JPH022537A - Optical learning and recognizing machine - Google Patents

Abstract

PURPOSE:To accomplish an optical learning and recognizing machine optically processing intellectual information by composing a perceptive layer, a joined layer and a response layer of nonvolatile space optical modulators and threshold processing circuits. CONSTITUTION:In the perceptron-structured, optical learning and recognizing machine provided with a learning function, the perceptive layer 11, the joined layer 12 and the response layer 13 are composed of the nonvolatile space optical modulator and the threshold processing circuits. Namely, the recognition processing layers 11-13 in the learning and recognizing machine are designed to mixingly employ the nonvolatile space optical modulators made of optical elements and the threshold processing circuits of an electronic device and to wire the layers by means of optical space wiring and electrical signal wiring, which passes through the electronic device of a part of the threshold processing circuits. As a result, the complicated layer wiring for recognition processing can be accomplished by optical space wiring, whereby trouble arising when entangled wiring for accomplishing complicated recognition processing is carried out by metallic wiring can be evaded.

Description

[Detailed Description of the Invention] [Field of Industrial Application] The present invention relates to an optical learning machine, and in particular to an element of an information processing device that processes information in parallel following the function of a human neural network, a so-called neurocomputer. This article concerns an optical station machine and its learning method.

[Conventional technology]

Research is underway into devices that simulate the superior information processing capabilities of the human brain. An example of such a device is a neurocomputer. Since neurocomputers process information in parallel by connecting a large number of processors called neurons, they are suitable for processing that targets the recognition (memory, perception, thinking) of a single human brain's functions. Furthermore, the neurocomputer can also have a learning function.

A recognition machine called a bersebutron has been proposed as one of the models of such a neurocomputer. The perceptron is composed of three basic units (layers), as shown in FIG.

FIG. 10 is a conceptual diagram of a perceptron. In FIG. 10, 71 is a sensory layer composed of a plurality of neuron elements, 72 is an association layer composed of a plurality of neuron elements, and 73 is a response layer composed of one neuron. Further, 74 is a neuron connection between the sensory layer 71 and the association layer 72, and 75 is a neuron connection between the association layer 72 and the response layer 73.

76 is an input signal, 77 is an output signal, and 78 is a teacher signal. Normally, the association layer 72 has a structure consisting of a plurality of layers each consisting of a plurality of neurons, and is a multilayer structure. The structure is as follows. Union layer 72 becomes 1!
J's is called a simple persebutron.

The mechanism of Bersebtron is as follows.

Feel it! The input signal received by the neuron 71 is received by the neuron of the association layer 72 via the neuron connection 74, and the excitation pattern of the neuron group of the sensory JPj 71 is converted into the excitation pattern of the neuron group of the association layer 72 and transmitted. . The signal then enters the response M73, where only one neuron of the response layer 73 discriminates and answers the received information. In such a perceptron, the correct answer to the input is learned in advance from a teacher signal, and the connections 74 of neurons between the sensory layer 71 and the association layer 72 are made so that the output signal pattern approaches the correct answer. This recognition machine is characterized by having a so-called learning function that modifies the neuron connections 75 between the association layer 72 and the response layer 73.

[Problem to be solved by the invention]

By the way, a neurocomputer like this Persebtron can actually be constructed using electronic device hardware,
In order to construct a practically meaningful machine with somewhat complex cognitive functions, it is necessary to interconnect processors equivalent to billions of neurons, which are connected to the processors of electronic devices such as transistors. If you try to configure it with , the number of wires will be enormous. In addition, not only does the number of wires become enormous, but various problems arise, such as signal propagation delay time in the wires, mutual interference, and insufficient aF range. Currently, hardware that can perform massively parallel execution has not been realized due to the difficulty of connection.

On the other hand, if optical signal processing technology is used in neurocomputers, it is expected that neurocomputers can be constructed by taking advantage of the inherent parallelism of optical signals, ease of spatial wiring, high density, and noncoherence. . Based on this idea, several proposals have been made to optically realize various neurocomputer models such as perceptrons. However, optical transistor-like elements for constructing optical signal processing processors of neuron elements are still in the development stage, and the number of devices that can be used is limited. In addition, due to reasons such as the immaturity of optical signal wiring technology,
Regarding this type of neurocomputer, the current situation is that there is still little progress in the development of hardware devices.

