Ministère délégué à la Recherche et aux Nouvelles Technologies

- Direction de la Recherche -

Action Concertée Incitative

Neurosciences intégratives et computationnelles

Appel à propositions 2003

Les réponses se conformeront au plan proposé.

Les dossiers, rédigés en anglais et accompagnés d'un résumé d'une page en français, devront être envoyés, sous forme de document papier, en 4 exemplaires (ainsi que 15 exemplaires des fiches résumé) à l’adresse suivante :

Ministère délégué à la Recherche et aux nouvelles Technologies

Direction de la Recherche

Cellule ACI

ACI Neurosciences Intégratives et Computationnelles

1, rue Descartes

75231 PARIS Cedex 05

Ils seront également fournis sous forme de fichier électronique, format RTF, adressés à :

Cellule.aci@recherche.gouv.fr

Date limite de réception des documents

L’exemplaire électronique devra être parvenu au Ministère

au plus tard le 14 mars

La date limite pour les versions papier (4 exemplaires complets du dossier, 15 exemplaires des fiches résumé français/anglais) est le 14 mars, la date du cachet de la poste faisant foi.

Des renseignements peuvent être obtenus en envoyant un courrier électronique à l’adresse suivante :

Cellule.aci@recherche.gouv.fr

Constitution du dossier

Dossier complet agrafé incluant dans l'ordre suivant :

les fiches résumé en anglais et en français, servant de page de garde au dossier;

le descriptif scientifique, 6 pages maximum pour un projet, 3 pages pour un pré-projet ;

les renseignements administratifs ;

l'estimation financière.

Dossier à envoyer en 4 exemplaires papier, 1 exemplaire version électronique + 15 exemplaires des fiches résumé

Neurosciences intégratives et computationnelles

Appel à propositions 2003

Thème : Cohérence et intégration dans les réseaux neuronaux

Fiche résumé

(15 exemplaires)

Titre du projet : RIVAGe
 Rétroaction lors de l'Intégration Visuelle: vers une Architecture Générique

Durée du projet  : préprojet sur 1an  (ayant vocation à aboutir à un projet sur deux ans en 2004)

Mots-clefs  : Cortex visuel, Rétroaction, Modélisation, Simulation

Responsable scientifique : FAUGERAS Olivier , Responsable du Projet Odyssee,

INRIA BP93 06 902 Sophia, +33 4 92 38 78 30 fax + 33 4 92 38 78 45, Olivier.Faugeras@sophia.inria.fr

Discipline du responsable scientifique :

      Vision par Ordinateur et Mathématiques appliquées

Organisme demandeur  (gestionnaire) :

INRIA Sophia 2004 rt des Lucioles  BP93 F-06902 Sophia, +33 4 92 38 78 30 fax + 33 4 92 38 78 45

Noms et coordonnées des équipes partenaires éventuelles :

CERCO Univ Paul Sabatier, 133 rt de Narbonne, F-31063 Toulouse Cdx, Tel : +33 5-62-17-37-75

ENPC Ecole Nationale des Ponts et Chaussées,
                                          CitéDescartes, F-77455 Marne La Vallée Cedex 2 Tel: +33 1 64 15 35 72

ENS Ecole normale supérieure, 45, rue d'Ulm, F-75230 Paris cedex 05 Tel: +33 1.44.32.21.54

Disciplines couvertes par ces équipes partenaires :

Structure, fonction et développement du cortex visuel

Modélisation de l'activité cérébrale

Résumé du projet: Le but de ce pré-projet est de construire une relation étroite entre un laboratoire des neurosciences de la perception visuelle d’une part et de la vision artificielle (aussi appelé vision par ordinateur) d’autre part. Le but à long terme est d’aboutir à une théorie commune débouchant sur des questions précises dans le champ des neurosciences et des architectures et des algorithmes utilisables en vision artificielle et ses applications.

