Classification device and classification method

Abstract

This classification device is configured to comprise: an optical operation unit (1) that, when M (M is an integer of 2 or more) feature amount signals having mutually different feature amount types are given as feature amount signals indicating feature amounts of classification targets, performs weighting matrix operation on the M feature amount signals and outputs N (N is an integer less than the number of classifications of the classification targets) weighting matrix operation results; and an activation function operation unit (2) that has mutually different K (K is the same number as the number of classifications of the classification targets) activation functions as functions taking the N weighting matrix operation results outputted from the optical operation unit (1) as input variables, performs operations using the respective activation functions on the basis of the N weighting matrix operation results, and outputs operation results of the K activation functions. The classification device also comprises a classification unit (4) that performs classification processing of the classification targets on the basis of the operation results of the K activation functions outputted from the activation function operation unit (2), and outputs classification results of the classification targets.

Description

Classification device and classification method

[0001] This disclosure relates to a classification device and a classification method.

[0002] There are classification devices that perform weighting matrix operations on feature signals representing the features to be classified, and then perform classification processing on the target based on the results of the weighting matrix operations. As an example of such a classification device, Non-Patent Literature 1 discloses a classification device comprising: an optical processing unit that, given M (M is an integer of 2 or more) feature signals of different types, performs weighting matrix operations on the M feature signals and outputs weighting matrix operation results for the number of classifications of the target; and a classification unit that performs classification processing on the target based on the weighting matrix operation results for the number of classifications output from the optical processing unit. The optical processing unit disclosed in Non-Patent Literature 1 is implemented using optical circuits to speed up the weighting matrix operations. When the classification unit is implemented using digital circuits, each of the weighting matrix operation results for the number of classifications output from the optical processing unit is output to the classification unit via a photoelectric conversion circuit and an analog-to-digital conversion circuit.

[0003] “Generalized robust training scheme using genetic algorithm for optical neural networks with imprecise components”, Vol. 10, No. 8 / August 2022 / Photonics Research

[0004] In the classification device disclosed in Non-Patent Document 1, there was a problem in that, in order for the optical processing unit to output the weighting matrix calculation results for the number of classifications to the classification unit, the classification device needed to be equipped with a number of photoelectric conversion circuits and a number of analog-to-digital conversion circuits for the number of classifications.

[0005] This disclosure was made to solve the above-mentioned problems, and aims to provide a classification device that can classify objects using fewer weighting matrix calculation results than the number of classifications.

[0006] A classification device according to the present disclosure is configured to include: an optical operation unit that, when M (M is an integer of 2 or more) feature quantity signals of mutually different feature quantity types are received as feature quantity signals indicating the feature quantity of a classification target, performs weighted matrix operation on the M feature quantity signals and outputs N (N is an integer less than the number of classifications of the classification target) weighted matrix operation results; and an activation function operation unit that has K (K is the same number as the number of classifications of the classification target) mutually different activation functions as functions using the N weighted matrix operation results output from the optical operation unit as input variables, operates each activation function based on the N weighted matrix operation results, and outputs the operation results of the K activation functions. The classification device further includes a classification unit that performs classification processing on the classification target based on the operation results of the K activation functions output from the activation function operation unit, and outputs the classification result of the classification target.

[0007] According to the present disclosure, the classification target can be classified using a number of weighted matrix operation results that is smaller than the number of classifications.

[0008] FIG. 2 is a configuration diagram showing a classification device according to Embodiment 1. FIG. 3 is a hardware configuration diagram showing hardware of the classification device according to Embodiment 1. FIG. 4 is a hardware configuration diagram of a computer when the classification device is implemented by software, firmware, or the like. FIG. 5 is an explanatory diagram showing an example of a neural network. FIG. 6 is an explanatory diagram showing a configuration example of an optical operation unit 1. FIG. 7 is an explanatory diagram showing an example of an operation layer that performs weighted matrix operation on M feature quantity signals. FIG. 8 is a flowchart showing a processing procedure of the classification device during learning. Operation result o of the optical operation unit 1 １ , o ２ For the two-dimensional space related to, two-dimensional activation function f ０ (o １ , o ２ ) is an explanatory diagram showing a distribution example of indices of classification label (0) output from. Operation result o of the optical operation unit 1 １ , o ２ For the two-dimensional space related to, two-dimensional activation function f １ (o １ , o ２ ) is an explanatory diagram showing a distribution example of indices of classification label (2) output from. Operation result o of the optical operation unit 1１ ,o ２ A two-dimensional activation function f for a two-dimensional space related to this ２ (o １ ,o ２ This is an explanatory diagram showing an example of the distribution of the indicators for the classification label (2) output from ). It is a two-dimensional activation function f showing a Gaussian distribution with respect to the argument angle. ０ (o １ ,o ２ ), f １ (o １ ,o ２ ), f ２ (o １ ,o ２ This is an explanatory diagram of the classification method, which is the processing procedure of the classification device during inference. It shows the two-dimensional activation function f, which exhibits a linear distribution with respect to the argument. ０ (o １ ,o ２ ), f １ (o １ ,o ２ ), f ２ (o １ ,o ２ This is an explanatory diagram of the one-dimensional activation function f. ｉ (o １ This is an explanatory diagram showing an example where the distribution function is maximized at the center of a number line divided into four sections. 3D activation function f ｉｊｌ (o １ ,o ２ ,o ３ This is an explanatory diagram showing an example where the function exhibits a Gaussian distribution with maximum values ​​at the center of each grid cell for a grid of 4 × 4 × 4 = 64 classification labels, obtained by dividing each side of a cube into four sections. This is a configuration diagram showing a classification device according to Embodiment 2. This is a hardware configuration diagram showing the hardware of the classification device according to Embodiment 2. This is an explanatory diagram showing an example where a two-dimensional space is divided into multiple regions based on the argument angle. This is an explanatory diagram showing an example where a two-dimensional space is divided into grid-like regions.

[0009] To provide a more detailed explanation of this disclosure, the forms for implementing this disclosure will be described below with reference to the attached drawings.

[0010] Embodiment 1. Figure 1 is a configuration diagram showing a classification device according to Embodiment 1. Figure 2 is a hardware configuration diagram showing the hardware of the classification device according to Embodiment 1. The classification device shown in Figure 1 comprises an optical processing unit 1, an activation function processing unit 2, a learning unit 3, and a classification unit 4.

