Atelier sur la genèse de la perception et de la notion d'espace chez le robot et l'homme
Workshop on the genesis of perception and the notion of space in machines and humans
jeudi 18 et vendredi 19 octobre 2001
Comité d'organisation : Kevin O'REGAN (LPE, Paris 5) et Jean-Pierre
avec l'aide administrative d'Aline Bompas (LPE, Paris 5)
Géométrie et Cognition
(Responsable Giuseppe Longo)
groupe de travail du CENECC :
CENtre d'Etudes des systèmes Complexes et de la Cognition
Avec le soutien de l'
Action Cognitique du Ministère de la Recherche
(Directrice Catherine Fuchs)
(Interventions : 40 mn + discussion 10 mn)
jeudi 18 octobre
chairman : JP. Nadal (ENS, Paris)
vendredi 19 octobre
chairman : M. Auvray (LPE, Paris)
"Sensory prostheses to study perceptual learning"
Laboratoire Neurophysique et Physiologie du Système Moteur CNRS FRE 2361
Université René Descartes – UFR Biomédicale des Saints Pères, 45, rue des Saints Pères – 75270 Paris CEDEX 06 (email@example.com)
Abstract: Sensory prosthesis does not substitute a sense to a deficient one. They induce new ways to sense and explore the environment trough the emergence of new sensation/action coupling. This can be achieved only after a learning period that leads to the forgetting (embodiment ?) of the prosthesis. We address the question of the use of sensory prostheses as experimental tools for the study of perceptual learning. We present some preliminary results and discuss the adequacy of classical models of perceptual processes to our observations. We finally emphasize the need of new theoretical issues that could help to account for the behavior of subjects.
"Dispositifs de Substitution Sensorielle visuo-tactiles
et Perception de l’espace: Mesures d’Acuité et Navigation spatiale"
Hervé Segond (1) & Eliana Sampaio (2)
(1) Université Louis Pasteur et Conservatoire National des Arts et Metiers (firstname.lastname@example.org)
(2) Conservatoire National des Arts et Metiers (email@example.com)
Abstract: Les dispositifs de substitution sensorielle non invasifs, destinés à permettre à des sujets aveugles d'accéder à des informations visuo-spatiales par l’intermédiaire de modalités sensorielles autres que la vision, ont considérablement évolué depuis leurs précurseurs que furent la canne blanche et le Braille. Les évolutions scientifiques, technologiques et techniques, se sont caractérisées non seulement par le recours à l’informatique, mais également par la miniaturisation, et l’utilisation de matériaux amagnétiques offrant l’opportunité d’études conjointes au niveau comportemental et au niveau des activations cérébrales.
A l’heure actuelle, on distingue deux catégories essentielles de systèmes de substitution sensorielle: les systèmes visuo-auditifs, qui recueillent de l’information sur l’emplacement et la distance des objets, et les systèmes visuo-tactiles qui fournissent des informations optiques sur la forme des objets et la profondeur (perspective, parallaxe). L’utilisation de ces systèmes, par des sujets aveugles de naissance ou voyants aux yeux bandés, suppose une capacité d’adaptation, ou plasticité, du système nerveux central.
Les recherches menées par notre équipe, dont les efforts se concentrent dans le domaine du développement des systèmes visuo-tactiles, portent sur: 1. les dimensions perceptive et cognitive inhérentes à la perception de l’espace extra-corporel par des systèmes sensoriels non visuels (cf. utilisation immédiate chez le bébé, en raison d’une plasticité cérébrale importante vs. procédures d’apprentissage requises chez l’adulte), 2. la mesure de « l’acuité » visuo-tactile, meilleure que celle permise par le système de vision artificielle développé par l’Institut Dobelle (système invasif connecté directement au cerveau), 3. le développement des changements au niveau cérébral, occasionnés par ces dispositifs (en IRMf), 4. les possibilités de navigation dans l’espace 3D via des informations substitutives visuo-tactiles (dans un labyrinthe réel 3D).