   On s'intéresse à l'étude comparative de l'intégration des processus visuels au sein de deux systèmes : l'un biologique (précisément les voies pariéto-ventrales et pariéto-dorsales du système visuel cortical chez le primate) et l'autre artificiel. Ce dernier délivre des estimations: « where » du mouvement et de la structure de la scène observée et « what » du groupement perceptuel et de l’identification d'objets de la scène. Dans ce cadre, le rôle et le fonctionnement des mécanismes adaptatifs de rétroaction est au centre des recherches actuelles dans le domaine biologique.
   Au coeur de cette étude est l'idée que le traitement visuel effectue :
                   -  une première vague de calculs qui permet de détourer l'information reçue, de fournir une estimation initiale, de faire des hypothèses sur les objets de la scène, etc..
                   - un raffinement  itératif pour arriver à une perception effective de la scène.
   De tels mécanismes adaptatifs sont omniprésents, parfois implicitement, dans les processus de vision artificielle où ils sont bien souvent traités dans le cadre du calcul des variations qui permet, outre de garantir que les problèmes sont mathématiquement bien posés, de répondre à des questions très importantes relatives à la  fusion d’informations telles que  (i) la combinaison d'attributs visuels calculés par des modules différents et utilisés ensuite pour les tâches de « what » et « where »et (ii) l'utilisation d'information a priori issue de modules visuels de plus haut niveau. Ce sont des « modèles » des données à traiter, ces modèles étant soit donnés a priori  (e.g. rigidité, régularité de forme, etc..),  soit issus de l’étiquetage d'objets reconnus par « la première vague » de calculs du système.
   Dans le cadre du thème 3 de cette ACI nous proposons un pré-projet: une étude théorique en trois étapes
         (a) une analyse systématique de la littérature en neurosciences portant sur les mécanismes adaptatifs de rétroaction au sein du cortex visuel.
         (b) une interprétation des résultats de cette analyse à la lumière des apports de l'approche variationnelle utilisée en vision arficielle pour traiter ces questions.
         (c) une définition du cahier des charges d'un outil de simulation informatique de certains aspects du fonctionnement du cortex visuel, qui pourrait être réalisé dans une seconde phase après ce pré-projet.
Action Concertée Incitative

Neurosciences intégratives et computationnelles

Appel à propositions 2003

Summary

Title of the project: RIVAGe

Feedback during Visual Integration : towards a Generic Architecture

Duration of the project: 1 year pre-project (with the goal of a two years project in 2004)

Key words : Visual cortex, Feedback, Modelization, Simulation

Scientific coordinator :  FAUGERAS Olivier , Odyssee, Research team leader

INRIA BP93 06902 Sophia, +33 4 92 38 78 30 fax + 33 4 92 38 78 45, Olivier.Faugeras@sophia.inria.fr

Field of expertise:  Computer Vision and Applied Mathematics

Institution (managing the project) :

INRIA Sophia 2004 rt des Lucioles  BP93 F-06902 Sophia, +33 4 92 38 78 30 fax + 33 4 92 38 78 45

Names and addresses of scientific partners:

   CERCO Univ Paul Sabatier, 133 rt de Narbonne, F-31063 Toulouse Cedex, Tel : +33 5-62-17-37-75

   ENPC Ecole Nationale des Ponts et Chaussées, 6 av Blaise Pascal, F-77455 Marne la Vallée, +33 1.64.15.36.64

   ENS Ecole Normale Supérieure, 45, rue d'Ulm, F-75230 Paris cedex 05 Tel: +33 1.44.32.21.54

Partners field of expertise:

  Development, function and structure of the visual cortex   

  Modelization of the cortical activity

Project summary:  The goal of this pre-project is to build a strong relationship between

a research team working in neurosciences of the visual perception and a research team working in artificil vision (i.e. computer vision). The long term objective is to elaborate a common theory about precise questions in both neurosciences and algorithms and their architecture in artificial vision, including computer vision applications.