[0011] The optical computing unit 1 is implemented, for example, by the optical computing circuit 21 shown in Figure 2. The optical computing circuit 21 is integrated, for example, by silicon photonics technology. The optical computing unit 1 has a computing layer that performs weighting matrix operations on M feature signals. Specifically, the optical computing unit 1 has computing layers U and V that perform unitary matrix operations and computing layer Σ that performs diagonal matrix operations as computing layers that perform weighting matrix operations on M feature signals.

[0012] The calculation layers U and V, which perform unitary matrix operations, include Mach-Zehnder interferometers. The arrangement of the Mach-Zehnder interferometers included in calculation layers U and V is, for example, either a Clements type or a Reck type. However, the arrangement of the Mach-Zehnder interferometers is not limited to the Clements type or the Reck type. In the classification device shown in Figure 1, the optical calculation unit 1 has two calculation layers U and V that perform unitary matrix operations and one calculation layer Σ that performs diagonal matrix operations, resulting in a total of three calculation layers. However, this is merely an example, and the number of calculation layers in the optical calculation unit 1 may be one or two, or four or more. For example, the optical calculation unit 1 may have only calculation layers U and V that perform unitary matrix operations, or it may have only calculation layer V. The optical computing unit 1 includes, in addition to the computing layer, photoelectric conversion circuits (hereinafter referred to as "PD") 1a-n and analog-to-digital conversion circuits (hereinafter referred to as "ADC") 1b-n. n = 1, ..., N.

[0013] The optical calculation unit 1 is supplied with M (M is an integer of 2 or more) feature signals, each with different types of features, from an external source, for example. These feature signals are optical signals in which the features of the target to be classified are reflected in either the intensity or amplitude of the light. If the target to be classified is, for example, an iris, the types of features could include, for example, the length of the sepals, the width of the sepals, the length of the petals, and the width of the petals. The optical calculation unit 1 performs a weighting matrix operation on the M feature signals and outputs N (N is an integer less than the number of classifications of the target to be classified) weighting matrix operation results to the activation function calculation unit 2. The weighting matrix operation on the M feature signals is an N × M matrix operation. Each of the N calculation results from the N × M matrix operation performed by the optical calculation unit 1 is output to the activation function calculation unit 2 via PD1a-n (n=1, ..., N) and ADC1b-n. PD1a-n converts the optical signal representing the nth calculation result out of M calculation results into an electrical signal and outputs the electrical signal to ADC1b-n. ADC1b-n converts the electrical signal output from PD1a-n into digital data and outputs the digital data to the activation function calculation unit 2.

[0014] The activation function calculation unit 2 is implemented, for example, by the activation function calculation circuit 22 shown in Figure 2. The activation function calculation unit 2 has K (where K is the same number as the number of classification targets) distinct activation functions as functions that take N weighting matrix calculation results output from the optical calculation unit 1 as input variables. The activation function calculation unit 2 obtains N weighting matrix calculation results from the optical calculation unit 1. The activation function calculation unit 2 calculates each activation function based on the N weighting matrix calculation results. During the learning phase of the optical calculation unit 1, the activation function calculation unit 2 outputs the calculation results of the N activation functions to the learning unit 3 and the classification unit 4, respectively. During the inference phase for classifying the classification targets, the activation function calculation unit 2 outputs the calculation results of each activation function to the classification unit 4.

[0015] The learning unit 3 is implemented, for example, by the learning circuit 23 shown in Figure 2. The learning unit 3 includes a loss function calculation unit 3a, a weight coefficient update unit 3b, and a digital-to-analog conversion unit (hereinafter referred to as "DAC") 3c. The learning unit 3 obtains the true classification result of the object to be classified, for example, from an external source. The true classification result is the training data. The learning unit 3 obtains the calculation results of K activation functions from the activation function calculation unit 2. Using the true classification result of the object to be classified, the learning unit 3 calculates a loss function for the calculation results of the K activation functions, and updates the weight coefficients for M feature signals based on the calculation result of the loss function.

[0016] The loss function calculation unit 3a obtains the true classification result of the object to be classified, for example, from an external source. The loss function calculation unit 3a obtains the calculation results of K activation functions from the activation function calculation unit 2. Using the true classification result of the object to be classified, the loss function calculation unit 3a calculates a loss function for the calculation results of the K activation functions and outputs the calculation result of the loss function to the weight coefficient update unit 3b. The weight coefficient update unit 3b obtains the calculation result of the loss function from the loss function calculation unit 3a. Based on the calculation result of the loss function, the weight coefficient update unit 3b calculates weight coefficients for M feature signals and outputs the weight coefficients to the DAC 3c. The DAC 3c converts the weight coefficients output from the weight coefficient update unit 3b into a weight adjustment voltage and outputs the weight adjustment voltage to the optical calculation unit 1. As a result, the weight coefficients for the M feature signals are updated.

[0017] The classification unit 4 is implemented, for example, by the classification circuit 24 shown in Figure 2. The classification unit 4 obtains the calculation results of K activation functions from the activation function calculation unit 2. The classification unit 4 compares the K calculation results and performs the classification process for the items to be classified based on the comparison result of the K calculation results. The classification unit 4 outputs the classification result of the items to be classified to, for example, a display device. This allows the user to confirm the classification result of the items to be classified by looking at the display content of the display device.

[0018] In Figure 1, the optical processing unit 1, activation function processing unit 2, learning unit 3, and classification unit 4, which are components of the classification device, are assumed to be implemented by dedicated hardware as shown in Figure 2. That is, the classification device is assumed to be implemented by an optical processing circuit 21, an activation function processing circuit 22, a learning circuit 23, and a classification circuit 24. The optical processing circuit 21, activation function processing circuit 22, learning circuit 23, and classification circuit 24 can be, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), or a combination thereof.

[0019] The components of a classification device are not limited to those implemented by dedicated hardware; the classification device may also be implemented by software, firmware, or a combination of software and firmware. The software or firmware is stored as a program in the computer's memory. A computer refers to the hardware that executes the program, and includes, for example, a CPU (Central Processing Unit), GPU (Graphics Processing Unit), central processing unit, processing unit, arithmetic unit, microprocessor, microcomputer, processor, or DSP (Digital Signal Processor).

