JP3275311B2 - Creative devices and neural networks - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a creation device for creating a device having a new function by artificial intelligence technology, and a neural network suitable for realizing the creation device.

<Conventional technology> Creation of a new device having a new function due to new needs, etc., up to now, has been performed by a person having basic knowledge of the device by a personal or group creative activity. Is being done.

<< Problems to be Solved by the Invention >> Conventionally, when a new device is originally created, it is necessary to rely on a superior human brain. Time and money are needed.

The present invention has been made in view of the above-described problems, and provides a creative device that automatically and creatively creates a new functional device mechanically by utilizing artificial intelligence technology, and a neural network suitable for realizing the creative device. It is an object.

<< Means for Solving the Problems >> In order to achieve the above object, the present invention provides an internal structure by receiving input data relating to a conceived new functional device and expected output data corresponding to the input data. Learning means for systematically learning; state information detecting means for detecting state information of the learning means; a state information pattern corresponding to a known basic function and a state information pattern corresponding to a basic combination form of the known basic function; A matching pattern detecting means for detecting a region similar to the state information pattern stored in the basic pattern storing means in the state information detected by the state information detecting means; FIG. 9 shows a configuration of a new function device that generates an output equivalent to the expected output data by inputting the input data based on a detection result by the means. And a configuration data generating unit for outputting data, wherein the learning unit is a neural network, and is configured to self-organizely learn an internal structure by changing a synaptic load between neurons, and the neural network has a multilayer structure. None, configured to execute a back propagation learning method, wherein the synapse of the neural network includes a light-emitting element that changes the amount of light emission according to a synapse load, and the state information detecting unit detects a light-emitting state of the light-emitting element. The image processing apparatus further includes imaging means for imaging as state information in the neural network, wherein the basic pattern storage means includes a state information pattern corresponding to a known basic function and a state information pattern corresponding to a basic combination form of the known basic function. Is stored as an image pattern, and the matching means And it is characterized in that it is configured to detect a similar area and the state information pattern state the basic pattern storage means at the information stored in the imaged in the neural network by.

Further, in addition to the means as described above, when a region similar to the state information pattern stored in the basic pattern storage unit is detected by the matching unit, the creation device according to the present invention may generate a state information pattern similar to the region. It may have pattern writing means for writing.

In the above-described creative apparatus, the synapse of the neural network includes a light receiving element that changes a synapse load according to the amount of received light, and further includes a state information in the neural network imaged by the imaging unit. When a region similar to the state information pattern stored in the basic pattern storage unit is detected by the matching unit, the state information pattern stored in the basic pattern storage unit for the light receiving element in that region is detected. An optical state information pattern writing means for irradiating light according to the similar state information pattern may be provided.

According to the present invention, it is an object of the present invention to provide a neural network suitable for realizing a creative device as described above, and a light-emitting element that changes a light emission amount according to a synapse load and a light-receiving element that changes a synapse load according to a light reception amount And a synapse that optically displays state information in the neural network and has an optically variable synapse load.

<< Operation >> According to the configuration as described above, the input structure and the expected output data corresponding to the input data of the new functional device envisioned in the learning means such as a neural network are given, so that the internal structure of the learning means is changed. Self-organized learning is performed, state information of the learning means is detected by the state information detecting means, and state information patterns (model patterns) stored in the basic pattern storing means are stored in the state information detected by the state information detecting means. ) Is detected by the matching means, and based on the result of the detection, data indicating the configuration of the new functional device which produces an output equivalent to the expected output data by inputting the input data by the configuration data generating means. Is output.

The synapse of the neural network includes a light emitting element that changes a light emission amount according to a synapse load, and the state information detecting means includes imaging means for imaging a light emitting state of the light emitting element as state information in the neural network. If the basic pattern storage means stores a state information pattern corresponding to a known basic function and a state information pattern corresponding to a basic combination form of the known basic functions as an image pattern, the matching means In the state information in the neural network imaged by the imaging means, an area similar to the state information pattern stored in the basic pattern storage means is detected by pattern recognition.

Furthermore, when a region similar to the state information pattern is detected by the matching unit, the optical state information pattern writing unit stores the state information pattern stored in the basic pattern storage unit for the light receiving element in that region. Irradiating light according to the similar state information pattern to change the synapse load is performed, whereby the area of the neural network is replaced with the similar known state information pattern, and the known state information pattern is changed. Creation of new functional devices by combination is promoted.