We consider the comparative study of visual process integration within either a biological system, i.e. the parieto-ventral and parieto-dorsal pathways of the cortical visual system in the primate or an artificial system. Both systems deliver an estimation of [where], that is to say the motion and structure of the observed scene and of [what], i.e. the perceptual grouping and labeling of objects in the scene. Within this framework, the function and behavior of adaptive feedback

mechanisms is a key point and on the leading edge of biological studies. The core of this idea is that visual processing is build around:

1) a first computational step allowing to pre-process the input information, provide initial estimates, generate hypotheses about which models to use, …

2) a  refinement step using iterative mechanisms of optimization of the visual perception.

      Such mechanisms occur, sometimes implicitly, in artificial vision processes. They are mainly related to such problems as the combination of visual attributes, computed from different sources and then fused for ``what'' and ``where'' perceptual tasks; the use of  a-priori information, obtained from higher-level visual modules. These modules define models estimated from the data. These models are  either given  a-priori (e.g. rigidity, shape regularity, ..) or chosen thanks to object labeling obtained during the first computational step.

Within the scope of the 3rd topics of this ACI, this project consists in three steps:

       (a) a systematic analysis of existing results in neuro-science,

       (b) an interpretation of these results from the viewpoint of the variational approach widely used in computer vision

       (c) a specification of a simulation tool of parts of the visual cortex,  the actual development of this simulator being the goal of a second phase after this pre-project.

       
Action Concertée Incitative

Neurosciences intégratives et computationnelles

Appel à propositions 2003

Descriptif scientifique / Scientific description

1. Situation actuelle du sujet proposé / Position of the problem

The interactions between the communities of researchers studying the biological bases of vision and those interested in developing artificial vision systems have gone up and down during the last twenty years. In the years 1980's, there was a consensus between the two communities highlighted by the publication of the book Vision by David Marr [16]. In those days, processing of visual information was thought to be done in a series of steps, progressing from a local analysis of borders to 3D surface elements and then to the identification of 3D objects. This model was consistent with the hierarchy of cortical areas that has just been proposed by Maunsell and Van Essen [8] : borders were analyzed in area V1 and successively more elaborate degrees of processing were achieved in the higher order areas.

Twenty years later, the situation has changed considerably : biological visual systems can no longer be considered as purely feedforward systems consisting of sets of filters of progressively higher degree of sophistication as one penetrates the different layers of processing. In the field of artificial vision, numerous other models have been proposed to process information. These models work well within their context but a general framework for processing visual information is no longer in sight. The purpose of this pre-project is to examine the possibilities of laying new foundations for interactions between the two communities.

Two of the major advances in the field of biological vision in recent years have been the realization that 1) information does not progress only bottom-up but that there is a very dense network of top-down connections and 2) that there are in the brain internal representations of the external world that are embodied in neuronal networks. This change in perspective has lead biologists to view the visual system no longer as processing information transferred by the reti, nal ganglion cells (the computer analyzing the output of the camera), but more as processing its own internal representations and checking whether these representations are consistent the messages sent by the sensory neurons (checking the validity of the models). It is therefore of major importance to evaluate the role of the top-down (also called feedback) connections that are likely to be used to compare the representations of higher levels with those located at lower levels of processing.

The roles attributed to top-down influences are numerous but very few have been tested in detail : feedback connections are thought to be involved in directing attention, in memory retrieval, in comparing internal models with sensory inputs, in combining global and local processing. Our intention is to examine the role of top-down influences not in isolation but within a general model of the brain such as that proposed by Friston [9]. This model stresses the point that feedback connections are essential when the relationship between retinal inputs and the stimuli that generate them is not invertible. This is the case for practically all situations of vision outside the laboratory because of the interactions between the stimuli and the importance of the context for a given stimulus. To resolve this difficulty, Friston proposes that the feedback connections send back information to the early levels of processing and that this information is compared to the input vectors. Thus, as in the model of Rao and Ballard [22], the way by which internal representations and incoming signals are combined is mainly subtractive : only differences are transmitted to higher levels. Although such models are popular among theoreticians, they do not fit with the results of biological experiments : all studies of feedback connections so far [4] converge to conclude that feedback influences act to potentiate the responses of neurons at lower levels. One of the goals of our collaboration will therefore be to develop models that incorporate the notions of internal models, non invertible relationship and interactions between top-down and bottom-up influences but are more compatible with results from biological vision, in the hope that this search might reveal some particularly interesting strategies used by biological systems to solve the problems posed by Friston [9].