[0020] Figure 3 is a hardware configuration diagram of a computer when the classification device is implemented using software or firmware. When the classification device is implemented using software or firmware, a program that causes the computer to execute the respective processing procedures in the optical processing unit 1, the activation function calculation unit 2, the learning unit 3, and the classification unit 4 is stored in memory 41. The computer's processor 42 then executes the program stored in memory 41.

[0021] Furthermore, Figure 2 shows an example where each component of the classification device is implemented by dedicated hardware, and Figure 3 shows an example where the classification device is implemented by software or firmware, etc. However, this is only one example, and some components of the classification device may be implemented by dedicated hardware, while the remaining components may be implemented by software or firmware, etc.

[0022] The optical processing unit 1 performs the weighting matrix operations performed by the neural network shown in Figure 4. Figure 4 is an explanatory diagram showing an example of a neural network. Figure 5 is an explanatory diagram showing an example configuration of the optical processing unit 1. The optical processing unit 1 shown in Figure 5 realizes the operation of the neural network shown in Figure 4. In the neural network shown in Figure 4, the calculation content of the weighting matrix operations for the M feature signals is updated by updating the weight coefficients at each node. In the optical processing unit 1 shown in Figure 5, the calculation content of the weighting matrix operations for the M feature signals is updated by updating the weight coefficients of the interference portion and the weight coefficients of the phase shift portion.

[0023] The optical processing unit 1, as shown in Figure 6, for example, has a first processing layer, a second processing layer, and a third processing layer as processing layers that perform weighting matrix operations on M feature signals. Figure 6 is an explanatory diagram showing an example of processing layers that perform weighting matrix operations on M feature signals. Figure 6 shows an example where M = 4 and N = 2. The first processing layer is processing layer U that performs unitary matrix operations, the second processing layer is processing layer Σ that performs diagonal matrix operations, and the third processing layer is processing layer V that performs unitary matrix operations. Processing layers U and V that perform unitary matrix operations include multiple Mach-Zehnder interferometers. In Figure 6, the Mach-Zehnder interferometers are denoted as "MZI".

[0024] In the example in Figure 6, the arrangement of multiple MZIs included in the computational layers U and V is a mesh configuration called the Clements type. Each MZI has the function of shifting the phase of the feature signal. The arrangement of multiple MZIs is not limited to the Clements type; for example, it could be a Reck type. In the example in Figure 6, the multiple attenuators (hereinafter referred to as "ATTs") included in the computational layer Σ have the function of attenuating the optical amplitude of the feature signal. The computational layer Σ may also include amplifiers that have the function of amplifying the optical amplitude of the feature signal instead of ATTs.

[0025] Next, we will explain the operation of the classification device shown in Figure 1. For the sake of explanation, we will describe an example where M=4, N=2, and K=3. First, we will explain the operation of the classification device when the optical processing unit 1 is learning. Figure 7 is a flowchart showing the processing procedure of the classification device during learning. The optical processing unit 1 receives, for example, M=4 feature signals a, which are of different types from each other, as feature signals indicating the features of the object to be classified, from an external source. ０ , a １ , a ２ , a ３ The following is given: Feature signal a ０ , a １ , a ２ , a ３ This is an optical signal in which the feature quantity to be classified is reflected in either the intensity or amplitude of the light. Feature quantity signal a during training ０ , a １ , a ２ , a ３ This is the training data. The optical processing unit 1 calculates M = 4 feature signals a as shown in the following equation (1). ０ , a １ , a ２ , a ３ As a weighted matrix operation on this, an N×M matrix operation is performed (step ST1 in Figure 7).

[0026]

[0027] PD1a-n of the optical computing unit 1 are the calculation results of M = 4 of the weighted matrix calculation o ０ ,o １ ,o２ ,o ３ The optical signal representing the nth calculation result is converted into an electrical signal, and the electrical signal is output to ADC1b-n. In the example in Figure 6, PD1a-1 represents the calculation result o １ The optical signal indicating the result is converted into an electrical signal, and the electrical signal is output to ADC1b-1, and PD1a-2 calculates the result o ２ The optical signal indicating the result is converted into an electrical signal, and the electrical signal is output to ADC1b-1. ADC1b-n converts the electrical signal output from PD1a-n into digital data, and outputs the digital data to the activation function calculation unit 2. In the example in Figure 6, ADC1b-1 calculates the result o １ The electrical signal indicating this is converted into digital data, and the ADC1b-2 calculates the result o ２ The electrical signals shown are converted into digital data. In the example in Figure 6, the result of the weighted matrix operation M = 4 is o ０ ,o １ ,o ２ ,o ３ Of these, two calculation results o １ ,o ２ It is converted into an electrical signal and then into digital data. However, this is just one example, and the calculation result o １ ,o ２ The other two calculation results may be converted into electrical signals and then into digital data. Furthermore, the number of calculation results to be converted does not need to be more than the number classified by the classification unit 4, and is not limited to two.

[0028] The activation function calculation unit 2 receives the calculation result o from the ADC1b-1 of the optical calculation unit 1. １ Digital data showing this is obtained, and the calculation result o is obtained from ADC1b-2. ２ Digital data showing this is obtained. The activation function calculation unit 2 has K distinct activation functions as functions that take the N=2 weighted matrix calculation results output from the optical calculation unit 1 as input variables. When N=2, each of the K activation functions is a two-dimensional activation function. When K=3, the activation function calculation unit 2 has three two-dimensional activation functions f ０ (o １ ,o ２ ), f１ (o １ , o ２ ), f ２ (o １ , o ２ ). The two-dimensional activation function f ０ (o １ , o ２ ) is a function that outputs an index of the classification label (0) as an operation result, as shown in the following formula (2). The two-dimensional activation function f １ (o １ , o ２ ) is a function that outputs an index of the classification label (1) as an operation result, as shown in the following formula (3). The two-dimensional activation function f ２ (o １ , o ２ ) is a function that outputs an index of the classification label (2) as an operation result, as shown in the following formula (4).