As described above, the future development of optical neurocomputers that have a learning function and can perform complex cognitive processing is awaited.

The present invention has been made to solve this problem.

An object of the present invention is to provide an optical recognition machine that has a learning function and optically performs intelligent information processing such as recognition of a plurality of patterns.

Another object of the present invention is to provide an optical learning mechanical device that optically performs intelligent information processing such as complex pattern recognition that requires a learning function by utilizing the advantages of light such as parallelism and high speed. It's about doing.

The above and other objects and novel features of the present invention will become apparent from the description of this specification and the accompanying drawings.

[Means to solve the problem]

In order to achieve the above object, the present invention includes a sensory layer consisting of a plurality of neuron elements that receive input signals, and a plurality of neurons that convert the excitation pattern received from the sensory layer according to a learned transmission amount and connect it to the next layer. In a learning recognition machine that has a learning function, it has a perceptron structure consisting of an association layer consisting of elements and a response layer consisting of neuron elements that discriminate and judge the excitation pattern received from the association layer and output an answer output signal. Each of the layers, the association layer, and the response layer is configured with a nonvolatile spatial light modulator and a threshold processing circuit.

This nonvolatile spatial light modulator is a microchannel spatial light modulator that stores a synaptic connection matrix whose elements are the weights of connections between pattern pixels corresponding to neuron elements, and the threshold processing circuit is composed of an electronic device. This is a threshold processing device that connects a non-volatile spatial light modulator and a threshold processing circuit using a light emitting element array that converts an electrical pattern signal into an optical signal, and a light receiving element array that converts the optical signal into an electrical signal. This is done via signal paths of electrical signal connections and optical signal connections.

[Effect]

According to the above means, in the learning recognition machine having a perceptron structure and a learning function, each of the sensory layer, association layer, and response layer is configured by a nonvolatile spatial light modulator and a threshold processing circuit. As a result, learning P! Each recognition processing layer of the recognition machine is configured to use a mixture of a nonvolatile spatial light modulator of an optical element and a threshold processing circuit of an electronic device,
The wiring connecting each layer can be configured using optical space wiring and electrical signal wiring via some electronic devices of the threshold processing circuit. For this reason, all the wiring for complex recognition processing in each layer can be realized using optical spatial wiring, which eliminates the complications that arise when the complicated wiring for realizing complex recognition processing is done using metal wiring. You can avoid this.

In other words, when a processor corresponding to the neuron elements of a neurocomputer is constructed using an electronic device, the neuron elements are interconnected using metal wiring, resulting in complicated wiring.
Using a mixture of optical elements such as spatial light modulators and electronic devices, the wiring for coupling between the main layers of recognition processing is achieved entirely optically using spatial optical wiring, which eliminates the difficulties that arise with metal wiring. You can avoid this.

〔Example〕

An embodiment of the present invention will be specifically described below with reference to the drawings.

In addition, in all the figures for explaining the embodiment, the same elements are given the same reference numerals, and repeated explanations thereof will be omitted.

FIG. 1 is a block diagram showing the overall general configuration of an optical recognition machine according to an embodiment of the present invention. In FIG. 1, 1 is a first nonvolatile spatial light modulator, and 2 is a first threshold processing circuit. The first non-volatile spatial light modulator stores a synaptic connection matrix of connection elements between the sensory layer and the association layer, and the output therefrom is sent to the first threshold processing circuit 2.
Threshold processing is performed. Further, 3 is a second nonvolatile spatial light modulator, and 4 is a second threshold processing circuit. The second non-volatile spatial light modulator 3 stores a synaptic connection matrix of connection elements between the association layer and the response layer, and the output from this is the second non-volatile spatial light modulator 3.
A threshold processing circuit 4 performs threshold processing. These non-volatile spatial light modulators 1, 3 and threshold processing circuits 2, 4,
Each recognition processing layer of the optical recognition machine is configured.

