CoSy project

COSY MEETING-OF-MINDS WORKSHOP
PARIS 16-18 SEPT 2007

Speakers, Titles and Abstracts

For Detailed Schedule, see this URL:
http://www.cs.bham.ac.uk/research/projects/cosy/conferences/mofm-paris-07/mofmschedule.html

Name Institution Title Abstract and Home Page

Henrik Christensen
College of Computing, Georgia Institute of Technology, USA, and KTH, Stockholm, Sweden
Cosy Speaker: Challenges in artificial cognitive systems Artificial Cognitive Systems is the study of embodied systems that can perceive, represent, acquire and reason about their own and others activities in the world to enable goal achievement. The design of systems that are truly cognitive would be a major departure from how IT systems are designed, implemented and utilized today. At the heart of this problem are system organization and representations that facilitate communication, generalization, learning, reasoning and fusion. At the same time to be meaningful these systems must be embedded in the real world and as such be
faced with the richness and diversity of the external environment. To study such systems there is a need for a multi-disciplinary approach that allows holistic consideration of all the aspects across systems theory, AI, perception, statistical models, language theory and formal methods. In this
presentation a number of general challenges for cognitive systems are outlined and the approach adopted to these problems in CoSy is also discussed.

- - http://www.cc.gatech.edu/~hic/

Annette Karmiloff-Smith
Developmental Neurocognition Lab, Birkbeck College, London, UK
Built-in modules vs a developmental process of gradual modularisation: insights from genetic disorders In this talk, I will argue that data from typically developing infants demonstrate how human brains become progressively specialised and localised for specific cognitive functions over developmental time, rather than starting out that way. Moreover, rather than pointing to genetically specified cognitive-level modules as some would claim, data from developmental disorders illustrate how despite proficient behaviour on some tasks, cognitive and brain processes are atypical involving a lack of progressive specialisation and localisation of function. In general, I will stress the importance of building full developmental trajectories of cognitive functions from the earliest basic-level cross-domain processes in infancy through childhood to the emergent higher-level domain-specific cognitive processes in the adult endstate.

- - http://www.psyc.bbk.ac.uk/research/DNL/personalpages/annette.html

Yuko Munakata
Department of Psychology, University of Colorado, Boulder, USA
Developing cognitive flexibility: Learning mechanisms and representational change Children show striking developments in cognitive flexibility, becoming increasingly able to overcome habits to respond appropriately to the changing environment. What drives these changes? This talk explores the roles of increasingly active and abstract representations, and of error-driven and self-organizing learning mechanisms. We test behavioral predictions from neural network models regarding effects of language, practice, and feedback on children's cognitive flexibility.

- - http://psych.colorado.edu/~munakata/

Jeremy Wyatt
School of Computer Science, University of Birmingham, UK
CoSy Speaker:
On Architectures for Intelligent Robots: Distributed Representation & Shared Workspaces
There are several strands of work on architectures for intelligent agents, from cognitive science, psychology, robotics and artificial intelligence. Each emphasises a different aspect of the problem, and offers a differently flavoured solution. Some emphasise a single, central model of the world, others the distributed nature of control, or the need for minimal representation, or hierarchy. I will talk about yet another architecture suitable for intelligent agents, or more strictly, an architectural schema. This emphasises the use of specialised representations, which are distributed across several work spaces, and refined collaboratively by several processes. This way of thinking allows us to explore answers to several problems such as: (what we call) the filtering problem; the binding problem; and the process management problem. I will support my argument by describing some case studies in which an implementation of the schema (CAS - CoSy Architecture Schema) is used to create specific architectural instantiations.

- - http://www.cs.bham.ac.uk/~jlw

Karen Adolph
Department of Psychology,
New York University, USA
Developing flexibility Behavioral flexibility is the essence of goal-directed action. This presentation focuses on three aspects of flexibility: perceiving changing possibilities for action under variable and novel conditions, modifying ongoing movements to suit the current constraints on action, and finding new means to achieve a desired outcome. Flexibility will be described across the life-span, from infants to elderly, as participants cope with various challenges to reaching, balance, and locomotion.