[0029] In formulas (2) to (4), σ ２ indicates the variance of the Gaussian distribution. Note that σ ２ may be set to any arbitrary value in each two-dimensional activation function.

[0030] FIG. 8 is an explanatory diagram showing a distribution example of the index of the classification label (0) output from the two-dimensional activation function f for the two-dimensional space related to the operation results o of the optical operation unit 1 １ , o ２ . ０ (o １ , o ２ ). FIG. 9 is an explanatory diagram showing a distribution example of the index of the classification label (1) output from the two-dimensional activation function f for the two-dimensional space related to the operation results o of the optical operation unit 1 １ , o ２ . １ (o １ , o ２ ). FIG. 10 is an explanatory diagram showing a distribution example of the index of the classification label (2) output from the two-dimensional activation function f for the two-dimensional space related to the operation results o of the optical operation unit 1 １ , o ２ . ２ (o １ , o ２This is an explanatory diagram showing an example of the distribution of indicators for the classification label (2) output from ). In Figures 8 to 10, the horizontal axis represents the calculation result o １ The vertical axis represents the calculation result o ２ The shades of gray represent the magnitude of the index for the classification label (indicated as "index value" in the figure).

[0031] Calculation result of optical calculation unit 1 o １ and the calculation result of the optical calculation unit 1 o ２ The ratio (o ２ / o １ If ) is the argument, then the argument at which the index of classification label (0) is maximized is 0 degrees, and the argument at which the index of classification label (1) is maximized is 45 degrees. Also, the argument at which the index of classification label (2) is maximized is 90 degrees. Two-dimensional activation function f ０ (o １ ,o ２ ), f １ (o １ ,o ２ ), f ２ (o １ ,o ２ ) is a Gaussian distribution function that shows a Gaussian distribution with respect to the argument angle, as shown in Figure 11. Figure 11 shows the two-dimensional activation function f that shows a Gaussian distribution with respect to the argument angle. ０ (o １ ,o ２ ), f １ (o １ ,o ２ ), f ２ (o １ ,o ２ This is an explanatory diagram of the system. In Figure 11, the horizontal axis represents the argument angle, and the vertical axis represents the index of the classification label. In the classification device shown in Figure 1, the activation function calculation unit 2 calculates a two-dimensional activation function f such that the argument angles at which the index of each classification label is maximized are equally spaced at 45 degrees. ０ (o １ ,o ２ ), f １ (o １ ,o ２ ), f ２ (o １ ,o ２) has. However, this is just one example, and the activation function calculation unit 2 has a two-dimensional activation function f such that the argument angles at which the index of each classification label is maximized are unequally spaced. ０ (o １ ,o ２ ), f １ (o １ ,o ２ ), f ２ (o １ ,o ２ ) may also be present. Furthermore, the classification device shown in Figure 1 has equally spaced angles of 45 degrees. However, this is merely an example, and the equally spaced angles may be angles other than 45 degrees.

[0032] The activation function calculation unit 2 calculates the result o １ Digital data and calculation results showing the results ２ Using digital data that shows this, K = 3 2D activation functions f ０ (o １ ,o ２ ), f １ (o １ ,o ２ ), f ２ (o １ ,o ２ The activation function calculation unit 2 calculates each of the following (step ST2 in Figure 7). The activation function calculation unit 2 calculates K = 3 two-dimensional activation functions f ０ (o １ ,o ２ ), f １ (o １ ,o ２ ), f ２ (o １ ,o ２ As a result of the calculation, the indicators for classification labels (0), (1), and (2) are output to the learning unit 3 and the classification unit 4, respectively.

[0033] The loss function calculation unit 3a of the learning unit 3 obtains, for example, an external source, true label c that indicates the true classification result of the object to be classified. The loss function calculation unit 3a obtains K = 3 activation functions f from the activation function calculation unit 2. ０ (o １ ,o ２ ), f １ (o １ ,o ２ ), f ２(o １ ,o ２ As a result of the calculation, the index of classification labels (0), (1), and (2) is obtained. The loss function calculation unit 3a uses the true label c to calculate K = 3 activation functions f ０ (o １ ,o ２ ), f １ (o １ ,o ２ ), f ２ (o １ ,o ２ The loss function H(p,q) is calculated for the result of the calculation (step ST3 in Figure 7).

[0034] In equation (5), if the true label c is "0", then p(0) = 1, p(1) = 0, and p(2) = 0. If the true label c is "1", then p(0) = 0, p(1) = 1, and p(2) = 0. If the true label c is "2", then p(0) = 0, p(1) = 0, and p(2) = 1.

[0035] Therefore, if the true label c is "0", the result of the loss function H(p,q) is log(f ０ o １ ,o ２ )) and if the true label c is "1", the result of the loss function H(p,q) is log(f １ (o １ ,o ２ )) and if the true label c is "2", the result of the loss function H(p,q) is log(f ２ (o １ ,o ２ The loss function calculation unit 3a outputs the calculation result of the loss function H(p,q) to the weight coefficient update unit 3b.

[0036] The weight coefficient update unit 3b obtains the calculation result of the loss function H(p,q) from the loss function calculation unit 3a. Based on the calculation result of the loss function H(p,q), the weight coefficient update unit 3b calculates weight coefficients w for the M feature signals (step ST4 in Figure 7). Specifically, the weight coefficient update unit 3b calculates the weight coefficients w that result in the smallest calculation result of the loss function H(p,q) by executing an optimization algorithm. Examples of optimization algorithms include gradient descent or simulated annealing. The weight coefficient update unit 3b outputs the weight coefficients w to the DAC 3c.

[0037] The DAC3c obtains the weight coefficient w from the weight coefficient update unit 3b. The DAC3c converts the weight coefficient w into a weight adjustment voltage and outputs the weight adjustment voltage to the optical calculation unit 1 (step ST5 in Figure 7). The weight adjustment voltage controls the phase shift by the MZI and the attenuation of the optical amplitude by the ATT of the optical calculation unit 1, thereby updating the weight coefficients for the M feature signals.