5 is a first modified signal generation circuit, and 6 is a second modified signal generation circuit. Further, 7 is an input signal, 8 is an output signal, and 9 is a teacher signal. The first correction signal generation circuit 5 generates a correction signal 5a for the synaptic connection matrix stored in the first nonvolatile spatial light modulator 1, and the second correction signal generation circuit 6
generates a correction signal 6a for the synaptic connection matrix stored in the second non-volatile spatial light modulator 3. Further, the second correction signal generation circuit 6 generates an error signal 6b for the correction No. 13 5a, and generates an error signal 6b for the correction No. 13 5a,
The correct signal 6 corresponding to the teacher signal is obtained through signal processing through
c, the signal is supplied to the first correction signal generation circuit 5.

The general operation of the optical learning recognition machine configured as described above is as follows. First, the input signal 7 is converted into a signal via the first non-volatile spatial light modulator 1, subjected to threshold processing by the first threshold processing circuit 2, and then converted again via the second non-volatile spatial light modulator 3. The signal is converted and threshold processing is performed in the second threshold processing circuit 4 to obtain an output signal 8. In this optical learning recognition machine recognition.

The learning process for setting the synaptic connection matrix (matrix of values of connection elements between each layer) stored in each nonvolatile Sky+111 optical modulator is performed by a learning method called back propagation method.

The learning method called back propagation method is performed as follows.

First, a correction signal 6a for the synaptic connection matrix between the association layer and the response layer stored in the second nonvolatile optical spatial modulator 3 is outputted from the second correction signal generation circuit 6.
The correction signal 5a for the synaptic connection matrix between the sensitive ft layer and the association layer stored in the first non-volatile spatial luminous intensity FlID1 is the correction signal 5a between the association layer and the response layer obtained by the second correction signal generation circuit 6. Error signal 6 for the synaptic connection matrix of
After passing through the second non-volatile spatial light modulator 3, the signal b is fed back to the first correction (No. II generation circuit 5) and generated. In this way, the error signal is transferred from the output side of the signal to the input side. is back-propagated to generate modified signals 5a and 6a.The generated modified signal 5a is input to the first non-volatile spatial light modulator 1, and the modified signal 6a is input to the second non-volatile spatial light modulator 3. Input,
Modify each synaptic connection matrix.

In this embodiment, in order to facilitate understanding, an optical learning recognition machine having a so-called two-layer bersebutron structure (a structure in which the association layer is one layer) is described.

It is easy to extend a general perceptron structure consisting of multiple association layers to an optical learning recognition machine, and since it does not require a special explanation, it will be omitted.

FIG. 2 is a diagram illustrating the concept of connection of neuron elements in each layer of an optical learning recognition machine having a perceptron structure according to an embodiment of the present invention. As shown in FIG. 2, the connections between neuron elements in each layer are in this case connections between pixels in each layer. In FIG. 2, 11 is a sensory layer, 12 is an association layer, and 13 is a response layer. 14 is the sensory layer 11 and the association layer 12
and 15 is the synaptic connection matrix between the association layer 12 and the response layer 13. The connection between 0 pixels is the i-th pixel of the sensory layer 11 and the j-th pixel of the association layer 12. , and the case where there is a connection between the pixel of the response layer 13 is shown.

According to perceptron theory, at this time, the input signal is 1 tI, the output signal is 031, the teacher signal is T1°, the synaptic connection matrix of the sensory layer and the association layer is W□1, and the synaptic connection matrix of the association layer and the response layer is W , W, correction signals ΔW', and ΔW for modifying the synaptic connection matrix W,1 and the synaptic connection matrix WIK, respectively.
``JK and 62.62 for generating these modified signals.
is known to be given by the following equation.

S'', = ΣW1□Ii O", = f, (S1□) S211 = ΣW2□08゜03m = fz (S"a) = 62 ・ O'' = (T, -o'g) ・ f'z ( s"w)=61.Ol;(Σδ2.W",J).f', (Sl,) The meaning of the above equation is as follows.