L’ensemble de ces recherches offrent des perspectives nouvelles pour la compréhension des mécanismes intermodaux de détection des invariants perceptivo-moteurs (inhérents au principe de substitution sensorielle et permettant la reconnaissance des objets perçus), l’utilisation précoce de ces dispositifs, chez le bébé, en vue de prévenir les troubles du développement fréquemment observés chez l’enfant aveugle et l’absence de prise en charge, et plus généralement, la possibilité pour des sujets aveugles d’accéder à des informations optiques jusqu’alors inconnues (cf. profondeur).
"Implications of a 'sensorimotor' theory of visual phenomenality
for change blindness, color perception, sensory substitution and the notion
of space and object."
Kevin O'Regan & Aline Bompas
Laboratoire de Psychologie Expérimentale, Université René Descartes, 71 avenue Edouart Vaillant,
Abstract: Under a recent "sensorimotor" theory of visual sensation (O'Regan & Noë, Behavioral and Brain Sciences, 2001, in press), the qualitative nature of our visual experience is constituted by our implicit knowledge of the laws of co-variance linking motor actions and the resulting sensory changes. One of the implications of this way of thinking about vision is that information about a visual scene need not be stored in detail in the brain. All that need be stored is 'recipes' that allow an observer to check that attended objects are actually present. The phenomenon of 'change blindness' is compatible with this. Another application of the theory concerns color vision. When the eye moves, or when an object moves, certain predictable changes occur which are typical of colors. Under the sensorimotor theory, it is the knowledge of such laws which constitute color sensation. If this is correct, it should be possible to modify perceived color by changing these sensorimotor contingencies. Experiments are reported where we attempt to do this.
"I. Deducing the dimensionality of space and the nature
of objects: towards a mathematical formulation. II. Human attempts at learning
the contingencies of space"
David Philipona (1), Malika Auvray (2,3), Kevin O'Regan (3) & J-P Nadal (4)
(1) Ecole Polytechnique (firstname.lastname@example.org)
(2) Babel@stal, 25 rue du Petit Musc, 75004 Paris (email@example.com)
(3) Laboratoire de Psychologie Expérimentale, Université René Descartes, 71 avenue Edouart
Vaillant, 92774 Boulogne-Billancourt (firstname.lastname@example.org)
(4) Laboratoire de Physique Statistique, Ecole Normale Supérieure (email@example.com)
Abstract (à confirmer): We show mathematically that when an observer or robot moves his body and receives sensory input, he can extract from the co-variance of the input-output statistics information about dimensionality of space, and the number of attributes necessary to describe objects within it. We present some preliminary experimental tests where humans are required to learn new sensorimotor contingencies.
"Ways of worldmaking: Exploring perception and
behavior in humans inhabiting simple virtual environments"
Computer Science Department, Electrical Engineering Faculty, SS Cyril and Methodius University in Skopje, Macedonia (firstname.lastname@example.org)
Abstract: The work described here explores the hypothesis that perception is a proactive process, resulting out of the active stay of the agent in a given environment. In our usual environments (me here in the office, looking at the computer monitor, through the window...) where we have long ago learned the sensory-motor contingencies, it is hard to see the numerous processes underlying perception. Therefore, the aim was to create an environment for human subjects where new sensory motor contingencies hold and where in order to perform a task subjects were to learn and exploit them.
Since 1996 we have created several such environments and some are described in (Stojanov, G., Gerbino, W., Investigating Perception in Humans Inhabiting Simple Virtual Environments, European Conference on Visual Perception, Trieste, 1999). The simplest set-up is the following one: human subjects sit in front of the computer and are told that they can perform certain simple actions (finite number of button presses) and that while doing so they would here auditory stimuli (beeps of varios frequences). Their task was to succeed to produce a beep of certain frequence (a goal beep, which is very different from all the other beeps). What subjects don't know is that in fact the button presses move them in a 2D environment (grid) where while in different regions they hear different beeps. Moreover, additional structure is imposed by populating the 2D environment with obstacles which emmit a distinct beep if one tries to go through them. With dimensions of the grid of 20x20 subjects are usually able to produce the goal beep at will ("to come to the goal place") within 30 min.
