JPH07120347B2 - Spatial information relocation device - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a spatial information rearrangement device applied to a pattern information processing device, a data processor, a spatial filter device, a parallel computer, a systolic array processor and the like. Here, "spatial information" is a collection of data that exists in a spatially distributed form such as image data, and that the relative spatial relationship between the constituent data also has an important meaning. Is meant.

[0002]

2. Description of the Related Art It is well known that the visual function of animals is extremely advanced through the process of evolution because it is an important function related to the conservation of the species.

Research on artificially imitating such pattern recognition functions represented by visual sense started in the 1950s on the basis of analog technology, and with the development of digital technology thereafter, research on computer processing systems is now under way. Is being promoted in various fields.

The Rosenblatt perceptron is well known as a pioneering research example, but the schematic configuration is as shown in the correspondence between the physical configuration of the retinal tissue and the three-layer perceptron in FIG.

Perceptron is a kind of hierarchical neural network composed of three layers of S, A and R, and corresponds to S layer (sensory layer) which corresponds to a receptor, A layer (association layer) corresponding to a bipolar cell, and ganglion cell. It is composed of R layer (reaction layer) and the random connection corresponding to the synapse group that connects them. A series of input patterns and a large number of expected outputs are prepared, and while presenting these repeatedly, R so that the actual output is asymptotic to the expected output
It adopts a configuration that controls the parameters of layers to improve the rate of correct answers.

The perceptron is an extremely epoch-making system as a learning machine, and has shown considerable performance for a limited range of problems, but it is impossible to discriminate with this system from the results of theoretical analysis thereafter. It turns out that there are quite a few problems. For this reason, the perceptron itself has not been researched at present, but research to improve it based on this is still being carried out in various fields.

Visual information usually has a remarkably large amount of information as compared with voice, but when processing this as digital data, since the input information is an analog signal,
A huge amount of AD conversion operation corresponding to the number of pixels is indispensable.

Further, since the processing of the acquired data becomes enormous, it is necessary to take measures such as speeding up by using a plurality of processors in a complicated processing, and in the processing by the conventional Neumann type computer. It is said that the limit has already been reached.

Digital processing has a great advantage of being able to access arbitrary data within a given data range and to process them accurately, but on the other hand, if the data amount becomes huge like visual information, the processing speed becomes faster. It is customary to cause a rapid increase in costs in terms of both hardware and software.

According to the results of research on the visual sense of animals, a considerable amount of information processing is already applied at the stage of the retina, which is a receptor, and the summarized result called essence is transmitted to the visual center of the brain. Is seen. From this point of view, research on simulating the retina with analog electronic circuits has also been conducted in various fields, and considerable research results have already been reported.

[0011]

It is said that the complex synaptic connection between a plurality of nerve cells generally constitutes the whole not only in the retina but also in the nerve tissue. However, when actually simulating with an electronic circuit, Achieving this connection is quite difficult and requires a large amount of spatial connections. For example, in the perceptron, the synapse connection is actually simulated by such a method, but it is not easy to change the hard connection, so the system has a low degree of freedom.

In the case of complete digital processing, this is realized by software by memory addressing, so if it is prepared to take time as described above, there are relatively few obstacles in this respect and the degree of freedom of the system is high. Can be said.

[0013]

SUMMARY OF THE INVENTION The present invention has been made in view of the above-mentioned viewpoints, and is a group of paths composed of charge transfer element arrays arranged along a plurality of paths that do not intersect each other. A plurality of charge transfer elements arranged at positions intersecting each other, and the charge transfer elements positioned at the intersections of the resulting paths,
(EN) A spatial information rearrangement device configured to be shared among a plurality of charge transfer element arrays involved in the intersection so that each path group performs a transfer operation along the path group in different time intervals. It is what

[0014]

According to the present invention, a large amount of these spatial connections required by the neural network are pseudo-realized through a spatial data bus (hereinafter referred to as SDB) composed of a CCD or the like.