In other words, the correction signal ΔW''llg for the weight of the connection between the n-th neuron element of the response layer and the j-th neuron element of the association layer is the error signal δ□ of the k-th neuron element and the j-th neuron element of the association layer. output o2 from the jth neuron element to the response layer, and a correction signal ΔW11 for the weight of the connection between the jth neuron element of the association layer and the ith neuron element of the sensory layer. ．．
is the sum of the outputs (Σδ2.・) and the second layer output (02,) is the error signal δ1. Input ■1□ for the i-th neuron element in the sensory layer, that is, the signal 01. from the i-th neuron element to the i-th neuron element in the association layer. It is expressed as the product of

FIG. 3 is a diagram illustrating the configuration of a microchannel spatial light modulator used as a nonvolatile spatial light modulator of an optical device that realizes such a neuron element. In FIG. 3, 21 is a half mirror, 122 is a photoelectric cathode surface, 23
is a multi-channel plate, 24 is a mesh electrode, 25
is LiNb0. A crystal plate, 26 an analyzer, and Vc and Vb bias power supplies. Also.

27a is a pattern image written optically, 27b is a pattern image optically read out, and 28 is a readout light.

Such a microchannel spatial light modulator is capable of optically writing a pattern image, and is also capable of optically reading out the written pattern image. In this optical learning recognition RA machine, this microchannel spatial light modulator is used as a nonvolatile spatial light modulator. In this case, the pattern image written and read becomes a synaptic connection matrix.

Next, an outline of the operation of this microchannel spatial light modulator will be explained. The data of the synaptic connection matrix will be stored in the microchannel spatial light modulator as a pattern image. When the optical signal of the pattern image 27a written here is applied as a writing light from the photoelectric cathode surface 22, the photoelectric cathode surface 22 and the multichannel plate 23
．． and the mesh electrode 24, LiNb0. A distribution of charges 25a having a density proportional to the writing light intensity of the added optical signal of the pattern image 27a is generated on the crystal plane of the crystal plate 25, and an electric field is generated within the crystal plate 25. The electric field caused by this charge distribution induces a birefringent portion in the crystal, and the readout light 28 irradiated onto the back surface of the crystal plate 25 undergoes phase modulation according to the intensity of the electric field. Therefore, the light intensity of the readout light after passing through the analyzer 26 is modulated according to the intensity of the optical signal of the written pattern image.

FIG. 4 and FIG. 5 are circuit block diagrams showing the configuration of the correction signal generation circuit, respectively. In FIGS. 4 and 5, 31 is an addition circuit that performs an addition operation, 32 is a multiplication circuit that performs a product operation, and 33 is a subtraction circuit that performs a difference operation.
34 is a nonlinear function calculation circuit that performs nonlinear function calculation. The circuits of these calculation elements are composed of electronic device circuits. The correction signal generation circuit 5 in FIG. 4 outputs a correction signal 5a expressed by the above-mentioned arithmetic expression, and
The correction signal generation circuit 6 in FIG. 5 outputs the correction (g signal 6a and error signal 6b) shown by the above-mentioned arithmetic expression.These correction signal generation circuits 5 and 6 execute the above-mentioned arithmetic expression. As an electronic device for performing the functional operation f'4, the functional operation fJ2, the exclusive OR, and the logical product, the operational circuits are constructed from existing analog or digital electronic circuits using nonlinear input/output of transistors.

FIG. 6 is a block diagram showing the entire optical signal system and electrical signal system in an optical learning recognition machine according to an embodiment of the present invention. In FIG.

7a is an input back-turn optical signal; 7b is a driving electric signal for forming the input back-turn optical signal 7a; 8 is an output signal; 9
is the teacher signal. 10 is a first synaptic connection matrix section;
20 is a first threshold processing section, 30 is a second synaptic connection matrix section, and 40 is a second threshold processing section.

50 is a first modified signal generation section, and 60 is a second modified signal generation section. First synaptic connection matrix section 10. The second synaptic coupling matrix section 30 is composed of a microchannel spatial light modulator. Also, 41.42.43.44°45,
46 is a light emitting element array in which light emitting diodes are arranged in a two-dimensional matrix. Also, 47.48
．． Reference numeral 49 denotes a light receiving element array, in which photodiodes are similarly arranged in a two-dimensional matrix. Also,
21a to 21d are half mirrors, and 26a to 26d are analyzers. In Figure 6.