In other words, the creative device according to the present invention having the above-described configuration generally has a unique process for creating a device having the following four steps when a human creates a device having a new function. It is constructed based on various theories.

First Step Assume the desired, or in other words, the required inputs and outputs of the device with the new function.

Second step By intensively thinking about the inputs and outputs that should be, a place where the relationship between the inputs and outputs is simulated in the human right brain is temporarily self-organized. This field is called the sensuous field, and the sensuous field is a spatial distribution of synaptic connection strength coefficients (synaptic loads), which are connections between neurons in the right brain, and expresses an information processing function.

Third step Pattern recognition is performed on the part of the right brain other than the sensibility field using the sensibility field as a recognition target pattern.

Fourth step In the distribution of the known basic functional blocks with respect to the local partial pattern of the sensitivity field, the pattern of the most similar known basic functional block is forcibly applied to the local partial pattern of the sensitivity field. Replace the pattern with its similar basic function block pattern. Then, returning to the second step, the self-organization is executed again. As a result, if the sensation field is represented by a combination of known basic functional blocks, creation of a new functional device is completed.

<< Example >> Hereinafter, the present invention will be described in detail with reference to the accompanying drawings.

FIG. 1 shows an embodiment of a creation device according to the present invention. The creation device has a neural network 1 having a multilayer structure.

In this embodiment, the neural network 1 is configured as a back propagation type neural network having a four-layer seven-stage structure, and the neurons Θij of the first to third layers are adjacent as shown in FIG. A variable synapse connection is established between all the neurons in the next layer and a synapse load (weight) W (n) jk. Where the code Θij
And the meaning of W (n) jk are defined.

Θij: the j-th (stage) neuron of the first layer W (n) jk: the connection synapse between the j-th neuron ｎnj of the n-th layer and the k-th neuron （(n + 1) k of the (n + 1) -th layer Synapse weight The back propagation type neural network 1 is composed of seven neurons Θ _{11 to} Θ _{17 in} the first layer (input layer).
Each of the new features apparatus was conceived, ie given the input data X ₁ to X ₇ of the new features devices to be created, the final layer 7 neurons theta ₇₁ through? The fourth layer is a (output layer) ₇₇
Each was from output data Y ₁ to Y ₇ to the output port 3 ₁ to 3 _7, also as teacher input, the input data X ₁ to X ₇ output port 3 ₁ expected output data D ₁ to D ₇ corresponding to It is given to 3 _7,
The learning signal obtained based on the sum of errors of the expected output data D _{1 to} D ₇ and the output data Y _{1 to} Y ₇ based on the sum of error squares is back-propagated to each joint synapse from the output layer side, thereby reducing the internal structure. It functions as a learning means for self-organized learning.

FIG. 3 shows a specific embodiment of a device for variable synapse connection used in the neural network 1. This variable synapse coupling device has a field-effect transistor FET, a light emitting diode LED, and a photodiode PD, and a terminal Te is provided with a learning signal by backpropagation of a synapse load, and the terminals T ₁ and T ₂
Is connected to the neuron.

In the variable synapse coupling device shown in FIG. 3, when a positive voltage is applied to the terminal Te, a new charge is accumulated in the capacitor C in response to the applied voltage. The gate voltage of the type transistor FET rises, and the resistance between the drain D and the source S changes according to the gate voltage, so that the synapse load is variably set, and the light emitting diode LED is connected to the gate of the field effect transistor FET. When a source-to-source voltage is applied, light is emitted with luminance according to the synapse load.
Further, in the variable synapse coupling device, when light is incident on the photodiode PD, the resistance value of the light is reduced, and the electric charge is newly stored in the capacitor C in response to this. Field-effect transistor FE
The gate voltage of T rises, and its drain D and source S
And the resistance value changes. Note that the wavelength of light received by the photodiode PD and the emission wavelength of the light emitting diode LED are different from each other, so that the light of the light emitting diode LED does not change the resistance value of the photodiode PD.

The synapse load W (n) jk in FIG. 1 may be interpreted as being replaced by the light emitting portion of the light emitting diode LED indicating this, and the photodiode PD is denoted by the symbol P (n) in FIG. ) Jk. Here, the meaning of the code P (n) jk is newly defined.

P (n) jk: Light receiving part of optical signal for forcibly changing and setting synapse load W (n) jk from outside Light emitting part (synapse load) W of light emitting diode LED
As shown in FIG. 1, the (n) jk and the light receiving portion P (n) jk of the photodiode PD are adjacent to each other with the same number and are regularly arranged in a matrix form for each layer and each stage. Is fixedly arranged. In FIG. 1, neurons are indicated by ◎, light emitting portions of light emitting diodes are indicated by ○, and light receiving portions of photodiodes are indicated by ●.