The CCD, which is widely used in video cameras and the like at present, is a typical charge transfer device. This device moves an input signal converted into the amount of charges on a semiconductor surface by a driving voltage from an external electrode. It is an element that performs transfer by confining it in a “potential well”. In addition to this, there is an element of a mechanism that transfers a signal via electric charge, but since the operation principle is different but the functions are similar, the description will be given centering on the CCD.

The SDB is distributed by utilizing a signal transfer function of a CCD or the like and performing transfer alternately not in a single direction like an image pickup device but in a plurality of transfer paths having a crossing relationship with each other. Collecting data on the plane at one point,
On the contrary, it realizes the function of distributing one point of data to the periphery.

The operation of the SDB is very similar to how people in the town gather on this bus system at several points in the city where city buses that circulate through a fixed route come and go. It is easy to understand if you think that they are very similar.

The driving path of the SDB has various forms depending on the driving amount and driving order of each path, and for example, the signals at the points located at the vertices of a regular hexagon are converged to the center thereof or arranged in a substantially straight line. It is possible to flexibly select according to the application, such as converging the signals of the six points to one point.

The input signal of the SDB may be any charge signal, but in the case of optical input in particular, the input signal can be easily obtained by arranging a photoelectric conversion element around the SDB like the image pickup element. it can.

In addition to the transfer function, the CCD element has a normal binary logic operation, an analog operation such as a sum, a multivalued logic operation, a multivalued AD conversion function, a maximum value selection function, and a fuzzy logic operation. It is possible, and by installing an arithmetic unit configured with these functions in the SDB, it is possible to aggregate a series of signals converged from a peripheral to one point as described above, perform local processing, and output the result. .

The calculation processing result can be directly transmitted to the outside through the wiring or to the outside through the SDB, or can be directly modulated to an output signal of an electromagnetic wave such as an infrared LED and directly transmitted to the outside. .

The gap created between the paths of the SDB will be referred to as a cell here, but such cells should be arranged in a distributed manner according to the use of the above-mentioned three functions of input, operation and output. You can

In such arithmetic processing, the processing result is again S.
By returning it to the DB, it is possible to transfer it to a cell having another arithmetic function and add a higher arithmetic processing, which is suitable for developing a more advanced processing.

[0024]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The structure of an embodiment of the present invention will be described in detail below with reference to the drawings. First, the definition of terms will be described. Number of route groups: (L) Of the total number of routes, the number of all route groups when the route in which the shift operation is always performed at the same time is regarded as one route group. Intersection order: (X) When considering one element,
The number of input / output elements is not included in the number of other SDB elements connected thereto. Therefore, at the crossroads, X = 4
For a simple shift register with no intersections, X = 2. Intersection ratio: (Y) The ratio of intersections of a specific order among all SDB elements. Stage: Indicates an individual charge transfer element forming SDB. Node: A stage on which two or more paths intersect. Ark: A stage or series of stages that involves only one path. Node signal: Information held by a node at a reference timing such as the start of a series of shift operations. Arc signal: Information held in the arc at the above points.

As a relatively simple example of the structure of the SDB, the number of route groups L = 3 the maximum intersection degree Xmax = 4 the fourth-order intersection ratio Y4 = 100% (all are nodes,
FIG. 2 shows an example of the SDB structure in the case of (without arc). The whole of FIG. 2 is composed of repeating hexagons shown by dotted lines.

In FIG. 2, in order to avoid confusion, the charge transfer mechanism, which is already obvious, is omitted and only the direction of the transfer output is shown by hatching. Of course, each SDB element is driven in accordance with the transfer direction. Electrodes are attached.

FIG. 2 is composed of three route groups P, Q, and R macroscopically intersecting each other at an angle of 120 degrees, and all the intersections are of degree 4. Further, a shift direction opposite to the other routes belonging to the same route group adjacent to each other is assigned to the routes forming each route group.

Route groups P, Q, R constituting the above SDB
For S, the shift operation is sequentially added, but in SDB, the shift is performed according to the following principle. Principles of SDB's shift operation "Do not shift along a route that has intersections with each other at the same time."

Therefore, P, Q, and R suspend the shift operation independently of the time intervals assigned to the respective route groups.