The optical signal system is shown by a dashed line, and the electrical signal system is shown by a solid line.

Next, the operation will be explained. First, the operation will be explained along the flow of signal processing to obtain the output signal 8 from the input back-turn optical signal 7a. A drive electrical signal 7b for forming an input back-turn optical signal 7a is applied to the light emitting element array 41, and an input back-turn optical signal 7a of the optical signal is formed. This optical signal passes through the half mirror 21a to the first synaptic coupling matrix section 1.
irradiated to o. The first synapse connection matrix section 10 is configured with a microchannel spatial light modulator, and the irradiated optical signal becomes readout light, and the pattern image of the stored synapse connection matrix is read out as reflected light. The optical signal of this reflected light is branched by a half mirror 21a, analyzed by an analyzer 26a, and inputted to a light receiving element array 47.

The light receiving element array 47 converts the pattern image of the received optical signal into an electrical signal, and after threshold processing is performed by the first threshold processing section 20, the output of the threshold processing is converted back into an optical signal by the light emitting element array 42. , by changing the optical path with a half mirror 21b.

The second synaptic connection matrix section 30 is irradiated and applied.

The second synaptic connection matrix section 30 is also constituted by a microchannel spatial light modulator, and the irradiated optical signal becomes readout light, and the pattern image of the stored synaptic connection matrix is read out as reflected light. The reflected light from the second synapse coupling matrix section 30 is analyzed via the half mirror 21b and the analyzer 26b, and then added to the light receiving element array 48. The light receiving element array 48 converts the pattern image of the received optical signal into an electrical signal and applies it to the second threshold processing section 40 . The second threshold processing section 40 subjects the applied signal to threshold processing and outputs it as an output signal 8. Further, this output signal is outputted by the light emitting element array 43 as a pattern image of an optical signal.

Next, the first synaptic connection matrix section 10. The procedure for modifying the synaptic connection matrix stored in the second synaptic connection matrix section 30 will be explained. First, the correction signal for the second synaptic connection matrix section 30 is generated by converting the output signal from the second threshold processing section 40, the teacher signal 9, and the electrical signal from the first threshold processing section 20 into the second correction signal generation. The output signal 60a obtained by inputting to the section 60 is converted into an optical signal pattern by the light emitting element array 45, and the pattern is irradiated and written into the second synaptic connection matrix section 30.

As a result, the pattern of the synaptic connection matrix stored in the second synaptic connection matrix section 30 is corrected. Further, the correction signal for the first synaptic connection matrix section 10 is
The error signal 60b from the second correction signal generation section 60 is converted into an optical signal pattern by the light emitting element array 44, and is irradiated onto the second synaptic coupling matrix section 30 via the half mirrors 21d and 21Q. , the reflected light of the optical signal that has been converted by the memory pattern of the second synaptic connection matrix unit 30 for error (a) is converted into an electrical signal by the light receiving element array 49 via the half mirrors 21c and 21d and the analyzer 26d. , to the modified signal generating section 50. On the other hand, the modified signal generating section 50 receives the signal from the first threshold processing section 20, and generates a modified signal 50a using these (a).

This electrical signal correction signal 50a is converted into a light pattern by the light emitting element array 46, and the first synapse connection matrix section 1
0 and corrects the synaptic connection matrix stored in the first synaptic connection matrix section 10. In addition, the first
Although the initial values of the synaptic connection matrices stored in each of the synaptic connection matrix section 10 and the second synaptic connection matrix section 30 are arbitrary, a weak writing light is initially set, and the intensity is gradually increased to the required intensity through learning. It is modified to output a desired pattern that matches the teacher signal for a specific input pattern.

Next, optical wiring between each element corresponding to each element of the synaptic connection matrix when such a spatial light modulator is used will be explained.

FIG. 7 is a diagram illustrating optical wiring between each element. 7th
In the figure, 51 is an input pattern generation section, 52 is a synaptic connection matrix section, 53 is a threshold processing section, 54 is a synaptic connection matrix section, and 55 is a threshold processing section. Optical wiring is performed by inserting a microlens array between the light emitting diode array and the photodiode array.