- - http://www.psych.nyu.edu/adolph/

Robert Vickerstaff
University of Canterbury, New Zealand
Modeling navigation in desert ants and vision in jumping spiders I will talk about two invertebrate species displaying a surprisingly rich set of behaviours for their small size, and my attempts to produce dynamical system models to mimic and explain these abilities. Path integration in Saharan desert ants allows surprisingly accurate homing in a largely featureless environment. I will present a neural network-based model showing a simple way to reproduce many features of the ant's behaviour. Jumping spiders of the genus Portia have an unusual visual system giving them very high visual acuity in a very small field of view. I will introduce work aimed at understanding how they use their vision as part of their rich and varied hunting behaviour.
http://www.informatics.sussex.ac.uk/research/groups/ccnr/people/RobertVickerstaff.html

- - http://www.ladysmith.eclipse.co.uk/

Thomas Collett
Department of Biology and Environmental Science, University of Sussex, UK
Insect-level intelligence as displayed in arthropod spatial behaviour I will explore this topic, concentrating on learning and memory, using examples from the navigational behaviour of social insects and crabs. Navigational behaviour is particularly informative, because the animals' rich motor output gives many hints about likely mechanisms underlying the behaviour.

- - http://www.sussex.ac.uk/biology/profile547.html

Aaron Sloman
School of Computer Science, University of Birmingham, UK
CoSy Speaker: Understanding causation in robots, animals and children: Hume's way and Kant's way. For Hume causation was just a matter of correlation, and current Bayesian theories of causation fit that general idea, adding conditional probabilities. For Kant causation implied a kind of necessity, analogous to the necessity in mathematical reasoning. There have been disputes as to who was right, whereas I'll argue that both notions of causation are needed and are used by intelligent systems. Roughly Humean causation is all you have when you merely have strong evidence regarding what causes what, whereas in some cases you know *why* something causes something, e.g. why going round a house in one direction produces one series of experiences and going round in the opposite direction produces another, and why when two centrally pivoted gear wheels made of rigid impenetrable material are meshed if one turns clockwise clockwise the other *must* turn counter-clockwise. The history of science is full of cases where Humean causation is replaced by Kantian causation as a result of deeper understanding. Kantian causal understanding, when available, is more powerful, e.g. because it can be used to deal with novel situations. Currently no robots that I know of have any Kantian understanding, and this is a very serious deficiency. Human children acquire Kantian understanding in a piecemeal and idiosyncratic way. It is not clear whether any other animals have this ability, but there is prima-facie evidence that some do. I shall base my presentation on a subset of the topics presented by Jackie Chappell and myself at a recent workshop on natural and artificial cognition. Our slides are available here: http://www.cs.bham.ac.uk/research/projects/cogaff/talks/wonac/

- - http://www.cs.bham.ac.uk/~axs/

Neil Berthier
Department of Psychology,
University of Massachusetts, USA
Development of reaching in infants The development of human infant reaching will be discussed from a computational viewpoint. Significant challenges to the infant will be discussed as well as aspects of the task that allow for infants to use simple algorithms to learn to reach effectively.

- - http://people.umass.edu/neb

Laurie Santos
Department of Psychology, Yale University, USA
The representations that underlie primate social cognition (Revised 23/8/2007) My talk will discuss how primates represent other objects and, in particular, if there is evidence that other primates can represent unobserved aspects of their physical and social world. I'll discuss recent work on monkey social cognition and some new work exploring whether monkeys share a human essentialist bias.