[0038] The classification unit 4 obtains, for example, an external source, a true label c that indicates the true classification result of the object to be classified. The classification unit 4 obtains K = 3 activation functions f from the activation function calculation unit 2. ０ (o １ ,o ２ ), f １ (o １ ,o ２ ), f ２ (o １ ,o ２ As a result of the calculation, the indices for classification labels (0), (1), and (2) are obtained. The classification unit 4 compares the indices for classification labels (0), (1), and (2). The classification unit 4 classifies the object to be classified as classification label (0) if the largest indice among the indices for classification labels (0), (1), and (2) is for classification label (0), classifies the object to be classified as classification label (1) if the largest indice is for classification label (1), and classifies the object to be classified as classification label (2) if the largest indice is for classification label (2).

[0039] The classification unit 4 compares the classification result of the classification label with the true label c. If the two match, it determines that the answer is correct; if they do not match, it determines that the answer is incorrect (step ST6 in Figure 7). The classification unit 4 stores the results of whether the answer is correct or incorrect. If the number of stored judgment results is less than a preset number G (G is an integer greater than or equal to 1) (step ST7 in Figure 7: if YES), the classification unit 4 repeats the process of steps ST1 to ST6. If the number of stored judgment results reaches the preset number G (step ST7 in Figure 7: if NO), the classification unit 4 calculates the correct answer rate CAR of the G judgment results as shown in the following formula (6) (step ST8 in Figure 7).

[0040] In equation (6), P is the number of correct answer results.

[0041] If the accuracy rate CAR of the classification result is less than the threshold (in step ST9: YES in Figure 7), the classification unit 4 outputs a command to continue learning to the optical processing unit 1. This causes the processing of steps ST1 to ST6 to be repeated. The threshold may be stored in the internal memory of the classification device or may be provided from outside the classification device. If the accuracy rate CAR of the classification result is greater than or equal to the threshold (in step ST9: NO in Figure 7), the learning of the optical processing unit 1 is terminated.

[0042] Next, we will explain the operation of the classification device when classification inference is performed. Figure 12 is a flowchart showing the classification method, which is the processing procedure of the classification device during inference. The optical processing unit 1 receives, for example, M = 4 feature signals a from an external source, which are of different types to each other, as feature signals indicating the features of the object to be classified. ０ , a １ , a ２ , a ３ The following is given: Feature signal a ０ , a １ , a ２ , a ３ This is an optical signal in which the feature quantity to be classified is reflected in either the intensity or amplitude of the light. Feature quantity signal a during inference ０ , a １ , a ２ , a ３This is not training data, but a signal representing feature quantities calculated by some device. The optical calculation unit 1, as shown in equation (1), has M = 4 feature quantity signals a ０ , a １ , a ２ , a ３ As a weighted matrix operation on this, an N×M matrix operation is performed (step ST11 in Figure 12).

[0043] PD1a-n of the optical computing unit 1 are the calculation results of M = 4 of the weighted matrix calculation o ０ ,o １ ,o ２ ,o ３ The optical signal representing the nth calculation result is converted into an electrical signal, and the electrical signal is output to ADC1b-n. In the example in Figure 6, PD1a-1 represents the calculation result o １ The optical signal indicating the result is converted into an electrical signal, and the electrical signal is output to ADC1b-1, and PD1a-2 calculates the result o ２ The optical signal indicating the result is converted into an electrical signal, and the electrical signal is output to ADC1b-1. ADC1b-n converts the electrical signal output from PD1a-n into digital data, and outputs the digital data to the activation function calculation unit 2. In the example in Figure 6, ADC1b-1 calculates the result o １ The electrical signal indicating this is converted into digital data, and the ADC1b-2 calculates the result o ２ It converts electrical signals that indicate something into digital data.

[0044] The activation function calculation unit 2 receives the calculation result o from the ADC1b-1 of the optical calculation unit 1. １ Digital data showing this is obtained, and the calculation result o is obtained from ADC1b-2. ２ Digital data showing the result is obtained. The activation function calculation unit 2 calculates the calculation result o １ Digital data and calculation results showing the results ２ Using digital data that shows this, K = 3 2D activation functions f ０ (o １ ,o ２ ), f １ (o １ ,o ２ ), f ２ (o １ ,o ２The activation function calculation unit 2 calculates each of the following (step ST12 in Figure 12). The activation function calculation unit 2 calculates K = 3 two-dimensional activation functions f ０ (o １ ,o ２ ), f １ (o １ ,o ２ ), f ２ (o １ ,o ２ As a result of the calculation, the classification labels (0), (1), and (2) are output to the classification unit 4.

[0045] The classification unit 4 receives K = 3 activation functions f from the activation function calculation unit 2. ０ (o １ ,o ２ ), f １ (o １ ,o ２ ), f ２ (o １ ,o ２ As a result of the calculation, the indices for classification labels (0), (1), and (2) are obtained. The classification unit 4 compares the indices for classification labels (0), (1), and (2). If the largest index among the indices for classification labels (0), (1), and (2) is the index for classification label (0) (in step ST13 of Figure 12: YES), the classification unit 4 classifies the object to be classified as classification label (0) (in step ST14 of Figure 12). If the largest index is the index for classification label (1) (in step ST13 of Figure 12: NO, step ST15: YES), the classification unit 4 classifies the object to be classified as classification label (1) (in step ST16 of Figure 12). If the largest index is the index for classification label (2) (in step ST13 of Figure 12: NO, step ST15: NO), the classification unit 4 classifies the object to be classified as classification label (2) (in step ST17 of Figure 12). The classification unit 4 outputs the classification result of the object to be classified, for example, to a display device. This allows users to check the classification results of the items to be classified by looking at the display content of the display device.

[0046] In the above embodiment 1, the classification device is configured to include an optical calculation unit 1 that, when given M (M is an integer of 2 or more) feature signals of different types representing the features of the object to be classified, performs weighting matrix calculations on the M feature signals and outputs N (N is an integer less than the number of classifications of the object to be classified) weighting matrix calculation results, and an activation function calculation unit 2 that has K (K is the same number as the number of classifications of the object to be classified) different activation functions as input variables, which take the N weighting matrix calculation results output from the optical calculation unit 1 as input variables, calculates each activation function based on the N weighting matrix calculation results, and outputs the calculation results of the K activation functions. Furthermore, the classification device includes a classification unit 4 that performs classification processing on the object to be classified based on the calculation results of the K activation functions output from the activation function calculation unit 2 and outputs the classification result of the object to be classified.Therefore, the classification device can classify the object to be classified using a number of weighting matrix calculation results less than the number of classifications.