Referring to FIG. 7, the synaptic connection matrix section that stores the synaptic connection matrix, the threshold value processing section, etc. are shown as having two-dimensional expansion. In the explanation so far, the target is 1
Since it was a dimensional pattern, the input and output signals were represented by one-dimensional vectors, and the synaptic connection matrix was second-order. However, in order to expand the scope of application, we will expand the input to a two-dimensional image pattern. In this case, the synaptic connection matrix is of fourth order, and can be applied as is by rewriting i in the above formula in the perceptron theory as lJ+J as kl and k as mn. In addition, the number of pixels of the nonvolatile spatial light modulator that stores the synaptic connection matrix in this case is mkX
n1 and kiX1j are required.

In this case, the light emitting diode array of the input pattern generating section 51 and the light emitting diode array S3c of the threshold value processing section are arrays in which kXl and mXn units are arranged, with iXj and kXl light emitting diodes being one unit, respectively. Fetdiode arrays S3a and 55a in the light receiving section of the threshold processing section are arrays in which kXl and mXn photodiodes are arranged, respectively. 55c is m
It is an array of n light emitting diodes. The optical wiring between each array uses a microlens array (not shown) in which microlens arrays are two-dimensionally arranged in a flat plate, and one light-emitting diode or one photodiode is connected to the microlens array (not shown).
Optical wiring between elements can be realized with low loss by arranging two microlenses with their optical axes aligned so that they correspond spatially, and inserting a microlens array between the arrays so that each element surface becomes a focal point. . It is also possible to perform optical wiring using optical fibers instead of microlenses.

Input pattern generation section 51. All of the light emitting diode units of the threshold processing section 53 emit light in the same pattern, and are each driven by an input signal Iij and an output signal Okl. Further, the light emitting diode array of the threshold value processing section 53 is driven by the output signal ○mn of the threshold value processing main body section 55b.

To generate the pattern of correction signals, light emitting diode arrays (44, 45, 46 in FIG. 6) are used, respectively, for kXl,
iXj, kXl is 1 unit 1 ~, mXn, kXl
, mXn units are arranged, and the light receiving element array (49 in FIG. 6) in this case is a photodiode array of kXl pieces.

Next, a description will be given of a method of optically performing matrix calculations regarding the equation ij, which is a two-dimensional extension of the crowd pattern, using the array element described above.

FIG. 8 is a diagram explaining the procedure of the calculation (ΣΣWklij, l1j). In FIG. 8, 61 is a light emitting diode array, 62 is a nonvolatile spatial light modulator that stores a synaptic connection matrix, and 63 is a light receiving element array. iX of the unit in the first row and column of the synaptic connection matrix Wklij
For each of the j elements, a light beam from a light emitting diode is irradiated onto a corresponding position in the pattern of ij, and the output of the entire unit is received by the photodiode in the first row and first column of the light receiving element array. Actually, the light emitting diode array is i
One unit is made up of Xj units, and one unit is arranged with the same light emission pattern, but one unit is optically branched into kX1 units to irradiate 62.

FIG. 9 is a diagram illustrating an example of optically performing the calculation of (ΣΣδ"mnW"nankl) in the back propagation method, which is another matrix calculation. As shown in Figure 9, m
The light beam of the light emitting diode in the row 1st column of all the
It is obtained by receiving light with the photodiode in the column.

The present invention has been specifically explained above based on examples, but
It goes without saying that the present invention is not limited to the embodiments described above, and can be modified in various ways without departing from the spirit thereof.