Light-emitting diode LEDs fixedly arranged in a matrix
The whole light-emitting unit W (n) jk is imaged by the two-dimensional imaging device 5 as the state information detecting means.

The two-dimensional imaging device 5 may be a CCD television camera or the like, which captures the entire light-emitting portion W (n) jk of the light-emitting diode LEDs fixedly arranged in a matrix as the above-described sensible field, and Each of the light emitting units W (n) jk captures a two-dimensional image pattern created by emitting light at a luminance related to the magnitude of the synaptic load.

Photodiode PD fixedly arranged in a matrix
The light receiving portions P (n) jk are individually and selectively irradiated with light from the optical pattern writing device 7.

The optical pattern writing device 7 individually transmits light to each of the light receiving portions P (n) jk of the photodiodes PD fixedly arranged in a matrix based on a writing instruction to be described later in order to change a synaptic load. It is configured to selectively irradiate, and may be configured by, for example, a digital illuminating device or a scanning laser beam irradiating device as disclosed in JP-A-61-226808.

The creation device includes a model pattern database 9 as basic pattern storage means for storing a plurality of state information patterns corresponding to known basic functions and a plurality of state information patterns corresponding to basic combinations of known basic functions in image patterns. have.

The image pattern of the state information pattern corresponding to the known basic function and the state information pattern corresponding to the basic combination form of the known basic function stored in the model pattern database 9 is the basic pattern of the known basic function or the basic function of the known basic function. When data relating to the connection form is given to the neural network 1, the light emitting diode according to each synapse load is represented by a matrix formed by the light emitting portions W (n) jk of the light emitting diode LED.
It is obtained based on the light emitting pattern obtained by the light emitting portion W (n) jk of the LED emitting light.

FIGS. 4 to 7 show examples of image patterns corresponding to the known basic functions, and FIGS. 8 and 9 show examples of image patterns corresponding to the basic combination form of the known basic functions. ing. The image pattern shown in FIG. 4 is a model pattern of the basic function A, the image pattern shown in FIG. 5 is a model pattern of the basic function B, and the image pattern shown in FIG. The model pattern, the image pattern shown in FIG.
The image pattern shown in the figure is a model pattern of the basic combination form N1, and the image pattern shown in FIG.
The model pattern is N2. FIG. 10 shows the actual connection state of the basic connection form N1 in the model pattern shown in FIG. 8, and FIG. 11 shows the actual connection state of the basic connection form N2 in the model pattern shown in FIG. Actual connection states are shown.

Information on the two-dimensional image pattern of the neural network 1 captured by the two-dimensional imaging device 5 and information on the above-described model pattern stored in the model pattern database 9 are input to an image processing device 11 as matching means.

The image processing device 11 performs image processing for expressing the structure of a two-dimensional image pattern (light emission pattern) of the neural network 1 captured by the two-dimensional imaging device 5 as a combination of known basic model patterns as described above. Specifically, in the state information in the neural network 1 imaged by the two-dimensional imaging device 5, an area similar to the above-described model pattern stored in the model pattern database 9 is detected. As shown in FIG. 12, image structure data indicating the matching coordinate position and the matching value is output to the writing pattern setting device 13 and the image structure data evaluation device 15. The pattern matching by the image processing device 11 may be performed in a two-dimensional image pattern of the neural network 1 in a scanning manner with a predetermined scanning range, thereby enabling high-speed pattern matching.

The writing pattern setting device 13 is given image structure data from the image processing device 11 and model pattern information from the model pattern database 9, and writes the image pattern to be forcibly written into the neural network 1 as a whole or Partially determined, a write instruction is output to the optical pattern writing device 7.

The image structure data evaluation device 15 is provided with the image structure data from the image processing device 11 and the signal E indicating the sum of error squares from the output unit of the neural network 1, and based on the error sum of signal E indicating the learning result, the image structure data is evaluated. The data is evaluated. If the evaluation result is equal to or more than a predetermined value, the image structure data is converted into functional configuration data as shown in FIG. 13 and output to a printer, a display, an external memory, etc. Output of an end signal due to completion to the image processing apparatus 11 is performed.

Converting image structure data to functional configuration data
From the matching center coordinate position of the basic function, the coordinate position of the terminal of the basic function is found, and similarly, the coordinate position of the terminal of the basic connection form is found. This is performed by connecting the terminals that are close to each other.