In most cases, the SDB shift operation achieves the purpose with a finite number of times, but when the shift direction is fixed for each path, this series of operations has a finite number of P, Q, R. It is uniquely represented by a list of
It is defined as a drive vector (V) of B.

For example, V = (P, Q, P, Q, Q, Q, Q, P, P, P) ........ 11 Equation 1 happens to be a vector that does not include any shift in the R direction. However, FIG. 3 shows a route followed by a signal at a certain point when this operation is repeatedly executed in the SDB of FIG. In FIG. 3, a locus in which the information at the initial position S (0,0) moves by the operation of V is shown.

If the output signal of the photoelectric conversion element arranged in the surrounding cells is taken in at the initial position S (0,0) at this point, this signal is output after 10 shift operations determined by V. Although reaching S (0,9), the above-mentioned operation of V causes the other five points S (1,9) -S (5) located around the cell C adjacent to the position S (0,9). , 9) initial position S
It can be seen that (1,0) -S (5,0) has a regular hexagonal arrangement including the position S (0,0).

That is, since the spatial information originally distributed is gathered around the cell of C, it is equivalent to the SDB pseudo-realizing six spatial wirings, and the arithmetic circuit is provided in this cell C. By installing, these dispersed input signals can be easily aggregated.

The cells marked with * in FIG. 3 are all cells C.
The above-mentioned calculation function can be added to a cell equivalent to the above-mentioned cell if necessary, and the result of the calculation is the position S (0,9) -S (5.
In the case of the rearrangement in any one of 9), this signal is transferred to another cell C ′ having a higher arithmetic function.

Also, the initial point S (N,
0) shows a variety of completely different forms when V other than Equation 1 is adopted, and this example is shown in FIG. In FIG. 4, the movement of the signal by the operation V = (Q, P, R, Q, P, P, P, P, P, P) is shown.

Although the above description is based on the case where the input signal is taken in only from the position of S (N, 0), it is further expanded and all SDs are read immediately before the start point of the operation V.
Consider a case where optical signals from peripheral input cells are simultaneously captured in the B stage.

In this case, optical signals at different points sequentially arrive at each shift, starting with the signal of the input cell closest to the cell C along each signal locus, in the cell C.
Naturally, in the tenth shift, the positions S (0,0) -S (5,
The signal 0) will arrive. If all the arrival signals during this period are captured by the cell C, it is possible to aggregate signals from more points.

Next, FIG. 5 shows an example in which the number of route groups is L = 3, the maximum crossing degree is Xmax = 4, the second-order intersection ratio is Y2 = 80%, and the fourth-order intersection ratio is Y4 = 20%. Figure 5 SD with nodes and arcs
In the example of B, hatching indicates the output end of each arc.

As can be seen from FIG. 5, in this example, since four stages of shift registers are interposed between the SDB stages of the configuration of FIG. 2, the shifts in the P, Q, and R directions are performed 5 times each. When executed collectively, the path that the node signal follows is exactly the same as in the case of the configuration of FIG. 2, but the arc signal shows a completely different behavior locus. That is, since the arc signal is always arranged on the stage of X = 2 (therefore, it has only one directionality of P, Q, and R) at the time of switching the shift direction, a single SDB is arranged. It will move intermittently along the route.

Therefore, there are two types of signal groups showing different behavior loci by the same shift operation. It is easy to understand if this is similar to the case where city buses and subways are mixed.

FIG. 6 shows a signal locus when the shift specified by the equation 2 is applied to the example of FIG. V = (P5, P5, P5, P5, Q5, Q5, Q5, Q5, R5, R5, R5, R5) ···· 2 P5, Q5, R5 are each shifted 5 times in each direction Indicates.

As can be seen from FIG. 6, since the arc signals are radially gathered in one arithmetic cell C from the periphery, all input signals to points along this path can be accessed by the central arithmetic cell. . The number of stages that make up an arc is at least one, but when distributing data to a wider range, the arc is made up of several stages as shown in the example of FIG. You can use the method that you use.