〔Effect of the invention〕

As explained above, according to the present invention, the advantages of light such as parallelism and high speed can be applied to intelligent information processing such as multiple pattern recognition that requires learning functions that could not be realized with conventional computers. Since it is possible to realize a computer that performs optical processing using this method, we can expect a wide range of applications to various expert systems that are currently only possible with so-called artificial intelligence using large-scale computers.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the overall schematic configuration of an optical learning recognition machine according to an embodiment of the present invention, and FIG. 2 is a block diagram showing an optical learning recognition machine with a perceptron structure according to an embodiment of the present invention. FIG. 3 is a diagram illustrating the concept of connection of neuron elements in each layer as a whole, and FIG. 3 is a diagram illustrating the configuration of a microchannel spatial light modulator used as a nonvolatile spatial light modulator of an optical element that realizes neuron elements. 4 and 5 are circuit block diagrams showing the configuration of the corrected signal generation circuit, respectively, and FIG. 6 shows the entire optical signal system and electrical signal system in an optical learning recognition machine according to an embodiment of the present invention. Block diagram representing the system. FIG. 7 is a diagram for explaining optical wiring between each element, FIGS. 8 and 9 are diagrams for explaining an example of optical calculation, and FIG. 10 is a conceptual diagram of a perceptron. In the figure, 1.3... non-volatile spatial light modulator, 2, 4...
- Threshold processing circuit, 5, 6... Correction signal generation circuit, 7.
... Input signal, 8 ... Output No. 4B, 9 ... Teacher signal, 11 ... Sensory layer, 12 ... Association layer, 13 ...
Response layer, 14.15... Synaptic connection matrix, 21...
・Half mirror 122...Photoelectric cathode surface, 23...
・Multi-channel plate mesh electrode, 25...L
iNbO, crystal plate, 26... analyzer, 27a, 27b
...Pattern image, 28...Reading light, 31...
Addition circuit, 32... Multiplication circuit, 33... Subtraction circuit,
34...Nonlinear function calculation circuit, 10. 30... Synaptic connection matrix section, 20, 40... Threshold processing section, 5
0. 60... Modified signal generation unit, 41. 42, 4
3, 44, 45. 46...Light emitting element array. 47, 48. 49... Light receiving element array, 218~
21d...half mirror, 26a-26d...analyzer, 71...sensory layer, 72...association layer, 73...
response layer, 74. 75...Neuron connections, 76.
··input signal,? ? ...Output signal, 78...Teacher signal. My father-in-law

Claims (1)