Apart from the quantitative analysis (e.g. how the number of button presses decrease after each successful trial) what is interesting is to explore the strategies people use to solve the task. After the initial seemengly random button presses, they seem to impose some structure in the flow of the incoming beeps by trying to repeat certain sequences of their actions. Indeed, while doing so, they usualy end in some sort of repetitive cycles of button presses and beeps, and when they "break the cycle" and again try the same button sequences end up in another cycle. Eventually, when the cycle they are currently in, contains the goal beep, the subjects seem to perform a kind of backward chaining learning of the particular ordering among the cycles. So, when put in a random place of the 2D environment, they "situate" themselves in some of these cycle and "go" to the goal region.
Despite of the spatial metaphors that I use to describe the behavior of the subjects, almost none of them have reported that they had a feeling of beeing in a 2D maze. What seems to be happening is that subjects learn isolated chunks of sensory-motor couplings that they can repeat at will, and then put additional ordering on these chuncs according to their proximity to the goal beep place in this abstract beep-space.Currently, we are exploring the effects of the initial instruction on subjects behavior (e.g. "Find the goal place" instead of "Produce the goal beep").
At the end, I relate this work to the research, from where the whole idea of simple virtual environments originated. This research was about artificial agents who were to solve the problem of coming up with a useful environment model (useful in the sense that it could be used to plan a path from one place to another) )based solely on the local sensory information. Unlike in the classical AI they had no access to a bird eye view of their environment which actually reduces the problem to systematic search. Some of the resullta are published in ( Stojanov, G., Embodiment as Metaphor: Metaphorizing-in The Environment in C. Nehaniv (ed.) Lecture Notes in Artificial Intelligence, Vol. 1562, Springer-Verlag, 1999; Stojanov, G. Expectancy Theory and Interpretation of EXG curves in the Context of Machine and Biological Intelligence, PhD Thesis, University in Skopje, Macedonia, 1997. Stojanov, G., Bozinovski S., Trajkovski, G., Interactionist-Expectative View on Agency and Learning", IMACS Journal for Mathematics and Computers in Simulation, North-Holland Publishers, Amsterdam, Vol. 44, 1997; Stojanov, G., Trajkovski, G., Spatial Representations for Mobile Robots: An Algorithm for Detection of Learnable and Non-learnable Environments, 1st Congress of Mathematicians and Computer Scientists in Macedonia, Ohrid, 1996.Stojanov, G., Stefanovski, S., Bozinovski, S., Expectancy Based Emergent Environment Models for Autonomous Agents, 5th International Symposium on Automatic Control and Computer Science, Iasi, Romania, 1995).
"Learning Spatial Concepts by Learning Sensorimotor Contingencies"
James J. Clark, Fatima Drissi
McGill University, Montreal, Canada (email@example.com, firstname.lastname@example.org)
Abstract: We propose that the brain uses sets of sensorimotor basis functions to represent sensorimotor manifolds. Different manifolds represent different spatial concepts and quantities. Following the sensorimotor contingency theory of O'Regan and Noe, we propose that learning, and recognition or perception, is carried out by actively exploring the sensorimotor contingencies associated with a given spatial concept, under the assumption of a stable world.
"Modeling Rat Place Cells as a Basis for Navigation"
Swiss Federal Institute of Technology, Laboratory of Computational Neuroscience, DI-LCN Lausanne EPFL (email@example.com)
Abstract: Neurons in the hippocampus of rats respond preferentially when the animal is at a specific location in the environment. Thus activity of each single neuron `represents' a certain place in the environment (place field). Two questions will be addressed in this talk from a modeling point of view: First, how can we derive place fields from visual and proprioceptive information? We show that, after some suitable preprocessing, visual and proprioceptive information? We show that, after some suitable preprocessing, unsupervised learning yields a stable representation of the environment by many overlapping place field. Second, how can place fields be used for navigation? We show that the dense representation by place fields that results from exploration is an ideal basis for reinforcement learning. A learned target position can be reached after 5-10 trials. The model is demonstrated on a Khepera robot with video camera head.
