A Cortical Architecture for Parallel Anticipation of Sensorimotor Sequences* A. Heinze, V. Stephan, D. Surmeli, and H.-M. Gross Technical University Ilmenau, Germany Department of Neuroinformatics Andrea. HeinzeQinformatik. tu-ilmenau .de Abstract The basic idea of our anticipatory approach to perception is to avoid the common separation of perception and generation of behavior and t o fuse both aspects into a consistent neural pro- cess. Our approach is based on the prediction of the consequences of hypothetically executed actions. In this sense, perception of space and shape is assumed t o be a generative process o f anticipating the course of events resulting from different sequences of actions. We present a bi- ologically motivated computational model that is able to anticipate and evaluate,hypothetical sensorimotor sequences. Our Model for Antici- pation based on Cortical Representations (MA- COR) allows a completely parallel search at the neocortical level using assemblies of rate coded neurons for grouping, separation, and selection of sensorimotor sequences. 1 Introduction Based on findings for the sensorimotor char- acter of perception [l, 71, we have developed an alternative approach to perception that avoids the common separation of perception and generation of behavior and fuses both as- pects into a consistent neural process [6]. In this approach, perception of space and shape in the environment is regarded to be an active process which anticipates the sensory conse- quences of alternative hypothetical interac- ‘This work was supported by the DFG-Grants Gr 1378/1-1&2 to H.-M. Gross. tions with the environment, that could be performed by the sensorimotor system, start- ing from the current sensory situation. This point of view emphasizes the generative and anticipative character of perception consider- ing both sensory and motor aspects of the action-perception-c ycle. In the following, we turn to the question of where in the brain anticipating sensori- motor systems may be expected. Based on these findings, in Section 3, we present MA- COR. This model allows a completely par- allel search at the neocortical level using as- semblies of rate coded or spiking neurons for generation, grouping, separation, and selec- tion of sensorimotor sequences. It is intended as a general scheme for sensorimotor antici- pation in a neural architecture. It does not attempt to provide a detailed description of a specific cortical or subcortical structure, but we try to capture some general properties of architecture and processing that are relevant to our “perception as anticipation”-approach in brain-like systems (for details see [3]). 2 Biological evidence Figure 1 provides an overview of the main cortical structures used in our model includ- ing the brain regions involved and their con- nections. The visual information about an actual situation is carried from the retina to the cerebral cortex and reaches also the pos- terior parietal cortex (PPC). The PPC ef- fects the integration of different sensory in- Artificial Neural Networks, 7 - 10 September 1999, Conference Publication No. 470 0 IEE 1999 407 in: Proc. of the Int. Conference on Artificial Neural Networks ICANN'99, pp. 407-412