Another recent important change in biological vision is the realization that information is not processed in a step-by-step hierarchical fashion but that some higher order areas, that contain sophisticated representations, are activated extremely early in the processing of visual information. In particular, it has been shown that the entire dorsal stream, including the parietal cortex and the frontal cortex are activated only a few millisecond after area V1 which constitutes the initial processing stage of vision in primates [26]. This new perspective has led to a model in which a first-pass analysis is done by neurons in the dorsal stream and that this first-pass analysis is used, through feedback connections, to optimize the more detailed processing that occurs later (detailed shape, color). One of the goals of our collaboration is to determine whether such a model could be used in the framework of a general model of the visual system to improve and accelerate the processing of visual stimuli.

We consider the topic ``integrated model of visual processing''. Regarding the hierarchical organization of the visual cortical areas of the primate (see e.g. [27] for a comparative description of the organization of visual areas in macaque and human cerebral cortex) this corresponds to two main streams: the ventral and dorsal considered as the ``what'' (i.e. object recognition) and ``where'' (i.e. object localization and motion analysis) processing streams, respectively. Our goal is to contribute to models of biologically plausible neural computations (see e.g. [3] for details about the biological plausibility of linear and multiplicative computational steps, including motion detection and short-term memory, while [29] focuses on the biologically plausible implementation of non-linear operation such as minimum computation or comparisons between inputs).

Following [4] we emphasize the fact that cortical visual processing requires information to be exchanged between neurons coding for distant regions in the visual field. Feedback connections from upper-layers are best candidates for such interactions because magnocellular layers of the LGN very rapidly project a ``first-pass'' information used to guide further processing. More precisely, [2] demonstrate that the so called ``horizontal connections'' (i.e. within a cortical, e.g. retinotopic, map) are not fast enough to account for such a transfer, considering the known timings of information transfer [19].

(1) These cortical computations include stereo/binocular disparity. The work of [17] is a recent attempt to explain how complex cells can issue depth percepts for binocular but also monocular (i.e. da Vince stereopsis) cues, including induced effects from contrast changes (i.e. shape from illumination) as simulated in [12]. This also includes disparity tuning: [14] develops a model explaining how the LGN is involved in binocular disparity tuning, matching left and right images with the same contrast polarity but inducing feedback adaptation from signals of opposite polarities. In the present study, we would like to get a step further, considering V3 computations. More precisely, [1] shows the existence of disparity-selective columns including occlusions and proposes that V3 contributes to the processing of stereoscopic depth information and that the parietal areas to which it projects use this information for object depth and 3D shape analysis.

(2) A second aspect is perceptual grouping. For instance, [13] attempts to demonstrate how the known laminar architecture of the V1 and V2 areas of the visual cortex, assuming a functional role for this stratification, is involved in percepts generation, see also [18]. This includes pre-attentive/attentive aspects, as pointed out by [21] describing how the parvocellular stream of the visual cortex performs visual filtering, i.e. attention and perceptual grouping, using feed-forward, feedback and horizontal interactions. One key aspect is the figure / ground segmentation in the brain, as recently reviewed by [24] showing that shape perception depends critically on this cue, while its link to early-vision mechanisms is now relatively well understood [25] e.g. the role of junctions in surface completion and contour matching.

(3) Motion processing is also a key problem. It has been shown in [6] how responses to moving stimuli are derived from transient cell, speed-tuning cell responses of different sizes yielding visual motion perception and speed discrimination, involving both ventral and dorsal visual pathways. At a cognitive level, considering interpolation between prototypes of gestures, [11] attempts to interpret how the cortical ``what'' and ``where'' pathways of the visual cortex are involved in complex movements patterns. For instance, a complex spatio-temporal pattern is efficiently represented as a combination of prototypes whose coefficients are estimated using variational estimation methods [10]. Well-known early vision visual cues (e.g. the Koenderinck def value easily computed by an affine modelization of the retinal motion field [5]) are other relevant candidates for such a parameterization.