[0047] The classification device shown in Figure 1 performs weighting matrix calculations and other operations to perform classification processing on the objects to be classified. In the classification device shown in Figure 1, the optical calculation unit 1 that performs weighting matrix calculations is realized by an optical circuit. That is, the optical calculation unit 1 is realized by an optical calculation circuit 21 integrated using silicon photonics technology, for example. Because the optical calculation unit 1 is realized by an optical circuit, the calculation speed of the weighting matrix calculation by the optical calculation unit 1 is faster than that of an optical calculation unit realized by an electrical circuit. In the classification device disclosed in Non-Patent Literature 1, in order to increase the calculation speed of weighting matrix calculations, an optical calculation unit realized by an optical circuit performs weighting matrix calculations. However, in the classification device disclosed in Non-Patent Literature 1, in order for the optical calculation unit to output the weighting matrix calculation results for the number of classifications to the classification unit, it is necessary to have a photoelectric conversion circuit for the number of classifications and an analog-to-digital conversion circuit for the number of classifications. For this reason, the classification device disclosed in Non-Patent Literature 1 has a large circuit size and requires a lot of power to perform classification processing on the objects to be classified. As is clear from the above description, the classification device shown in Figure 1 can classify objects using fewer weighting matrix calculation results than the number of classifications. In other words, the classification device shown in Figure 1 can classify objects if it is equipped with fewer photoelectric conversion circuits and fewer analog-to-digital conversion circuits than the number of classifications. Therefore, the classification device shown in Figure 1 has a smaller circuit size than the classification device disclosed in Non-Patent Document 1, and the power required to perform the classification processing of the objects to be classified is reduced.

[0048] In Embodiment 1, the classification device is configured to include a learning unit 3 that performs an update process for the weight coefficients of M feature signals, which the optical calculation unit 1 uses when performing weighting matrix calculations based on the calculation results of K activation functions output from the activation function calculation unit 2. Therefore, the classification device can update the weight coefficients of the M feature signals so that the classification accuracy of the classification unit 4 is improved.

[0049] In Embodiment 1, the classification device is configured such that the classification unit 4 compares K calculation results output from the activation function calculation unit 2 and classifies the objects to be classified based on the comparison results of the K calculation results. Therefore, the classification device can classify objects even without the weighting matrix calculation results for each classification.

[0050] In the classification device shown in Figure 1, the activation function calculation unit 2 has K = 3 two-dimensional activation functions f ０ (o １ ,o ２ ), f １ (o １ ,o ２ ), f ２ (o １ ,o ２ This shows a Gaussian distribution function that exhibits a Gaussian distribution with respect to the argument angle. However, this is only one example, and the activation function calculation unit 2 has K = 3 two-dimensional activation functions f ０ (o １ ,o ２ ), f １ (o １ ,o ２ ), f ２ (o １ ,o ２ ) may be a linear distribution function that shows a linear distribution with respect to the argument angle, for example, as shown in Figure 13. Figure 13 shows a two-dimensional activation function f that shows a linear distribution with respect to the argument angle. ０ (o １ ,o ２ ), f １ (o １ ,o ２ ), f ２ (o １ ,o ２ This is an explanatory diagram.

[0051] In the classification device shown in Figure 1, the activation function calculation unit 2 calculates K = 3 two-dimensional activation functions f ０ (o １ ,o ２ ), f １ (o １ ,o ２ ), f ２ (o １ ,o ２As a result of the calculation, the indicators for classification labels (0), (1), and (2) are output to the learning unit 3 and the classification unit 4, respectively. However, this is just one example, and the activation function calculation unit 2 may also convert each of the indicators for classification labels (0), (1), and (2) into probabilities and output the probabilities for classification labels (0), (1), and (2) to the learning unit 3 and the classification unit 4, respectively. In the learning unit 3 and the classification unit 4, the probabilities for classification labels (0), (1), and (2) are treated in the same way as the indicators for classification labels (0), (1), and (2). The process of converting each of the indicators for classification labels (0), (1), and (2) into probabilities is performed, for example, by using the softmax function.

[0052] In the classification device shown in Figure 1, assuming N=2, the activation function calculation unit 2 is shown as having a two-dimensional activation function. However, this is only one example, and assuming N=1, the activation function calculation unit 2 may have a one-dimensional activation function f as shown in equation (7) below. ｉ (o １ It may also have the following characteristics:

[0053]

[0054] The one-dimensional activation function f shown in equation (7) ｉ (o １ As shown in Figure 14, the distribution function is maximized at the center of the number line divided into four sections. The activation function calculation unit 2 calculates the one-dimensional activation function f shown in equation (7). ｉ (o １ If it has ), it is possible to classify the object to be classified into four classification labels. Figure 14 shows the one-dimensional activation function f ｉ (o １ This is an explanatory diagram showing an example where the distribution function is maximized at the center of a number line divided into four sections. In this case, the one-dimensional activation function f ｉ (o １ The index output from ) is a mapping of output coordinates to regions delimited on a one-dimensional line. The regions delimited on the one-dimensional line are, for example, the region of classification label (0), the region of classification label (1), the region of classification label (2), and the region of classification label (3). Therefore, the one-dimensional activation function f ｉ (o １) is a Gaussian distribution function that shows the Gaussian distribution with respect to the mapping location.

[0055] The activation function calculation unit 2 calculates the one-dimensional activation function f ｉ (o １ If it has a one-dimensional activation function f ｉ (o １ If the index of ) is in the range of classification label (0) which is 0 or more and less than 0.25, classify the object to be classified as classification label (0) and use the one-dimensional activation function f ｉ (o １ If the index of ) is in the range of classification label (1), which is 0.25 or more and less than 0.5, the object to be classified is classified to classification label (1). In addition, the classification unit 4 uses a one-dimensional activation function f ｉ (o １ If the index of ) is in the range of classification label (2), which is 0.5 or more and less than 0.75, classify the object to be classified into classification label (2) and use the one-dimensional activation function f ｉ (o １ If the index of ) is in the range of classification label (3), which is 0.75 or more and less than 1, the object to be classified is classified as classification label (3). Note that the one-dimensional activation function f ｉ (o １ The function is not limited to a Gaussian distribution function that shows a Gaussian distribution with respect to the mapping location, but may also be a linear distribution function that shows a linear distribution with respect to the mapping location.