- - http://pantheon.yale.edu/~lrs32/santos.html

Felix Warneken
Department of Developmental and Comparative Psychology,
Max Planck Institute for Evolutionary Anthropology,
Leipzig, Germany
Acting for and with others: Developmental and comparative perspectives on prosocial behavior What are the psychological prerequisites for prosocial behaviors such as altruistic helping and joint cooperative activities? By integrating a developmental and comparative perspective, I aim at elucidating aspects of prosocial behavior which humans share with their closest primate relatives from aspects which emerge only in human ontogeny. I will provide evidence that young children as well as chimpanzees possess the cognitive and motivational prerequisites for helping: They are able to understand another individual's unfulfilled goal and have the altruistic motivation to act on their behalf. With regard to cooperative activities, experimental evidence suggests that chimpanzees are not so much different from young children in their ability to coordinate their actions with another individual in a cooperative situation, but only human children appear to form a representation of these actions as part of a joint plan with a collective goal.

- - http://email.eva.mpg.de/~warneken/

Ales Leonardis
Faculty of Computer and Information Science,
University of Ljubljana, Slovenia
Cosy speaker: Hierarchically Learned Representations of Object Categories: From Pixels towards Semantic Parts The question how to represent visual information in an artificial cognitive system to enable fast and reliable execution of various cognitive tasks has been discussed throughout the history of computer vision. The theories have converged towards hierarchical architectures of parts composed of parts, (the so-called compositional systems), starting with simple, frequent features that are gradually combined into more and more complex entities. However, the automatic design of parts in hierarchical layers has been hindered by a theoretically enormous number of possible compositions. In this talk, I will describe a novel approach that overcomes the exponential complexity of unsupervised learning by exploiting the favorable statistics of natural images in a sequential, hierarchical manner. The parts recovered in the individual layers of the hierarchy vary from simple to more complex ones and enable a fast indexing (bottom-up) and matching (top-down) scheme that can be efficiently used for a variety of cognitive tasks. I will show the results of the proposed approach obtained on different data sets, yielding important insights for designing compositional systems.

- - http://vicos.fri.uni-lj.si/alesl/

Thomas Shipley
Department of Psychology, Temple University, USA
Event segmentation and action recognition My talk will review recent research on how humans decompose and recognize events. We are beginning to understand some of the interplay between top-down and bottom-up processes in event understanding. I will outline a geometric model of low-level event segmentation, and discuss its potential role in children’s verb learning. I will also discuss perception – action interactions and the role of motor cortex in action recognition. Here I will review recent EEG work in which we use the mu rhythm as indicator of mirror system activity to study action recognition in infants and learning by imitation in adults.

- - http://astro.temple.edu/~tshipley/index.html

Cecilia Heyes
Department of Psychology, University College London, UK
Imitation from sensorimotor learning Imitation constitutes a special class of reactions to recalled or incoming stimulation, in which the action responses match or correspond with the action stimuli. However, even in cases where the ‘correspondence problem’ is most acute (e.g. when a system imitates a facial expression or a whole body movement), imitation may not require any special types of computation or innate knowledge. The associative sequence learning (ASL) hypothesis suggests that the development of imitation (and the mirror system) depends on task- and species-general processes of associative, sensorimotor learning. I shall discuss evidence in support of this view from behavioural and neurophysiological studies of imitation in adult humans.

- - http://www.psychol.ucl.ac.uk/celia.heyes/netintro.htm

Günther Knoblich
Psychology Department
Rutgers University, USA
17, 18
Two Challenges for Robotics: Mirroring and Joint Action Recent research on the links between perception, action, and cognition has led to a new understanding of how people perceive each other and how they cooperate with one another. In a nutshell, people use their own action capabilities, emotions, and sensations to perceive, predict, and understand others and to perform joint actions with them. Of course, this development has potentially important implications for robotics because it opens up a road for constructing robot societies where robots learn from each other or perform actions together. However, the recent progress in psychology and the cognitive neurosciences has also resulted in a renewed focus on intentionality, experience, and social relations. It is far from clear how these capabilities could be technically implemented in robots. In my talk I will provide an overview of important findings and discuss which aspects of these findings may be realistically transferable from humans to robots.