All three previous aspects of visual perception are easily formalized using the ``generative model'' approach developed in [7] and reviewed in [9]. This allows to relate what is actually known in statistical modelization with existing biological models. A general framework for the role of feedback connections is proposed in these publications. A step further, the link with variational approaches [28] has been already sketched out. Regarding the role of feedback connections, [22] describes a model of visual processing in which feedback connections from a higher- to a lower- order visual cortical area carry predictions of lower-level neural activities, whereas the feed-forward connections carry the residual errors between the predictions and the actual lower-level activities. In the scope of this approach receptive fields are Gabor-like filters with a sigmoid profile output, weights being optimized by a 1st order gradient optimization of a likelihood criterion. This is easily generalized to more plausible operators. But the relation with sensory-motor strategies, as discussed in [23] is very important: these authors point out the fact that visual cognition depends heavily on the gaze orientation mechanisms. They also analyze how the appearance information contained in the image is converted into a target position, using saliency maps and separating targeting ``what'' process and the localization ``where'' process. This is in close agreement with the functionalities of the dorsal (where) and ventral (what) visual cortical pathways.

As a conclusion, this first reading of the literature indicates that there are many but rather different frameworks to model the ventral/dorsal visual pathways functionalities, while these correspond to recent computer vision mechanisms, now formalized in a unifying ``variational'' framework (e.g. [15,20] for recent introductions). It is thus very tempting to investigate whether this unifying framework could also be relevant in modeling biological vision. This is the goal of this project.

Bibliography

[1]    D. Adams and S. Zeki.    Functional organization of macaque V3 for stereoscopic depth.    J. Neurophysiol., 86:2195-2203, 2001.

[2]    A. Angelucci and J. Bullier.    Reaching beyond the classical receptive field of V1 neurons: horizontal or feedback axons?    J. Physiol. (Paris), 2002.

[3]    G. Bugmann.    Biologically plausible neural computation.    Biosystems, 40:11-19, 1997.

[4]    J. Bullier.    Integrated model of visual processing.    Brain Res. Reviews, 36:96-107, 2001.

[5]    C. Caudek and N. Rubin.    Segmentation in structure from motion: modeling and psychophysics.    Vision Research, 41:2715-2732, 2001.

[6]    J. Chey, S. Grossberg, and E. Mingolla.    Neural dynamics of motion processing and speed discrimination.    Vision Res., 38:2769-2786, 1997.

[7]    P. Dayan and L. F. Abbott.    Theoretical Neuroscience : Computational and Mathematical Modeling of Neural Systems.    MIT Press, 2001.

[8]    D. V. Essen and J. Maunsell.    Hierarchical organization and functional streams in the visual cortex.    Trends in Neurosciences, 6(9), 1983.

[9]    K. Friston.    Functional integration and inference in the brain.    Prog Neurobiol, 68:113-143, 2002.

[10]    M. Giese and M. Lappe.    Measurement of generalization fields for the recognition of biological motion.    Vision Research, 38:1847-1858, 2002.

[11]    M. Giese and T. Poggio.    Neural mechanisms for the recognition of biological movements and actions.    Nature Neuroscience, 2003.    in press.

[12]    S. Grossberg and N. McLoughlin.    Cortical dynamics of three-dimensional surface perception: binocular and half-occluded scenic images.    Neural Networks, 10(9):1583-1605, 1997.

[13]    S. Grossberg, E. Mingolla, and W. D. Ross.    Visual brain and visual perception: how does the cortex do perceptual grouping?    Trends in Neurosciences, 20(3):106-111, 1997.

[14]    A. Grunewald and S. Grossberg.    Self-organization of binocular disparity tuning by reciprocal corticogeniculate interactions.    Journal of Cognitive Neuroscience, 10:199-215, 1998.

[15]    F. Guichard and J.-M. Morel.    Image analysis and P.D.E.'s.    Tutorials on Geometrically Based Motion, IPAM, Ucla, Los Angeles, 2001.