[0056] In the classification device shown in Figure 1, assuming N=2, the activation function calculation unit 2 is shown as having a two-dimensional activation function. However, this is only one example, and assuming N=3, the activation function calculation unit 2 may have a three-dimensional activation function f as shown in equation (8) below. ｉｊｌ (o １ ,o ２ ,o ３ It may also have the following characteristics:

[0057]

[0058] The three-dimensional activation function f shown in equation (8) ｉｊｌ (o １ ,o ２ ,o ３As shown in Figure 15, the function f is a Gaussian distribution that maximizes at the center of each grid cell for a grid of 4 × 4 × 4 = 64 classification labels obtained by dividing each side of a cube into four parts. Figure 15 shows the 3D activation function f ｉｊｌ (o １ ,o ２ ,o ３ This is an explanatory diagram illustrating an example where the function f(x) is a function that exhibits a Gaussian distribution with maximum values ​​at the center of each grid cell for a grid of 64 classification labels (4x4x4) obtained by dividing each edge of a cube into four sections. The three-dimensional activation function f(x) shown in equation (8) ｉｊｌ (o １ ,o ２ ,o ３ The index output from ) is a mapping of output coordinates to regions divided into three-dimensional space. Therefore, the three-dimensional activation function f ｉｊｌ (o １ ,o ２ ,o ３ ) is a Gaussian distribution function that shows the Gaussian distribution with respect to the mapping position. In Figure 15, the region of classification labels (i, j, l) is shown as an example of a region divided into three-dimensional space. Note that the three-dimensional activation function f ｉｊｌ (o １ ,o ２ ,o ３ The function is not limited to a Gaussian distribution function that shows a Gaussian distribution with respect to the mapping location, but may also be a linear distribution function that shows a linear distribution with respect to the mapping location.

[0059] Embodiment 2. Embodiment 2 describes a classification device in which the classification unit 5 maps the N weighting matrix calculation results output from the optical calculation unit 1 to an N-dimensional space, and classifies the objects to be classified based on the mapping positions of the weighting matrix calculation results in the N-dimensional space.

[0060] Figure 16 is a configuration diagram showing a classification device according to Embodiment 2. In Figure 16, the same reference numerals as in Figure 1 indicate the same or corresponding parts, so a detailed explanation is omitted. Figure 17 is a hardware configuration diagram showing the hardware of the classification device according to Embodiment 2. In Figure 17, the same reference numerals as in Figure 2 indicate the same or corresponding parts, so a detailed explanation is omitted. The classification device shown in Figure 16 comprises an optical processing unit 1, an activation function processing unit 2, a learning unit 3, and a classification unit 5.

[0061] The classification unit 5 is implemented, for example, by the classification circuit 25 shown in Figure 17. The classification unit 5 obtains N weighting matrix calculation results from the optical calculation unit 1. The classification unit 5 maps the N weighting matrix calculation results to an N-dimensional space. Based on the mapping positions of the weighting matrix calculation results in the N-dimensional space, the classification unit 5 performs classification processing on the items to be classified. The classification unit 5 outputs the classification results of the items to be classified to, for example, a display device. This allows the user to confirm the classification results of the items to be classified by looking at the display content of the display device.

[0062] In the classification device shown in Figure 1, the classification unit 4 compares K calculation results output from the activation function calculation unit 2 and performs classification processing of the items to be classified based on the comparison result of the K calculation results. The K calculation results from the activation function calculation unit 2 are based on N weighting matrix calculation results output from the optical calculation unit 1, and the classification unit 4 indirectly performs classification processing of the items to be classified based on the N weighting matrix calculation results. In contrast, in the classification device shown in Figure 16, the classification unit 5 directly performs classification processing of the items to be classified based on the N weighting matrix calculation results output from the optical calculation unit 1.

[0063] In Figure 16, it is assumed that the components of the classification device—the optical computing unit 1, the activation function calculation unit 2, the learning unit 3, and the classification unit 5—are each implemented by dedicated hardware as shown in Figure 17. That is, the classification device is assumed to be implemented by an optical computing circuit 21, an activation function calculation circuit 22, a learning circuit 23, and a classification circuit 25. The optical computing circuit 21, the activation function calculation circuit 22, the learning circuit 23, and the classification circuit 25 can each be, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC, an FPGA, or a combination thereof.

[0064] The components of the classification device are not limited to those implemented by dedicated hardware; the classification device may also be implemented by software, firmware, or a combination of software and firmware. When the classification device is implemented by software or firmware, a program that causes a computer to execute the respective processing procedures in the optical processing unit 1, the activation function calculation unit 2, the learning unit 3, and the classification unit 5 is stored in the memory 41 shown in Figure 3. Then, the processor 42 shown in Figure 3 executes the program stored in the memory 41.

[0065] Furthermore, Figure 17 shows an example in which each component of the classification device is implemented by dedicated hardware, and Figure 3 shows an example in which the classification device is implemented by software or firmware, etc. However, this is only one example, and some components of the classification device may be implemented by dedicated hardware, while the remaining components may be implemented by software or firmware, etc.

[0066] Next, the operation of the classification device shown in Figure 16 will be explained. For the sake of explanation, an example where M=4, N=2, and K=3 will be described below. The classification process of the classification unit 5 when the optical computing unit 1 is learning differs from the classification process of the classification unit 4 shown in Figure 1. However, the classification device when the optical computing unit 1 is learning is similar to the classification device shown in Figure 1 in that it calculates the accuracy rate of the judgment result and repeats the learning of the optical computing unit 1 until the accuracy rate exceeds a threshold. For this reason, the operation of the classification device when the optical computing unit 1 is learning will be omitted. The classification process of the classification unit 5 when the optical computing unit 1 is learning is the same as the classification process of the classification unit 5 when classification inference is performed. The operation of the classification device when classification inference is performed will be explained below.