 Note that, in FIG. 13, the circles indicate terminal connection.

The function configuration data is data on various functions necessary for the configuration of the new function device to be conceived, that is, the new function device to be created, and the interconnection of the functions. The device is embodied.

The structure of the function configuration data may be such that the connection destination of each terminal is indicated by a pointer, in addition to the matrix configuration as shown in FIG.

FIG. 14 shows a light-emitting portion W (n) jk of a light-emitting diode LED in a neural network 1 for learning by self-organization.
2 shows an example of a two-dimensional image pattern using a matrix. This two-dimensional image pattern shows the distribution of the synapse load, and in this case, there are partial patterns indicated by reference signs a to e.

The partial patterns a to e may correspond to the model patterns of the basic functions A to D or the model patterns of the basic combination forms N1 and N2, and this determination is made by the pattern matching as described above. . In this case, the partial pattern a is a model pattern of the basic function A, the partial pattern b is a model pattern of the basic function B, and the partial pattern c
Is similar to the model pattern of the basic function C, the partial pattern d is similar to the model pattern of the basic function D, and the partial pattern e is similar to the model pattern of the basic combination form N1 by a predetermined reference value or more. Functional configuration data indicating a combination of basic functions as shown in the figure is obtained.

Model pattern database 9, image processing device 11, writing pattern setting device 13, and image structure data evaluation device 15
May be constituted by a microcomputer,
One embodiment of this is shown in FIG. The microcomputer includes a CPU 21, a ROM 23 for storing a system program and the like, a RAM 25 for storing various data, a camera interface unit 27 to which a video signal is supplied from the two-dimensional imaging device 5, and a video signal from the camera interface unit 27. An image memory 29 for taking in and storing the same, a scan range setting memory 30 for storing a pattern matching scan range, and a write pattern setting for storing a write pattern write instruction and outputting the same to the optical pattern writing device 7 A memory 31; a latch 32 for outputting input data and expected output data to the neural network 1; a latch 33 for receiving an error square signal from an output of the neural network 1; Multiple sets of data and expected output data And also a hard disk unit 35 and a hard disk interface portion 37 which forms a model pattern database 9, and a display interface unit 39 is connected to the CRT of such a display 41 by the display interface 39.

Next, an example of the operation procedure of the creation device according to the present invention will be described with reference to the flowchart shown in FIG.

First, in step 10, the count value SS of the trial counter is returned to 1, and the entire range of the image memory 29 is initialized to a scanning range for pattern matching. Then, the process proceeds to step 15, The input data and the expected output data are input to the neural network 1,
Performing learning is performed.

Next, in step 20, the two-dimensional imaging device 5 captures a two-dimensional image pattern using a matrix of the light-emitting portions W (n) jk of the light-emitting diodes LED in the neural network 1.

Next, at step 30, data of the captured two-dimensional image pattern of the neural network 1 is stored in the image memory 29.

Next, in step 40, the hard disk unit
The model pattern of the basic function as described above stored in the model pattern database 9 is read out.

Next, in step 50, the data stored in the image memory 29 storing the captured two-dimensional image pattern of the neural network 1 is scanned with the model pattern of the basic function, and pattern matching detection is performed. This scanning range is specified by the scanning range setting memory 30.

Next, in step 60, the name of the model pattern of the basic function indicating the matching value equal to or greater than the predetermined threshold value and the data representing the matching coordinate position thereof are written in the RAM 25.

Next, in step 70, it is determined whether or not scanning for pattern matching detection has been completed for all of the basic function model patterns held in the model pattern database 9. If it is determined that the scanning for pattern matching detection has been completed for all of the basic function model patterns, the process proceeds to step 80; otherwise, the process returns to step 40.

In step 80, the model patterns in the above-described basic combination form stored in the hard disk unit 35 as the model pattern database 9 are sequentially read from the hard disk unit 35.

Next, in step 90, the data stored in the image memory 29 that stores the captured two-dimensional image pattern of the neural network 1 is scanned with the model pattern in the basic combination form, and pattern matching detection is performed. .

Next, in step 100, the name of the model pattern in the basic combination form having a matching value equal to or greater than a predetermined threshold value and data representing the matching coordinate position thereof are written in the RAM 25.

Next, in step 110, it is determined whether or not scanning for pattern matching detection has been completed for all of the model patterns in the basic combination form held in the model pattern database 9. If it is determined that the scanning for pattern matching detection has been completed for all of the model patterns in the basic combination form, the process proceeds to step 120; otherwise, the process returns to step 80.