[16]    D. Marr.    Vision.    W.H. Freeman and Co., 1982.

[17]    N. McLoughlin and S. Grossberg.    Cortical computation of stereo disparity.    Vision Res, 38(1):91-99, 1998.

[18]    H. Neumann and E. Mingolla.    Computational neural models of spatial integration in perceptual grouping.    In T. . P. Kellman, editor, From Fragments to Objects: Grouping and Segmentation in Vision, pages 353-400. Amsterdam: Elsevier, 2001.

[19]    L. Novak and J. Bullier.    The Timing of Information Transfer in the Visual System, volume 12 of Cerebral Cortex, chapter 5, pages 205-241.    Plenum Press, New York, 1997.

[20]    S. Osher and R. P. Fedkiw.    Level set methods : overview and recent results.    Tutorials on Geometrically Based Motion, IPAM, Ucla, Los Angeles, 2001.

[21]    R. Raizada and S. Grossberg.    Towards a theory of the laminar architecture of the cerebral cortex: Computational clues from the visual system.    Cerebral Cortex, 13:100-113, 2003.

[22]    R. Rao and D. Ballard.    Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects.    Nat Neurosci, 2(1):79-87, 1999.

[23]    R. Rao, G. Zelinsky, M. M. Hayhoe, and D. Ballard.    Eye movements in iconic visual search.    Vision Research, 42(11):447-1463, 2002.

[24]    N. Rubin.    Figure and ground in the brain.    Nature Neuroscience, 4:857-858, 2001.

[25]    N. Rubin.    The role of junctions in surface completion and contour matching.    Perception, 30:339-366, 2001.

[26]    S. Thorpe and M. Fabre-Thorpe.    Seeking categories in the brain.    Science, 291:260-263, 2001.

[27]    D. Van-Essen.    Organization of visual areas in macaque and human cerebral cortex.    In L. Chapula and J. Werner, editors, The Visual Neurosciences. MIT Press, 2003.

[28]    T. Vieville.    Biologically plausible regularization mechanisms.    RR 4625, INRIA, 2002.

[29]    A. J. Yu, M. Giese, and T. Poggio.    Biophysiologically plausible implementations of maximum operation.    Neural Computation, 14(12), 2003.

The variational approach: a unifying theoretical framework to analyze visual processes

The adaptive mechanisms described previously are always present -most of the time implicitly- in artificial visual processes. The so-called ``variational approach'' is a unifying theoretical framework to design and implement this formalism.

At a phenomenological level, this framework :

  - defines the estimation problem in terms of the optimization of a criterion. This criterion is usually built from two terms:

 (i) One term is related to the data input (e.g. looking for a solution as compatible as possible with this input)

 (ii) One term is related to the a-priori information (i.e. looking for a solution corresponding to plausible properties allowing to regularize the solution)

 - implements this global optimization using a local iterative scheme of a parametric or non-parametric map.

  This scheme arises from the partial differential equations (PDEs) that must be verified by the solution.

  The architecture of this implementation may correspond to what is processed in a cortical [maxi]column: here is the fundamental motivation to apply this formalism to neuronal computations, since we assume that it provides an alternative to usual artificial ``neural-nets'' as a basic generic biologically plausible estimation process.

More precisely, such a mechanism allows to model how a visual system, as a whole, solves perceptual tasks not only at a ``pixel'' level. A key aspect is that, under this assumption, there is a direct link between PDE parameterization and spatio-temporal maps of cortical activity. Therefore we plan to find biological networks of neurons which carry out some of the PDEs computations used in algorithmic vision and conversely, starting from neuro-physiological data, to try to define pertinent PDEs acting on visual inputs.

This should allow the development of not only anatomical but also functional models of activity in cortical areas during a visual task.

At a theoretical level, the variational paradigm is interesting because:

 - it provides concise mathematical models of many computer vision problems,

 - it allows to study in detail the problems of the existence and the uniqueness of solutions of the resulting equations, including wellposedness and

 - to design correct and usually efficient algorithms to calculate approximate solutions of these equations. 