On the other hand, in the operation of the above formula 2, since the node signal around the operation cell C causes the movement of a regular hexagon similar to the example of FIG. 2, the operation cell similar to the cell C is properly arranged as shown. Then, if the operation result is rearranged on the SDB as a node signal at the time of switching the transfer direction, the operation results from the six Cs are gathered around the next operation cell C ′ at the end of the next series of shifts. It is also possible to transfer the data to cells such as C ″ and C ′ ″ and stack other operations if necessary.

Of course, the output signal can be linearly transferred not only as a node signal but also as an arc signal, and the point is that it is possible to select whether to use a bus or to take a subway. is there.

The contents of the arithmetic cells are irrelevant to the present invention, so a detailed description thereof will be omitted. As an example, FIG. 7 shows an example of the arithmetic cells for the weighted sum operation conforming to the SDB having the configuration of FIG. FIG. 7 shows an example of a computing cell having both optical output and charge output. Reference numeral 1 in FIG. 7 indicates a stage for acquiring a signal from the arc.

Data taken in every five shifts are sequentially transferred to the shift register 2 connected to the stage 1,
When all the data are collected, the output of the weighted sum circuit 3 by the transversal filter is determined, and the charge output means 5
Is output to the node 6 as a charge signal.

Although not shown in particular, even if the weighted sum circuit 3 is directly installed on the arc of the SDB without passing through the stage 1 and the shift register 2, an equivalent function can be realized, and it is necessary. Accordingly, the signal moving on the arc can be used nondestructively.

In FIG. 7, the optical output is driven through the modulator 7 to drive the LED 8 to output the calculation result in the form of an optical signal. Further, FIG. 8 shows an example of an optical system in which this device is incorporated.

In FIG. 8, 1 is an optical filter device, 2
Is a primary imaging system for capturing external information, 3 is a circuit including the SDB according to the present invention, 4 is a mirror device for selectively reflecting output light, and 5 is a secondary connection for extracting output information to the outside. The image system 6 indicates a detection device for detecting the output signal.

In summary, according to one theory, it is considered that about 10 billion neurons are connected to other neurons by 1000 to 100,000 synaptic connections to constitute the whole human brain nerve tissue. ing.

The neural network aims at constructing an information processing system suitable for processing pattern information and the like, which the conventional Neumann computer is not good at, based on the structure and function of such neural tissue. .

At present, it is practically impossible to construct a system of a scale similar to that of an actual nerve tissue, but similar functions are configured on a smaller scale to simulate a living body and apply it. There is a situation where the possibility is being explored.

When such a highly parallel system is realized on hardware, since it has a structure for processing a large amount of data in parallel by imitating an actual neural tissue, an enormous amount of data equivalent to synapse connection is obtained. However, the amount of this wiring is one of the important factors that determine the scale of the system, and the resources such as wiring and arithmetic circuits corresponding to synapses are used in a time-divisional manner. Avoiding these difficulties is also being considered.

However, in the case of a conventional network using hardware, it is often difficult to change the wiring once the wiring has been formed, and it is considered that there are many difficulties in multiple use of resources.

In the present invention, these wirings are replaced by a spatial data bus using a charge transfer element, and the dispersed spatial data is rearranged in an array suitable for calculation.
In exchange for a slight time delay required for rearrangement, it is possible to drastically reduce wiring as a substance and avoid various difficulties caused by an increase in the number of wiring. It is possible to give a degree of freedom to change the connection equivalently through selection, and it is a great merit that resources such as arithmetic circuits can be used in multiple cases in some cases.

For example, the function of the horizontal cell layer, which is said to calculate the local average intensity of optical stimulation in the retinal tissue, is SD.
The operation realized via B can be easily realized by using the configuration shown in FIG. 6 and FIG. 7, and further, to realize the function of the bipolar cell layer connected thereto,
This can be easily realized by simply installing an arithmetic circuit for calculating the difference between the logarithmic output signal shown in FIG. 7 and the logarithmic output signal of the photoelectric conversion element provided in or around the cell. To be done.