In step 120, the model pattern having the matching value larger than SS among the model patterns of the basic function or the basic combination form showing the matching value equal to or more than the predetermined threshold value and the matching coordinate position thereof are represented. Extracting data is performed.

Next, in step 130, the data representing the name of the model pattern extracted in step 120 and its matching coordinate position are written in the write pattern setting memory 31.

Next, in step 140, the optical pattern writing device 7 according to the data written in the write pattern setting memory 31, the light receiving portion P (n) jk of the photodiode PD fixedly arranged in a matrix of the neural network 1 by the optical pattern writing device 7. Are individually and selectively irradiated with light in order to change the synapse load, that is, a model pattern similar to the neural network 1 is forcibly written in the matching area. The writing of the model pattern may be performed on an area matched around the matching position.

Next, in step 150, the sum of error squares E from the output unit of the neural network 1 is fetched.

Next, at step 160, it is determined whether or not the error sum of squares E has dropped below a predetermined threshold value. When it is determined that the error sum of squares E has fallen below a predetermined threshold value, the process proceeds to step 190; otherwise, the process proceeds to step 170.

In step 170, the count value SS of the counter is incremented.

Next, in step 180, the count value S of the counter
It is determined whether or not S has reached a predetermined maximum value. When the count value SS of the counter reaches the predetermined maximum value, the process proceeds to step 190, and otherwise, the process proceeds to step 185.

In step 185, the contents of the scan range setting memory 30 are rewritten so that the area in which the model pattern has been written in step 140 is set as the matching scanning prohibited area. After step 185, the process returns to step 15.

In step 190, the matched basic functions,
The function configuration data corresponding to the two-dimensional image pattern (distribution pattern) of the neural network 1 is created from the matching coordinate position of the model pattern in the basic combination form, and this routine is ended.

<< Effects of the Invention >> As is clear from the above description, in the creative device according to the present invention, the input data relating to the new functional device envisioned in the learning means such as a neural network and the expected output data corresponding to this input data Is given, the internal structure of the learning means is learned in a self-organizing manner, the state information of the learning means is detected by the state information detecting means, and the basic pattern is stored in the state information detected by the state information detecting means. Pattern matching for detecting an area similar to the model state information pattern stored in the means is performed by the matching means. Based on the detection result, the configuration data generating means inputs the input data and outputs the same output data as the expected output data. Data indicating the configuration of the new function device that causes the The data necessary for realization has been obtained, and this has been created.

The synapse of the neural network includes a light emitting element that changes a light emission amount according to a synapse load, and the state information detecting means includes imaging means for imaging a light emitting state of the light emitting element as state information in the neural network. If the basic pattern storage means stores a state information pattern corresponding to a known basic function and a state information pattern corresponding to a basic combination form of the known basic functions as an image pattern, the matching means In the state information in the neural network imaged by the imaging means, an area similar to the state information pattern stored in the basic pattern storage means is detected by optical pattern recognition, Reliable and efficient pattern matching by using the same technology as pattern recognition Ku made so that the line that may be in high speed.

Furthermore, when a region similar to the state information pattern is detected by the matching unit, the optical state information pattern writing unit stores the state information pattern stored in the basic pattern storage unit for the light receiving element in that region. Irradiating light according to the similar state information pattern to change the synapse load is performed, whereby the area of the internal structure of the neural network is forcibly replaced with the similar known state information pattern,
This facilitates creation of a new function device by combining known state information patterns.

[Brief description of the drawings]

