Neural network devices based on reaction–duffusion media: an approach to artificial retina

Neural network devices based on reaction–duffusion media: an approach to artificial retina

PII: S0968-5677(98)00121-7 Supramolecular Science 5 (1998) 765—767  1998 Elsevier Science Limited Printed in Great Britain. All rights reserved 0968...

397KB Sizes 0 Downloads 12 Views

PII: S0968-5677(98)00121-7

Supramolecular Science 5 (1998) 765—767  1998 Elsevier Science Limited Printed in Great Britain. All rights reserved 0968-5677/98/$19.00

Neural network devices based on reaction–duffusion media: an approach to artificial retina Nicholas G. Rambidi International Research Institute for Management Sciences, 9, Prospect 60-let Oktyabria, 117312 Moscow, Russia

Neural network architecture of devices based on reaction—diffusion media is discussed. Image processing operations performed by these media are similar to first steps of human vision mechanism.  Elsevier Science Limited. All rights reserved. (Keywords: neural network; reaction–diffusion; artificial retina)

INTRODUCTION Neural network devices are the best known alternative to von Neumann computers. Now, when problems of high computational complexity define important aspects of human activity, a neural net approach began to give practical, tangible results at both software and hardware levels. Practical versions of the neural network devices intensively elaborated now are based mainly on traditional semiconductor digital circuitry and planar technology of VLSI production. Owing to this, the manufacture of neurochips faces a number of fundamental complicated problems, such as large-scale integration limits, the ‘nightmare of interconnections’, implementation of gradually changing connection weights of neuron interaction, etc. Therefore, it seems natural to look for both fundamentally new ways for neural network realization and new technological approaches. One of these is the overall usage of both reaction—diffusion active media and biomolecular materials and technologies.

NEURAL NETWORK ARCHITECTURE OF REACTION—DIFFUSION MEDIA Neural network architecture of active reaction—diffusion media follows from general considerations. Let us restrict the consideration to pseudo-two-dimensional versions of active media and outline their general structural features that determine information processing characteristics. (i) Small cells of an excitable medium compared in their dimensions with diffusion length can be considered as primitive processors representing chemical systems having point wise kinetics. To consider these cells as independent, trigger and oscillatory regimes should be inherent in them.

In common parlance, the diffusion length is a distance where total intermixing of reaction components takes place because of diffusion. (ii) The coupling between cells is accomplished due to diffusion. It determines a number of complicated dynamic modes that are displayed in thin layers and in the volume of excitable media. (iii) The control of excitable medium regimes can be carried out by changing the composition and temperature of the medium and by physical stimuli (light radiation, electrical potentials and so on). (iv) Generally speaking, due to diffusion coupling each cell is connected to each other cell of the medium. Nonetheless this interaction is carried out with a time delay proportional to the distance between cells and the strength of the interaction decreases proportionally to this distance. (v) The model representing a system of cells coupled by diffusion does not take into account that excitable media are uniformly distributed systems. More adequate model should be invariant to infinitesimal shifts of the cell system along the surface of the medium. The detailed consideration of information processing capabilities of reaction-diffusion media shows that for the case of pseudo-two-dimensional versions, these media can be described in terms of neural networks having lateral connections. The main responses of shunting on-center off-surround feedback neural networks and reaction—diffusion active systems proved to be surprisingly similar.

INFORMATION PROCESSING CAPABILITIES OF ACTIVE MEDIA Given these features of excitable media, the comparison of dynamic and information processing characteristics of

SUPRAMOLECULAR SCIENCE Volume 5 Numbers 5—6 1998 765

Neural network devices: N.G. Rambidi excitable media and neural networks seems to be of a substantial value. One of the straightforward ways to simulate information processing primitive operations of biomolecular devices based on complex nonlinear dynamics is experimental modeling. In this case Belousov—Zhabotinsky-type excitable media are effective because of close similarity between dynamics of these media and known reaction—diffusion systems . The basic important feature of light-sensitive Belousov—Zhabotinsky media is that they store input information for a rather long period of time. Periodical process of stored image transformation begins after projecting an image on to a thin layer of the medium. This process represents a combination of three interlaced primitive responses to the light excitation: E the contour enhancement of image fragments, E the alternation of negative and positive images of an input picture, E the disappearance of small features of the picture. The responses of excitable medium to the light excitation determine the character of primitive information processing operations that could be performed by the medium. Image processing operations performed by active chemical media proved to be similar to the human visual capabilities and dependent on the state of the medium. There are two main sets of them. The first of them can be defined as ‘description of the general features of an object’. This set includes such Figure 2 Processing images corresponding to Kanizsa (A) and Kennedy (B) illusions by active media

primitive operations as concentration on the general outline of an image (Figure 1A), removing small immaterial features (Figure 1B) including random noise filtration, ‘addition to the whole’ operations, and, in particular, restoration of an image having defects (Figure 1C). The second set of image processing operations can be determined as ‘switching to the details of an image’. It includes contour enhancement (Figure 1D), segmentation (Figure 1E), that is division of an image into simple parts, image skeletonizing, italicizing small features of an image (Figure 1F ). In general, the investigation performed showed that even the simplest (homogeneous, primitive in their structure) excitable media are capable of performing complicated enough operations of image processing as primitive.

CONCLUDING REMARKS

Figure 1 The responces of active media to input of black and white pictures representing processes of: smoothing of immaterial features of an image (A), removing of small features (B), defect repair (C), contour enhancement (D), segmentation of an image (E), and enhancement of small features of an image (F)

The main results discussed above and detailed investigations performed earlier show that image processing operations of active media and the first steps of human vision are surprisingly adequate. Therefore, active medium seems to be a model of artificial retina.

766 SUPRAMOLECULAR SCIENCE Volume 5 Numbers 5—6 1998

Neural network devices: N.G. Rambidi Some interesting features of this model were found in the investigations performed. Illusory figures are widely known features of human vision. The mechanism of this phenomenon is rather sophisticated. It is possible to see based on image processing by active media (Figure 2) that some features of illusory figures seem to be displayed at the first steps of image transformation. Square (in the case of Kanizsa illusion) and circle (in the case of Kennedy illusion) are results of contour evolution in the process of image transformation by active media.

ACKNOWLEDGEMENTS The author gratefully acknowledges the support of K.C. Wong Education Foundation, Hong Kong.

REFERENCES 1 Rambidi, N.G. and Maximychev, A.V. BioSystems 1997, 41, 195 2 Rambidi, N.G. http://www.cs.wayne.edu/&ngr/biocomputing1/, 1997 3 Rambidi, N.G. and Maximychev, A.V. Advanced Materials for Optics and Electronics 1995, 5, 223

SUPRAMOLECULAR SCIENCE Volume 5 Numbers 5—6 1998 767