Anticipation and anticipatory behavior: II - Google Sites

1 downloads 190 Views 124KB Size Report
action execution and understanding, illustrating how the same anticipatory ... different time scales and granularities (
Cogn Process (2007) 8:149–150 DOI 10.1007/s10339-007-0177-8

EDITORIAL

Anticipation and anticipatory behavior: II Giovanni Pezzulo Æ Joachim Hoffmann Æ Rino Falcone

Received: 12 June 2007 / Accepted: 4 July 2007 / Published online: 24 July 2007 Ó Marta Olivetti Belardinelli and Springer-Verlag 2007

Introduction to the special section The three papers in the special section of Cognitive Processing continue and extend the June 2007 issue, which was devoted to ‘‘Anticipation and Anticipatory Behavior’’. The contributions in particular offer insights about generative, anticipatory mechanisms in the brain and in artificial systems, related to issues such as action control (focusing on gaze control), action recognition, prediction of intent, and inference of action’s causes given its observable effects. In philosophy the productive perspective of cognition originates from Kant’s idea that, representations, although rooted on empirical ground, are produced by the cognitive apparatus and are not passive reflections of the world. The papers included in this special section illustrate nicely this point, from the (highly complementary) computational and neurobiological perspectives. Balkenius and Johansson’s paper is a research report of computational work on the development of gaze control in humans. It illustrates how increasingly complex anticipatory mechanisms are progressively developed by children for the sake of gaze control, and how they interact and coordinate. The other two papers review the role of anticipatory mechanisms in G. Pezzulo (&)  R. Falcone Istituto di Scienze e Tecnologie della Cognizione, CNR, Via S. Martino della Battaglia, 44-00185 Rome, Italy e-mail: [email protected] R. Falcone e-mail: [email protected] J. Hoffmann University of Wu¨rzburg, Rontgenring 11, 7070 Wu¨rzburg, Germany e-mail: [email protected]

action execution and understanding, illustrating how the same anticipatory mechanism can be in play for both. Kilner, Friston, and Frith’s paper offers a theoretical perspective, grounded in neurobiological evidence, on the role of mirror neurons in the inference of intentions from actions. Demiris’s paper focuses on action understanding and imitation in cognitive robotics, reviewing actual robotic models.

Multiple anticipatory mechanisms in the brain, their development, and interaction One of the aims of the special issue is understanding the variety of anticipatory mechanisms in the brain. One important aspect is that there is not a one-to-one correspondence between anticipatory mechanisms and cognitive functions. On the contrary, the same anticipatory mechanism can be involved in the realization of different cognitive capabilities, and complex cognitive capabilities may require the integration and coordination of several anticipatory mechanisms. Moreover, complex anticipatory mechanisms can develop from simpler ones, and typically they do not replace but coordinate with the preexisting ones. This mechanism permits the refinement of one own’s cognitive functions as well as the development of novel ones. The paper by Christian Balkenius and Birger Johansson explores this issue by illustrating the coordination of multiple anticipatory mechanisms in gaze control. A computational model of gaze control is introduced according to which action execution is accompanied by predictions of its sensory consequences, which are produced by internal, generative models. In particular, the authors discuss how two generative models, specialized for predicting at

123

150

different time scales and granularities (target positions and perceptual events), can be developed in successive stages, and play complementary roles. The former learns continuous models of target motion, and the latter learns discrete events that change the motion of the target. The simulations demonstrate that a successful gaze control depends on an adaptive switch of anticipatory strategy, and then that anticipatory mechanisms having different complexity and scope can be successfully integrated. The study also describes an interaction between the dynamic control of gaze and the mechanisms for learning (and conceptualizing) discrete events. This aspect is particularly relevant since it illustrates that a pressure for anticipating at long time scales actually generates the possibility to discretize events as a novel cognitive capability.

Cogn Process (2007) 8:149–150

architectures used in cognitive robotics. The prototypical example illustrated in the paper is the HAMMER model that consists of multiple pairs of inverse and forward models operating in parallel. The main peculiarity of HAMMER is that the same internal models are exploited in order to both produce goal-oriented movement as well as perceive it when produced by others. HAMMER has been successfully used in robotic imitation applications, and its functioning has been related to a biological counterpart, the mirror system. Related themes such as the top-down control of attention when predicting and understanding actions, and the hierarchical organization of the inverse and forward models, are also investigated in the details.

Conclusions Anticipation as a unifying mechanism for executing and understanding movements, actions, and goals Recently and in opposition with the classical separation of the realms of action and perception, an action-centered view of perception and of cognition in general emerges. According to this view action and perception commonly base on the same anticipatory mechanisms of the brain. In this framework, perception (and recognition) are realized by ‘‘inverting’’ the generative models that govern action execution and the prediction of its sensory consequences. The same brain mechanisms may also permit to bridge the gap between the realms of ‘‘self’’ and of ‘‘others’’, as suggested in the mirror system literature. A related theme is the distinction between descriptive and generative models in the recognition of goals and intentions. While the former operate by extracting perceptual features and using pre-existing rules, the latter introduce hidden variables and exploit motor mechanisms that can produce the observed data. James M. Kilner, Karl J. Friston and Chris D. Frith present a theory of how the mirror system permits to infer intentions from actions, and how the mirror neuron system can be interpreted as a predictive coding model. They propose that, like in predictive coding, each level of a hierarchy (representing intention, goal, motor and kinematic) employs a generative model to predict representations in the next subordinated level. The prediction error is used to revise the representation at the the superordinated level. Top-down and bottom-up signals are therefore, used, and when the prediction error is minimized, the most likely cause of the input has been generated. The authors also show that this scheme is equivalent to the statistical approach of empirical Bayes. Jiannis Demiris reviews approaches to action recognition and prediction of intent and surveys generative

123

Taken together, the three papers included in this special section show that an important part of how a cognitive agent engages with its world, at the individual and social levels, depends on its anticipatory capabilities. A cognitive agent can make sense of itself, the world and other agents since it is able to predict them. Anticipation permits to understand phenomena at the subjective level, including the effects that actions have on the world, and the boundaries of an agent’s autonomy, powers and limitations. Anticipation permits to understand phenomena at the objective level, including the existence and persistence of objects, and in which cases their behavior is related or unrelated to the agent itself. Anticipation permits to understand phenomena at the social level, including the actions and intentions of others. Moreover, the passage from simple to increasingly sophisticated anticipatory capabilities plays a role in cognitive development tout cour, since the coordination of multiple anticipatory mechanisms permits to refine existing cognitive functions, and the evolutionary pressure to predict permits to develop novel ones. By highlighting the multiple roles of anticipation, and in particular of generative mechanisms, the papers in this special section complement well those in the former issue, and contribute in a significant view to the general aims of the special issue: to advance the understanding of the phenomenon of anticipation in a multidisciplinary way, and to propose the view of anticipation as a unifying mechanism for cognition, that can stimulate further theoretical, empirical and computational work. Acknowledgments The authors are funded by the EU project MindRACES (2004–2007) ‘‘From Reactive to Anticipatory Cognitive Embodied Systems’’, FP6-511931.