Image Processing with Neuron-Like Branching Elements (POSTER)

The dendritic shape of neurons may be responsible for functional characteristics in information processing in a brain. Since neurons have redundant inputs and excitability thresholds for outputs, logic operations with neurons can be considered as a kind of the AND operation. Here, we propose two-dimensional excitable media of radial dendritic patterns that conduct the AND operation with time constraint. Further, we apply the dendritic elements to construct an image processing circuit. The investigation of the bio-inspired computation may give insights into how the combination of the AND operation processes information. It is known that excitation waves fail to propagate from a narrow path to a broad area, as seen in a diode. By this mechanism called the curvature effect, a signal given on a terminal branch of a dendritic pattern disappears at a central broad area where branches converge. However, when a sufficient number of neighboring branches receives signals around the same time, accumulation of excitation penetration at the central area can revive excitation waves, leading to signal transmission to the unstimulated side of the dendrite. We determine the condition of the excitation property and the geometry of dendrites including branching frequencies for the curvature effect and for the deep penetration of excitation waves by computer simulation. On the description of dendritic pattern formation, we adopt a cellular automaton model of self-organization, by which a multilayer single-electron devise to generate dendritic patterns has been proposed [1]. In the present study, we design a circuit where a number of random dendrites are tiled. One-dimensionalized signals of image information are given from one side of a queue of dendrites, and signal output are detected from the other side. Each element receives data from several pixels via terminals of branches, and only elements at dense signal areas can output signals, resulting in a similar function to characterization of the image. Interaction between adjacent elements tends to facilitate clustering in the output image. We will discuss the correlation between spatial features of input/output images and the geometry of the circuits.