- - http://psychology.rutgers.edu/~knoblich/

Luciano Fadiga
Neurolab, University of Ferrara, Italy
From Hand Actions to Language: Anatomo-Physiological Links between Monkey Mirror Neurons and Human Broca's Area The classical serial flow-diagram describing how sensory information is processed and eventually transformed into movements by the brain has becoming more and more implausible because of neuroanatomical and neurophysiological evidence.
In the first part of my presentation I will discuss how this association between stimulus and response, manifesting itself at single neuron level, might provide the goal to the movements. Thus, movements are transformed into actions and, perhaps more interestingly, may give origin to action representations, playing a crucial role in understanding the other individuals and the environment around us.
In the second part of my presentation, I will discuss the possible link between action representation, action understanding and interindividual communication, in the peculiar framework of human language.

- - http://ec.europa.eu/information_society/istevent/2006/cf/people-detail.cfm?id=1992
http://web.unife.it/progetti/neurolab/people.htm

Kim Plunkett
Department of Experimental Psychology, Oxford University, UK
Connectionnist approaches to language acquisition How is lexical structure based on the underlying speech and visual categorisation infants have already acquired during the first year of life? We use a neural network model to demonstrate that good categorisation of auditory and visual stimuli in self-organising maps, leads to extension of a label to other objects within the same category, even after a single presentation of a word-object pair. We show that the ability to generalise labels to objects of like kinds, commonly referred to as the taxonomic assumption, is an emergent property of the model. Finally, we examine the role of constraints imposed on the Hebbian associations in the pairing of objects and labels and show how generalisation can be prevented.

- - http://psyweb.psy.ox.ac.uk/babylab/personal/kim.html

Heinrich Buelthoff
Max Plank Institute for Biological Cybernetics,
Tübingen, Germany
Cosy Advisor: Panel http://www.kyb.mpg.de/~hhb

Ben Kuipers
Computer Science Department, University of Texas at Austin, USA
Cosy Advisor: Panel http://www.cs.utexas.edu/users/kuipers/

Linda Smith
Department of Psychological and Brain Sciences, Indiana University, USA
Words and objects from the first person view: A dynamic and embodied approach to attention Understanding early word learning requires an understanding of the real-world physical and social environment in which learning takes place. However, the relevant aspects of this environment for the learner are only those that make contact with the learner's sensory system. I will present new findings (from a collaborative effort with Chen Yu, Hanako Yoshida, and Alfredo Pereira) using a novel method that seeks to describe the visual learning environment from a young child's point of view. The method consists of a multi-camera sensing environment consisting of two head-mounted mini cameras that are placed on both the child's and the parent's foreheads respectively. The main results is that the adult and child's view are fundamentally different in that the child's view is more dynamic and centered on one object at time. These findings have broad implications for how one thinks about toddler's attentional task as opposed to adults. In one sense, toddlers have found cheap solution: Selectively attend not by changing internal weights by bringing the attended object close to your eyes so it is the only one in view.

- - http://www.indiana.edu/~psych/faculty/smith.html

Randall O'Reilly
Department of Psychology,
University of Colorado Boulder, USA
Abstract representations and embodied agents: prefrontal cortex and basal ganglia contributions Using biologically-based computational models of the prefrontal cortex and basal ganglia (together with posterior cortical areas), my colleagues and I have developed new insights into the development of abstract, rule-like representations, which are critical for higher level cognitive functions. We are currently applying these models to simulated robotic agents equipped with realistic sensory-motor systems, to understand how cognitive control might emerge in embodied agents engaged in developmentally-realistic tasks (e.g., playing with blocks, sorting by colors, etc). We are also exploring how a wide range of standard cognitive control/executive function tasks can be performed in a single system, by building upon more basic primitive skills that develop in naturalistic environments.

- - http://psych.colorado.edu/~oreilly/

Kevin O'Regan
Laboratory of Experimental Psychology, CNRS, Paris, France
CoSy Speaker: How to build phenomenal consciousness into a robot Abstract:
http://nivea.psycho.univ-paris5.fr/

Jacqueline Fagard, Laboratoire Cognition et Développement
Institut de Psychologie - Université Paris 5, France
Local Host, Co-organiser http://www.psycho.univ-paris5.fr/recherch/labo_cog/Personnel/JacquelineFagard.htm