Yutaka Sakaguchi

Topographic organization of nerve field with teacher signal

Abstract A number of neurons which respond selectively to specific stimuli are observed in many parts of the vertebrate nervous system. These neurons are developed by self-organization through sensory experiences. This is regarded as a mechanism for representing the external world structure. However, it is insufficient to think only of this learning mechanism being dependent on the input environment; other information representing mechanisms should exist in the brain. The present article proposes a neural learning model which reflects not only the input signal set but also the teacher signal. This model is constructed by adding a teacher signal layer to the conventional models for topographic organization. After learning, the network realizes the desired input-output relation. Moreover, a structure is organized in the network, not only so as to represent the input environment structure but also to fit well in the “system purpose.” The model behavior was investigated through computer simulations, and the formation of spatial maps in the brain is discussed from a viewpoint of purpose dependent representation.

Haptic sensing system with active perception

Human haptic perception is not caused by simple mechanical stimulation to the skin; it is achieved by integrating sensory information from various receptors (such as mechano-receptors and thermo-receptors) and by observing objects in various ways. The author constructed sensing systems which simulate human haptic processes in order to clarify the mechanism of such sensory integration and active perception. In this article, first the author formalizes these processes from a theoretical point of view and constructs an intentional observation algorithm on the basis of that formalization. This algorithm is to select appropriate sensors from many sensors based on information criteria to recognize objects more accurately by fewer observations. Second, the author describes two kinds of haptic sensing systems which recognize objects materials and surface textures utilizing actively several sensor devices, and shows that the proposed algorithm is effective in these actual systems. Some related problems are also...

Haptic recognition system with sensory integration and attentional perception

The authors constructed a haptic recognition system which discriminates feel of touch based on the principles of sensory integration and attentional perception. The system is equipped with several sensors and can push and rub the objects surface with several values of force and speed. It integrates the sensory information iteratively by selecting appropriate sensors and measurement conditions according to the proceeding of the recognition. The algorithms of sensory integration and of attentional perception are realized by Bayes inference and by an iterative experimental design based on an information criterion, respectively. The experimental result shows that the system can discern a subtle difference in feel of touch. It is also proved that the system selects appropriate sensors and conditions according to the situation, that is, the attentional perception algorithm realizes good recognition accuracy with fewer observations. In addition, it is shown that the characteristics used in the system correspond well to those human beings utilize in haptic perception. These results suggest that the constructed system is a faithful model for the human haptic mechanism.<<ETX>>

Figure recognition with intentional observation

When perceiving an object humans do not receive stimuli passively, but observe the object actively to obtain useful information for recognizing it. A neural network model which realizes such active perception by utilizing neural dynamics is proposed. The model consists of a sensory unit, a representation field, and recognition units. The model sequentially collects local information of the presented figure, gradually constructs its internal image, and recognizes it based on the constructed internal image. An intentional observation mechanism was constructed by which the model can selectively observe informative parts of the figure for constructing the internal image. Several computer simulation results are described. The model is discussed from the viewpoint of Bayesian inference and information theory.<<ETX>>

A robot organizing purposive behavior by itself

For studying the mechanism of the brain, the synthetic approach is effective. The synthetic approach is to conjecture the mechanism of the target through constructing its model. Some 20 models of the brain were constructed for this study. One of them is described. The model includes the faculties of perception, memory, and action. Having these three faculties enables the model to realize highly intellectual behavior or self-organizing ability that cannot be realized by a model having just one faculty. The model was realized in the form of a robot which organizes purposive behavior by itself. This robot forms effective behavioral patterns to achieve the purpose through trial and error.<<ETX>>

Motor planning according to reliability of internal model

The authors propose a motor planning algorithm which reflects reliability of the internal model. The algorithm is to calculate the destination and its expected error for every motor command using the internal model, and to select the most effective command whose expected error would not exceed a given limit. Utilizing this algorithm, the system behaves moderately before adaptation, and its performance is improved as the internal model becomes more reliable through adaptation. It is also shown that the proposed algorithm simulates well the human behavior in a shape-tracing task.