[Claims] (1) A sensory layer consisting of a plurality of neuron elements that receive input signals, and a plurality of neuron elements that convert the excitation pattern received from the sensory layer according to the learned transmission amount and connect it to the next layer. In a learning recognition machine with a learning function and a perceptron structure consisting of an association layer and a response layer consisting of neuron elements that identify and judge the excitation patterns received from the association layer and output an answer output signal, the sensory layer, the association layer, and an optical learning recognition machine, characterized in that each layer of the response layer is constituted by a nonvolatile spatial light modulator and a threshold processing circuit. (2) The nonvolatile spatial light modulator is a microchannel spatial light modulator that stores a synaptic connection matrix whose elements are the weights of connections between pattern pixels corresponding to neuron elements, and the threshold processing circuit is connected to the electronic device. The threshold processing device is composed of a light emitting element array that converts an electrical pattern signal into an optical signal, and a light receiving element that converts the optical signal into an electrical signal. 2. The optical learning recognition machine according to claim 1, wherein the optical learning recognition machine performs the optical learning recognition using a signal path of an electrical signal connection and an optical signal connection via an element array. (3) The optical learning recognition machine according to claim 1, further comprising a correction signal generation circuit configured from an electronic device that generates a correction signal, and a correction signal generation circuit configured to generate a correction signal based on an input signal to be recognized. The correct answer is given from the outside as a teacher signal, a correction signal is generated from the input signal and the difference between the output signal output from the threshold processing circuit and the teacher signal, and the correction signal is fed back from the output side to the input side in the opposite direction to respond. The weights of the connections between the neuron elements of the layer and the association layer, the weights of the connections between the neuron elements of different association layers, and the weights of the connections between the neuron elements of the association layer and the sensory layer are varied based on the modification signal. 1. A learning method for an optical learning recognition machine, characterized in that the output signal of the optical learning recognition machine is brought closer to a teacher signal by repeatedly performing back propagation learning processing for correction. (4) In the optical learning recognition machine according to claim 2, each light emitting element of the first light emitting element array is caused to emit light as a point light source in correspondence to each pixel of the input signal pattern, and each light emitting element is caused to emit light as a point light source. directing each light beam from the element to a corresponding pixel position on a readout surface of a first non-volatile spatial light modulator storing a synaptic connection matrix between neuronal elements of the sensory layer and the association layer; After the reflected light passes through an analyzer, it is photoelectrically converted by the first light receiving element array and inputted to the first threshold processing circuit, and the electrical output signal of the first threshold processing circuit is transmitted to the second light emitting element array. Applied as a drive current, the second
The light emitting elements of the light emitting element array are made to emit light as a point light source, and each light beam from each light emitting element is converted into a second nonvolatile spatial light modulator that stores the synaptic connection matrix between the neuron elements of the association layer and the response layer. The reflected light is irradiated onto the corresponding pixel position on the reading surface of the device, and after passing through an analyzer, the reflected light is photoelectrically converted by the second light receiving element array and inputted to the second threshold processing circuit. The electrical output signal obtained from the threshold processing circuit is further input to a third light emitting element array, and each light emitting element of the third light emitting element array emits light as a point light source to obtain an optical signal of an output pattern. A pattern recognition method for optical learning recognition machines. (5) In the optical learning recognition machine according to claim 2, a learning and storing method for storing connection weights between pattern pixels corresponding to neuron elements in a non-volatile spatial light modulator, A drive current proportional to the value representing the weight of the element of the synaptic connection matrix is generated in the light emitting element that illuminates the corresponding pixel of the light emitting element array, and the light beam of the light emitting element is emitted with a light intensity proportional to the drive current. The corresponding pixel portion on the writing surface of the nonvolatile spatial light modulator was irradiated to induce charges that cause birefringence on the readout surface, and the readout surface was irradiated with linearly polarized laser light of the readout light. When,
An optical learning and memory method for an optical learning recognition machine, characterized in that the plane of polarization of the reflected light is rotated by birefringence, and when it passes through an analyzer, the readout light is intensity-modulated in proportion to the weight of the coupling. . (6) In the optical learning recognition machine learning method according to claim 3, the correction signal generation circuit that generates a correction signal for correcting the weight of the connection between the neuron elements of the association layer; The output signal and the external input teacher signal are input to generate a correction signal for the electric signal, and a drive current proportional to the correction signal of the pixel corresponding to each light emitting element of the light emitting element array is applied to change the drive current. irradiating a light beam of a light emitting element emitted with a proportional light intensity to each pixel position on the writing surface of the nonvolatile spatial light modulator whose coupling weight is to be corrected;
When the nonvolatile spatial light modulator is irradiated with readout light, the plane of polarization of the reflected light is rotated due to birefringence, and when it passes through an analyzer, the readout light is intensity-modulated in proportion to the coupling weight. A learning method for an optical learning recognition machine, characterized by: (7) In the optical learning recognition machine learning method according to claim 3, the first step generates a correction signal for correcting the weight of the connection between the neuron elements of the sensory layer and the association layer. A second modification in which the sensory layer signal output from the threshold processing circuit of the sensory layer is given as an input to the modification signal generation circuit, and a modification signal is generated for modifying the weight of the connection between the neuron elements of the association layer. The modified signal fed back from the signal generation circuit is converted into a light pattern by a light emitting element array, and is illuminated on the readout surface of a nonvolatile spatial light modulator that records the weights of connections between neurons in the association layer and the response layer. An electric signal obtained by photoelectrically converting the reflected light from the light receiving element array is given as input, a correction signal is generated from the difference, and a drive current proportional to the intensity of the correction signal is applied to the light emitting element that illuminates each corresponding pixel of the light emitting element array. A light beam of the light emitting element is emitted with a light intensity proportional to the drive current by applying a current to the position of each pixel on the writing surface of the nonvolatile spatial light modulator whose coupling weight is to be corrected. When a nonvolatile spatial light modulator is irradiated with readout light, the plane of polarization of the reflected light rotates due to birefringence, and when it passes through an analyzer, the readout light is intensity-modulated in proportion to the coupling weight. A learning method for an optical learning recognition machine.