"From view cells and place cells to cognitive map learning:
a robotics perspective"
P. Gaussier (1), A. Revel (1), J.P. Banquet (2), M. Quoy (1)
(1) Neuro-cybernetic team, Image and Signal processing Lab. (ETIS), URA 8051, 6 av du Ponceau, 95014 Cergy Pontoise, France (firstname.lastname@example.org, email@example.com, firstname.lastname@example.org)
(2) Neurosciences and modelisation institute, INSERM 483, Jussieu, Paris (email@example.com)
Abstract: In this talk a model of the hippocampal system (Hs) that conciliates the presence of neurons that look like "place cells'' with the implication of the Hs in other cognitive tasks (complex conditioning acquisition, memory tasks...) is proposed. Robotics experiments and mathematical considerations the Hs in other cognitive tasks (complex conditioning acquisition, memory tasks...) is proposed. Robotics experiments and mathematical considerations show that ''place cells'' or view ''cells'' can be learned in the perirhinal and entorhinal cortex. The role of Hs would not be fundamentally dedicated to navigation or map building. In our model, Hs is used to learn, store and predict transitions between multimodal states. This transition prediction mechanism could be important for novelty detection but, above all, crucial to merge in a single and coherent system planning and sensory-motor functions. A neural architecture embedding this model has been successfully tested on an autonomous robot, during navigation and planning in an open environment showing the interest of using robots has a simulation tool to understand cognitive systems in the case of complex system/environment dynamics.
"Bayesian elements for perception and space representation"
CNRS - Laboratoire LEIBNIZ, IMAG, Grenoble (Pierre.Bessiere@imag.fr)
Abstract: We will briefly present a Bayesian theory of sensory-motor systems. We will then develop in some details some experiments done with mobile robots using this Bayesian approach: sensor fusion, object and situation recognition, night watchman behaviour, chasing an elusive target, navigation based on smell information, etc... We will illustrate with this experiments how some important issues in perception, navigation and space representation may be addressed without any geometrical models.
"Multisensory spatial representations"
Brain and Cognitive Science Dept, 9216 MelioraHall, University of Rochester, Rochester NY 14627 (firstname.lastname@example.org)
Abstract: Our ability to reach, or move our eyes, toward a sensory target, whether this target is seen, heard or touched, indicates that the brain can transform any sensory coordinates into motor coordinates. Most models assume that these coordinate transforms are broken down by the nervous system into a series of steps in which the position of the target is successively recoded into several intermediate frames of reference (such as head-centered, body-centered and so on). Moreover, each sensory modality are believed to enter this process at the level of their primary frame of reference (eg. eye-centered for vision, head-centered for audition, body-centered for touch). Unfortunately, as we will discuss, this view is incompatible with behavioral and neurophysiological data. We will suggest an alternative in which coordinates transforms and multisensory integration are performed by basis function networks with multidimensional attractors. We will show that this architecture is consistent with anatomical, neurophysiological and behavioral data while being computationally optimal.