Main publications of the partners related to this subject:

Hupé, J. M., James, A. C., Payne, B. R., Lomber, S. G., Girard, P., and Bullier, J. (1998) Cortical feedback improves discrimination between figure and background by V1, V2 and V3 neurons. Nature 394 : 784-787.

Hupé, J.M., James, A.C., Girard, P., Lomber, S., L., Payne, B., and Bullier, J. (2001) Feedback connections act on the early part of the responses in monkey visual cortex. J. Neurophysiol. 85, 134-145.

Angelucci, A., Levitt, J.B., Walton, E.J.S., Hup?, J.-M., Bullier, J., Lund, J.S. (2002) Circuits for local and global signal integration in visual cortex. J. Neurosci. 22: 8633-46.

Bullier, J. (2001) Integrated model of visual processing. Brain Res. Reviews, 36:96-107.

L. Novak and J. Bullier. (1997) The Timing of Information Transfer in the Visual System, Cerebral Cortex, vol 12, chapter 5, pages 205-241. Plenum Press, New York.

Rousselet GA, Fabre-Thorpe M, Thorpe SJ. (2002. Parallel processing in high-level categorization of natural images. Nat Neurosci 5: 629-30

Mikael Rousson, Thomas Brox and Rachid Deriche (2003) "Active Unsupervised Texture Segmentation on a Diffusion Based Feature Space", IEEE Conference on Computer Vision and Pattern Recognition, Madison, Wisconsin, USA,

Vieville, T., & Faugeras, O. (2002). La longue marche vers la vision cognitive. La Recherche, 2.

Vieville, T., & Crahay, S, O. (2003). A deterministic biologically plausible classifier. J.Computational

      Neuroscience, in review

Vieville, T, Lingrand, D. & Gaspard, F. (2001). Implementing a multi-model estimation method, International Journal of Computer Vision, 44.

Hermosillo, G., Chefdhotel, C. & Faugeras, O. (2002). Variational methods for multimodal image matching. International Journal of Computer Vision, 50, 329-343.

Thirion, B. & Faugeras, O. (2002). Fmri data modeling: from linear to functional dependence, in a multivariate framework. Proc. of the 8th Int. Conf. on Functional Mapping of the Human Brain.

2.  Description du pré-projet / Pre-project description

Purpose Within the scope of the 3rd topic of this ACI, this project consists in three steps:

       (a) a systematic analysis of existing results in neuro-science related to the feedback mechanisms in the visual cortex,

       (b) an interpretation of these results from the viewpoint of  the variational approach widely used in computer vision

       (c) a specification, based on the previous two steps, of a simulation tool of parts of the visual cortex,  the actual development of this simulator being the goal of a second phase after this pre-project.

Methods The partners will organize a working group for literature reading and criticizing, exchange of ideas, development of a set of concepts common to biological and artificial vision. Twice a month joint working sessions with oral presentations will be organized. Geographically this group is scattered over four sites. The work has already started and the method been validated. The work of  Friston, Dayan and Abott and Grossberg et al. will be analyzed first (see the enclosed bibliography).

In parallel with this activity, computer scientists in the pre-project will start implementing and simulating the most promising and relevant models proposed in the literature in order to understand to what extent they can predict experimental neuro-physiological or psycho-physiological data. These pieces of code are not to be considered as software development, but preliminary prototypes for feasibility purposes.

To prepare for this part of the work, a young computer engineer has done an extensive study of existing non-commercial software tools, and performed detailed evaluations on well-known (actually V1 and MT) neuronal computations.

3. Conséquences attendues (valorisation) / Outcome and valorisation

 {A} The development of a strong synergy between a laboratory in visual neurosciences and one in computational vision.

 {B} A review of the formalisms used by the visual neuroscience community in order to model and predict experimental data related to the visual cortex. Interesting consequences may be an evaluation of their relevance, a better understanding of their intrinsic complexity, and the construction of a mathematical typology.

 {C} An identification of the French and European emerging entities in this area, with the objective to join or create at the European level a multidisciplinary community in biological/computational vision.