Note that the above-mentioned function has already been performed by C.I. A simulation experiment using hardware configured by combining a CMOS analog circuit with a resistor network has been carried out by Mead, and steady results have been achieved, but since the circuit becomes complicated, it is currently 50 × 50 pixels. Since it remains in a relatively small-scale system, and the horizontal cell layer is approximated by a simple resistance network, there is also a concern that performance restrictions due to this approximation may be imposed.

On the other hand, in the structure using SDB, CCD technology capable of high-density integration can be adopted, which is extremely advantageous in constructing a large-scale system.
Further, not only adjacent signals but also spatially dispersed signals can be collected around the arithmetic circuit, so that it is possible to incorporate a function closer to an actual living body.

Even if the arithmetic circuit is spatially fixed, the signal is not connected to a specific input source by a fixed connection, so that the necessary input signals are gathered around the arithmetic unit. Since the output at that time can be obtained by executing the operation at that time, there is a degree of freedom to select a configuration in which the same kind of arithmetic processing is multiply applied to the data groups of different types of input sources. If a plurality of arithmetic functions are assigned to one cell and switching control of arithmetic processing is given,
Naturally, this degree of freedom will be further improved.

[0060]

As is apparent from the above description, the present invention has the following effects. (1) Spatially dispersed information can be gathered around the arithmetic unit without passing through fixed wiring, and at the same time the arithmetic result can be distributed to spatially separated places. It is possible to reduce a huge amount of wiring, which was inevitable with technology. (2) The mode of “collection” and “distribution” can be selectively changed by changing the system driving method, input sampling, operation timing, and the like. (3) Since it can be realized by a CCD or the like, direct input of optical information and high-density integration are relatively easy within the range of conventional techniques.

[Brief description of drawings]

FIG. 1 is a relationship diagram showing the correspondence between the physical configurations of a retinal tissue and a three-layer perceptron.

FIG. 2 is a structural diagram showing the structure of an SDB according to an embodiment of the present invention.

FIG. 3 is a trajectory diagram showing a route followed by a signal at a certain point in the SDB according to the embodiment of the present invention.

FIG. 4 shows an initial point S (N, 0) gathering in a cell C of an SDB according to an embodiment of the present invention, and shows movement of a signal by a predetermined operation.

FIG. 5 shows an example of an SDB having nodes and arcs of the present invention, showing a signal trajectory when a predetermined designated shift is applied.

FIG. 6 shows a signal trajectory when a predetermined designated shift is applied to an example of the SDB having nodes and arcs of the present invention.

FIG. 7 shows the output of one embodiment of the present invention as LE through a modulator.
It is a block diagram of the arithmetic cell which has both the optical output and the electric charge output which also equipped the structure which drives D and outputs the arithmetic result in the form of an optical signal.

FIG. 8: SD equipped with optical output means of one embodiment of the present invention
It is a block diagram which shows the output detection means of B.

Claims (6)

[Claims] 1. A plurality of route groups each including a plurality of charge transfer element arrays arranged along a plurality of routes which do not intersect each other are arranged at positions intersecting each other, and the resultant is located at an intersection of the routes. A plurality of charge transfer element columns involved in the intersection, and each path group is configured to perform a transfer operation along the path group in different time intervals. Relocation device for spatial information. 2. The number of the route groups is three or more, and each intersection of the routes is mainly constituted by the intersection of only two routes.
Relocation device for the described spatial information. 3. The spatial information rearrangement apparatus according to claim 1, wherein a photoelectric conversion element is provided in a part or all of the charge transfer element, and the optical stimulus from the outside is directly taken in as a charge signal. 4. The spatial information rearrangement apparatus according to claim 1, wherein arithmetic units for performing various calculations based on charge signals held by a plurality of adjacent charge transfer elements are dispersedly arranged. 5. The spatial information rearrangement device according to claim 4, further comprising a mechanism for returning the charge signal of the calculation result to the charge transfer element on the path. 6. The spatial information rearrangement apparatus according to claim 4, further comprising a mechanism for converting an electromagnetic wave output signal such as infrared rays, which is modulated by a signal of a calculation result, into the electromagnetic wave output signal.