"How can properties be ascribed to an object? Comparison
between psychophysical experiments and robots simulations"
Jacques Droulez & Nicolas Bullot
Laboratoire de Physiologie de la Perception et de l'Action, CNRS-Collège de France, 11 place Marcelin-Berthelot, 75005 Paris, France (email@example.com ,firstname.lastname@example.org)
Abstract: Classical experiments in psychophysics investigate the abilities of perceptual systems to recover information from sensory flow. This kind of research program often relies on an implicit hypothesis (the psycho-physical isomorphism hypothesis) according to which the ability to perceive objects relies on the extraction, binding and storage of different intrinsic properties, or features, belonging to the distal object. There now exists a growing literature which criticizes this hyptothesis and postulates that the explanation of perception requires the study of the interactions between agent and environment, and of the different ways sensorimotor information connects to cognition. Following this trend of research, we will develop the idea that, in perception, a crucial kind of agent/object-properties – which will be called «current-status property» – is actively attributed to objects by the agent. The current-status properties are not object-intrinsic and therefore cannot be directly extracted from the sensory flow ; they rather indicate the kind of pragmatic role the object plays within the course of the agent’s action plan. We assume that such an ascription plays a central role in object identification and action planning.
To illustrate this hypothesis, we have designed a task based on a Modified Traveling Salesman Problem (MTSP) which can applied to both psychophysical experimental protocol and robot simulations. As in standard TSP, the agent has to find the minimal path connecting a set of targets. In MTSP however, targets are not labelled and distance between targets are not given as a primary data to the agent. To solve MTSP, the agent has then to identify targets by ascribing current-status property (for instance, to distinguish between «already connected targets» from «not yet reached targets») and also to extract some relevant intrinsic properties such as velocity or distances from motion parallax. The results obtained in psychophysical experiments with up to 10 targets) show that subjects made only a few mistakes in target identification while the traveled distance was significantly greater than the optimal solution. These results were compared to the performances reached by various simulated robots. The simulations show that internal states corresponding to current-status properties are required to reach the subject level of performance. Finally, we will discuss the kind of common reference (pointer-based versus topographical) required for both object-intrinsic and current-status properties.
"Interactive perception: some implications for metaphysics
Centre for Logic and Philosophy of Science, Department of Philosophy, Free University Brussels (VUB) (email@example.com)
Abstract: This presentation has three intertwined aims:
1) to provide a general definition of theories that describe perception as the exercise of potential for action, such as the 'sensorimotor contingency theory' (O'Regan and Noë, Behavioral and Brain Sciences (in press);
2) to sketch how such theories require a different metaphysics than the traditional reductionist one;
3) to outline how such a theory requires a different view of the relation between neuroscience and perception than the traditional one.
I will argue that the conception of perception as the active exercise of know-how concerning relations between input and self-movement implies that perception cannot be identified with lower-level neurobiological, let alone physical, events. This creates a conflict with both traditional reductionist metaphysics and with the idea of there being 'neural correlates for consciousness'. I argue a more satisfactory metaphysics can be found in 'promiscuous realism' (J. Dupré, The Disorder of Things, Harvard UP, 1993) and I sketch how the relation between capacity theories of perception and neuroscience can be reestablished though focusing on neural plasticity rather than on snapshots of neural activity.
"Neural plasticity and consciousness: a sensorimotor approach"
Alva Noe (1), Susan Hurley (2)
(1) Department of Philosophy, University of California at Santa Cruz, Santa Cruz, CA 95064 (firstname.lastname@example.org) ; http://www2.ucsc.edu/people/anoe/
(2) University of Warwick, United Kingdom (email@example.com)
Abstract: Many scientists ask: why is brain activity in this particular region of cortex experienced as like this rather than like that? - as red rather than green, say, or as visual rather than auditory? It tends to be assumed that the answer must depend on intrinsic properties of the brain activity. We tend to think, that is, that what an experience is like is the direct result of the activity of populations of brain cells. But perhaps it is just this assumption that we can explain consciousness in terms of intrinsic properties of neural activity that is the source of our enduring puzzlement? In this paper we challenge the idea that the qualitative character of experience - consciousness - is a phenomenon that can be understood just in terms of intrinsic properties of neural activity. We propose, on the basis of considerations about the phenomenon of neural plasticity, that the qualititative expression of neural activity is determined not by the intrinsic features of the activity alone, but the way that neural activity is embedded within the conext of the sensorimotor organization of the animal. We also address broader metaphysical implications of this proposal.