 {D} A conclusion about whether or not the variational approach may provide a unifying formalism to describe the behavior and functionality of visual cortical layers. Although quite promising, this idea is still in a very preliminary stage and needs to be confronted to hard biological facts, this pre-project is precisely centered around this question.

 {E} The specification of a software tool for the simulation of ``macroscopic'' aspects of visual processing in the cortex, as detailed previously in this proposal. The precise mathematical framework of this simulator will heavily depend on the outcome of {D} and this is the reason why we want to proceed in two steps with a pre-project possibly followed by a full-project.

Ministère délégué à la Recherche et aux Nouvelles Technologies

- Direction de la Recherche -

Action Concertée Incitative

Neurosciences intégratives et computationnelles

Appel à propositions 2003

Renseignements administratifs

(4 exemplaires)

Nom du responsable scientifique :          FAUGERAS Olivier

Etablissement dont relève le responsable scientifique : INRIA

Laboratoire : INRIA Sophia, Projet Odyssee http://www-sop.inria.fr/odyssee

Directeur du laboratoire (nom, prénom et signature) : COSNARD Michel

Adresse complète : INRIA 2004 rt des Lucioles, BP93 06 902 Sophia

Téléphone :      +33 4 92 38 78 30                                    Télécopie :  + 33 4 92 38 78 45

Adresse électronique : Thierry.Vieville@sophia.inria.fr

Etablissement gestionnaire de l’opération

Nom :

INSTITUT NATIONAL DE RECHERCHE EN INFORMATIQUE ET EN AUTOMATIQUE

Statut juridique : EPST

Adresse complète : INRIA, Sophia 2004 rt des Lucioles, BP93 06 902 Sophia

Téléphone :      +33 4 92 38 77 77                           Télécopie :  + 33 4 92 38 78 45

Signature du représentant de l’organisme gestionnaire :

Composition de l’équipe du responsable

(*) Ces 5 chercheurs travaillent en Vision par Ordinateur avec des spécialités explicités ici.

Composition des autres équipes participant au programme de recherche

(*) L'équipe de recherche Odyssee est un projet commun entre l'INRIA, l'ENS et l'ENPC

Ministère délégué à la Recherche et aux Nouvelles Technologies

- Direction de la Recherche -

Action Concertée Incitative

Neurosciences intégratives et computationnelles

Appel à propositions 2002

Estimation financière

(4 exemplaires)

Titre du projet : RIVAGe

Rétroaction lors de l'Intégration Visuelle: vers une Architecture Générique

Nom du responsable scientifique : FAUGERAS Olivier

Demande financière : (Indiquer les montants en euros)

      Fonctionnement (TTC)

frais de laboratoire                   aucun    

prestations de service        aucune

frais de mission  (1)                4000 euros

frais de gestion                pris en charge par les organismes

Equipement (TTC) (2)              6000 euros                                                                                                

Montant total de l'aide demandée (TTC)  10000 euros

Autres financements du projet

demandés : post-doc pour l'étude et le développement des outils de simulation (2^nd phase du projet)

obtenus : projet Amiria / Robea, collaboration entre le Cerco et l'INRIA

http://www-sop.inria.fr/odyssee/contracts/robea/amiria.pdf

(sur la modélisation de la classification d'objets)

(1) Détails des missions:

           Séjour d'un jeune chercheur du CERCO à l'INRIA :  10 jours x 50 e/j + 200 e voyage = 700 e

           Séjours de travail de chercheurs CERCO <-> INRIA :  10 x (2 jours x 50 e/j + 200 e voyage)  = 3000 e

           Participation à un événement scientifique (congrès, .. ) d'un représentant du projet : 300e

(2) Equipement :.

Achat d'ouvrages pour l'étude bibliographique : 12 x environ 40 e = 500 e

Equipement de visio-conférence sur trois sites (le 4^ème est équipé) : 3 x 500 e = 1500 e

Deux stations de travail pour le déploiement des outils de télétravail et de simulation : 2 x 2000 e