VT/SVT discrimination. The
requirements of this algorithm make it appropriate for implementation in the simplest ICD system.
T. Hueber, E. Benaroya, G. Chollet, B. Denby, G. Dreyfus, M. Stone
Visuo-Phonetic Decoding using Multi-Stream and Context-Dependent Models for an Ultrasound-based Silent Speech Interface,
Interspeech, Brighton (2009).
Abstract
Recent improvements are presented for phonetic decoding of continuous-speech from ultrasound and optical observations of the tongue and lips in a silent speech interface application. In a new approach to this critical step, the visual streams are modeled by context-dependent multi-stream Hidden Markov
Models (CD-MSHMM). Results are compared to a baseline system using context-independent modeling and a visual feature fusion strategy, with both systems evaluated on a onehour, phonetically balanced English speech database. Tongue and lip images are coded using PCA-based feature extraction techniques. The uttered speech signal, also recorded, is used to initialize the training of the visual HMMs. Visual phonetic
decoding performance is evaluated successively with and without the help of linguistic constraints introduced via a 2.5kword decoding dictionary.
Keywords: silent speech interface, visual speech recognition, multi-stream modeling
Cliquez ici pour obtenir le document / Click here to download the .pdf. document
[ Retour à la liste des publications / Back to the list of publications ]
M. Toukourou, A. Johannet, G. Dreyfus,
Flash Flood Forecasting by Statistical Learning in the Absence of Rainfall Forecast: a Case Study,
Engineering Applications of Neural Networks EANN 2009, Londres (2009).
Abstract
The feasibility of flash flood forecasting without making use of rainfall predictions is investigated. After a presentation of the “cevenol flash floods“, which caused 1.2 billion Euros of economical damages and 22 fatalities in 2002, the difficulties incurred in the forecasting of such events are analyzed, with emphasis on the nature of the database and the origins of measurement noise. The high level of noise in water level measurements raises a real challenge. For this reason, two regularization methods have been investigated and compared: early stopping and weight decay. It appears that regularization by early stopping provides networks with lower complexity and more accurate predicted hydrographs than regularization by weight decay. Satisfactory results can thus be obtained up to a forecasting horizon of three hours, thereby allowing an early warning of the populations.
Cliquez ici pour obtenir le document / Click here to download the .pdf. document
[ Retour à la liste des publications / Back to the list of publications ]
Dernière modification / Last update :