FIG. 1 is a schematic configuration diagram showing one embodiment of a creation device according to the present invention. FIG. 2 is a schematic configuration diagram showing one embodiment of a back propagation type neural network used in the creation device according to the present invention. Is an electric circuit diagram showing a specific embodiment of a variable synapse coupling device used in a neural network, and FIGS. 4 to 7 each show an example of an image pattern (model pattern) corresponding to a known basic function. FIG. 8, FIG. 8 and FIG. 9 are explanatory diagrams each showing an example of an image pattern (model pattern) corresponding to a basic combination form of known basic functions, and FIG. 10 and FIG.
FIG. 9 is an explanatory view showing the actual connection state of the basic connection form in the model pattern shown in FIG. 9 and FIG. 9, FIG. 12 is an explanatory view showing the image structure data structure, and FIG. 13 is a functional configuration data structure FIG. 14 is an explanatory diagram showing an example of a two-dimensional image pattern in a neural network after learning by self-organization, and FIG. 15 is a diagram showing basic functions generated from the two-dimensional image pattern in the neural network. FIG. 16 is a block diagram showing one embodiment of a microcomputer used in the creation device according to the present invention, and FIG. 17 is a flowchart showing an example of an operation procedure of the creation device according to the present invention. . Θij... Neuron W (n) jk... Synapse load (light-emitting portion of light-emitting diode) P (n) jk... Light-receiving portion of photodiode 1... Neural network 31-37... Output port 5. Apparatus 7 Optical pattern writing apparatus 9 Model pattern database 11 Image processing apparatus 13 Writing pattern setting apparatus 15 Image structure data evaluation apparatus 21 CPU 23 ROM 25 RAM 27 ... Camera interface unit 29 ... Image memory 30 ... Scan setting range memory 31 ... Write pattern setting memory 32 ... Latch unit 33 ... Latch unit 35 ... Hard disk unit 37 ... Hard disk interface unit 39 ... Display interface Part 41: Display

Continuation of the front page (58) Field surveyed (Int.Cl. ⁷ , DB name) G06F 3/02-3/10 G06G 7/60 G06T 7/00 JICST file (JOIS)

Claims (4)

(57) [Claims] 1. A learning means for self-organizingly learning an internal structure by receiving input data relating to an envisaged new function device and expected output data corresponding to the input data; State information detecting means for detecting, basic pattern storing means for storing a state information pattern corresponding to a known basic function and a state information pattern corresponding to a basic combination form of the known basic function, and detecting by the state information detecting means Matching means for detecting an area similar to the state information pattern stored in the basic pattern storage means in the obtained state information; and inputting the input data based on a result of the detection by the matching means. Configuration data generating means for outputting data indicating the configuration of the new function device that generates an output equivalent to The learning means is a neural network, and is configured to self-organize learning of an internal structure by changing a synaptic load between neurons.The neural network has a multilayer structure and is configured to execute a back propagation learning method. The synapse of the neural network includes a light-emitting element that changes the amount of light emission according to a synapse load, and the state information detecting means images the light-emitting state of the light-emitting element as state information in the neural network. The basic pattern storage unit stores a state information pattern corresponding to a known basic function and a state information pattern corresponding to a basic combination form of the known basic function as an image pattern, and the matching unit stores the image pattern. Neural network imaged by the optimizing means Creation device at the state information, wherein the basic pattern storage means is configured to detect the state information pattern similar areas are stored. 2. A learning means for self-organizingly learning an internal structure by receiving input data relating to a conceived new function device and expected output data corresponding to the input data; State information detecting means for detecting; basic pattern storing means for storing a state information pattern corresponding to a known basic function and a state information pattern corresponding to a basic combination form of the known basic function; detection by the state information detecting means Matching means for detecting an area similar to the state information pattern stored in the basic pattern storage means in the obtained state information; and an area similar to the state information pattern stored in the basic pattern storage means. Pattern writing means for writing a similar state information pattern to the area when is detected by the matching means, A configuration data generating unit that outputs data indicating a configuration of a new function device that generates an output equivalent to the expected output data by inputting the input data based on a detection result by the switching unit, wherein the learning unit is a neural network. Yes, it is configured to self-organize learning of the internal structure by changing the synaptic load between neurons, the neural network has a multilayer structure, and is configured to execute the back propagation learning method, and the synapse of the neural network is A light emitting element that changes a light emission amount according to a synaptic load; the state information detecting means includes imaging means for imaging a light emitting state of the light emitting element as state information in the neural network; The storage means stores state information patterns corresponding to known basic functions. And a state information pattern corresponding to a basic combination form of a known basic function as an image pattern, wherein the matching means includes the basic pattern in state information in the neural network imaged by the imaging means. A creation device characterized by detecting a region similar to the state information pattern stored in a storage means. 3. The neural network according to claim 2, wherein said synapse includes a light receiving element for changing a synapse load according to a received light amount, and said basic pattern storage means based on state information in said neural network imaged by said imaging means. When an area similar to the state information pattern stored in the area information is detected by the matching means, the similar state information among the state information patterns stored in the basic pattern storage means for the light receiving elements in that area. 3. The apparatus according to claim 1, further comprising an optical state information pattern writing unit that irradiates light according to a pattern. And a light receiving element for changing a synapse load according to a received light amount, and optically displaying state information in the neural network. A neural network having a synapse in which a synapse load is variably set.