Marr's computational theory of vision is often criticized,
because it considered mainly the vision module of the brain and resulted in a
very static view about the brain mechanism. While his aim was to reconstruct
a 3D world from a 2D image, there was little consideration on
how the visual 3D information can be used for the other modules, for instance, memory, expectation,
motion planning and motion control. Likewise, the influence of the other
sensory modalities
on the vision module has been also neglected.
So far there have been only few research that has achieved an understanding
of the interaction among Marr's three levels in an interdisciplinary manner[1].
There is much experimental evidence with an analytic approach about the brain from anatomy, physiology and psychophysics on different animals. However, there are still many unclear points, in particular regarding function. Biological systems are examples of complex systems, in which dynamic self-organization phenomena play big roles and it is difficult to analyze the causal relations among individual parts and the whole system. To fill the gap between the neural level and psychological phenomena, a synthetic model was built. By considering the model, some feedback and prediction regarding physiology and psychophysics would be possible.
On the one hand, much the consideration has been done, about how the information is processed in the brain, on the other hand there is only few research considering, how information is created in the brain. A neural network is neither a feed-forward filter nor a feature pattern matching detector machine. Living things are not indeed machine at all. Each individual neuron changes its behavior and creates relationships to other neurons according to the global contextual consistency in the system. Therefore cognition has some analogy to Jerne immunity system, which interprets the antigens. The individual parts determine the action of the whole cognitive system, in turn the macroscopic state of the whole cognitive system determines the motion of the individual parts and modules according to the memory[2]. A metric modelling approach has been the norm in robot navigation literature[3]. However, in a big part of the literature, neural networks are used as a kind of filter with trigonal functions, but there is almost no research on how the metric information itself is self-organized. In a self-organizing system operation from the outside (computer programme) is not necessary. The operational information is also self-organized. By such a model some psychophysical experiments could be replicated with a mobile miniature robot Khepera (R) navigating in a hexagonal maze. These are, for instance, distance and direction estimation experiments to discuss why errors occur, the relative reaction time measurement and lesion experiments. It would be possible to discuss higher cognitive functions with one of the simplest models by controlling the environment as far as possible From experiments in rats there are some suggestions that view-specific information is necessary for the cumulative errors in the satial information from the head-direction cells. A rat probably relies on the integration of self-motion most of the time and uses only a few representative local views as the anchor points for calibration[4]. It would be too expensive to compute head direction from local-view information alone, because different scenes seen from all possible viewpoints with parallel sightlines must be considered as equivalent, whereas similar scene as sociated with different directions must be distinguished[5].A synergistic neural model of cortico-hippocampal interplay consisting of neural non-linear oscillators and an attractor neural network is proposed. With the model, spatial cognition or cognitive map was considered as an example of a higher cognitive function. In the model, while on the one hand the macroscopic state of the whole cognitive system determines the motion of the individual parts, the individual parts in turn determine the action of the whole cognitive system. In contrast to the traditional approach whereby the role of hippocampus for navigation is emphasized, in this model the cooperation of hippocampus and the other parts of the brain plays a big role. Moreover apart from conventional vision research, whereby the integration with the other sensory modalities tend to be ignored, the important role of egocentric movement in the cognition was discussed. For that the coupling or interaction between the vision and egocentric motion was examined.
In the learning mode the synergistic loops among the individual parts and the whole system were learnt. After the learning, relying on this synergistic loop macroscopic state of the whole cognitive system and the motion of the individual parts converge in a global consistency by dynamical linking through temporal coherence. Like the interpretation of a sentence, whereby unless the context of the sentence is clear, the meaning of the individual words making up the sentence is also not clear, the interpretation of the individual sensory modules is governed by the global cognitive state. In turn the individual sensory modules determine heuristic the global cognitive state. The comparison between the performance of the model and psychophysical evidence in a navigation task acquired from our group, and general literature on the subject were discussed to examine the plausibility of the model. In particularly the lesion experiments of cutting either the object or the location information path of the model were carried out.