[0067] The optical processing unit 1 receives, for example, four feature signals (a) from an external source, each with different types of features, as feature signals indicating the features of the object to be classified. ０ , a １ , a ２ , a ３ The following is given: Feature signal a ０ , a １ , a ２ , a ３ This is an optical signal in which the feature quantity to be classified is reflected in either the intensity or amplitude of the light. Feature quantity signal a during inference ０ , a １ , a ２ , a ３ This is not training data, but a signal representing feature quantities calculated by some device. The optical calculation unit 1, as shown in equation (1), has M = 4 feature quantity signals a ０ , a １ , a ２ , a ３ As a weighted matrix operation, an N×M matrix operation is performed.

[0068] PD1a-n of the optical computing unit 1 are the calculation results of M = 4 of the weighted matrix calculation o ０ ,o １ ,o ２ ,o ３ The optical signal representing the nth calculation result is converted into an electrical signal, and the electrical signal is output to ADC1b-n. In the example in Figure 6, PD1a-1 represents the calculation result o１ The optical signal indicating the result is converted into an electrical signal, and the electrical signal is output to ADC1b-1, and PD1a-2 calculates the result o ２ The optical signal indicating the result is converted into an electrical signal, and the electrical signal is output to ADC1b-1. ADC1b-n converts the electrical signal output from PD1a-n into digital data, and outputs the digital data to the activation function calculation unit 2 and the classification unit 5, respectively. In the example in Figure 6, ADC1b-1 outputs the calculation result o １ The electrical signal indicating this is converted into digital data, and the ADC1b-2 calculates the result o ２ It converts electrical signals that indicate something into digital data.

[0069] As shown in Figure 18, the classification unit 5 receives the calculation result of the optical calculation unit 1. １ ,o ２ The two-dimensional space relating to this is given by polar coordinates with an argument D (= o ２ / o １ Based on this, the space is divided into multiple regions. Figure 18 is an explanatory diagram showing an example in which a two-dimensional space is divided into multiple regions based on the argument. In the example in Figure 18, the two-dimensional space is divided into the region 0 ≤ D < tan(π / 8), the region tan(π / 8) ≤ D < tan(3π / 8), and the region tan(3π / 8) ≤ D < tan(π / 2).

[0070] The classification unit 5 receives the calculation result o from the ADC1b-1 of the optical calculation unit 1. １ Digital data showing this is obtained, and the calculation result o is obtained from ADC1b-2. ２ Digital data showing the result is obtained. The classification unit 5 calculates the result o １ and calculation result o ２ The data is mapped to the two-dimensional space shown in Figure 18. The classification unit 5 processes the calculation result o １ and calculation result o ２ Based on the mapping position D, the classification target is classified. Specifically, the classification unit 5 calculates the result o １ and calculation result o ２ If the mapping position D is within the region 0 ≤ D < tan(π / 8), the object to be classified is classified as classification label (0). The classification unit 5 calculates the result o １ and calculation result o ２If the mapping position D is within the range tan(π / 8) ≤ D < tan(3π / 8), the object to be classified is classified to classification label (1). The classification unit 5 calculates the result o １ and calculation result o ２ If the mapping position D is within the range tan(3π / 8) ≤ D < tan(π / 2), the object to be classified is classified to classification label (2). The classification unit 5 outputs the classification result of the object to be classified, for example, to a display device. This allows the user to confirm the classification result of the object by looking at the display content of the display device.

[0071] The classification device shown in Figure 16 is an example where N=2, but as with the classification device shown in Figure 1, N can be any value less than the number of categories to be classified, and is not limited to N=2.

[0072] In the above embodiment 2, the classification device is configured to include: an optical calculation unit 1 that, when given M (M is an integer of 2 or more) feature signals of different types representing the features of the object to be classified, performs weighting matrix calculations on the M feature signals and outputs N (N is an integer less than the number of classifications of the object to be classified); and a classification unit 5 that maps the N weighting matrix calculation results output from the optical calculation unit 1 to an N-dimensional space, performs classification processing on the object to be classified based on the mapping position of the weighting matrix calculation results in the N-dimensional space, and outputs the classification result of the object to be classified. Therefore, the classification device can classify the object to be classified using a number of weighting matrix calculation results less than the number of classifications.

[0073] In the classification device shown in Figure 16, the classification unit 5 calculates the calculation result of the optical calculation unit 1. １ ,o ２ The two-dimensional space relating to this is given by polar coordinates with an argument D (= o ２ / o １Based on this, it is divided into multiple regions. However, this is just one example, and the classification unit 5 may divide the two-dimensional space into a grid of regions, as shown in Figure 19. Figure 19 is an explanatory diagram showing an example in which the two-dimensional space is divided into a grid of regions. In the example in Figure 19, the two-dimensional space is divided into 16 regions. If each of the 16 regions corresponds to a classification label (i, j) (i = 1, 2, 3, 4: j = 1, 2, 3, 4), the classification unit 5 can classify the object to be classified into 16 classification labels (i, j). That is, the classification unit 5 calculates the result o １ and calculation result o ２ By identifying the region to which the mapping position D belongs, the object to be classified can be assigned a classification label (i, j) corresponding to the identified region.

[0074] Furthermore, this disclosure allows for free combination of each embodiment, modification of any component in each embodiment, or omission of any component in each embodiment.

[0075] This disclosure is suitable for classification devices and classification methods.

[0076] 1 Optical processing unit, 1a-n PD, 1b-n ADC, 2 Activation function calculation unit, 3 Learning unit, 3a Loss function calculation unit, 3b Weight coefficient update unit, 3c DAC, 4,5 Classification unit, 21 Optical processing circuit, 22 Activation function calculation circuit, 23 Learning circuit, 24,25 Classification circuit, 41 Memory, 42 Processor.

Claims

1. A classification device comprising: an optical calculation unit that, given M (M is an integer of 2 or more) feature signals of different types representing features of a target to be classified, performs weighting matrix calculations on the M feature signals and outputs N (N is an integer less than the number of classifications of the target to be classified) weighting matrix calculation results; an activation function calculation unit that has K (K is the same number as the number of classifications of the target to be classified) different activation functions as input variables to the N weighting matrix calculation results output from the optical calculation unit, calculates each activation function based on the N weighting matrix calculation results and outputs the calculation results of the K activation functions; and a classification unit that performs classification processing of the target to be classified based on the calculation results of the K activation functions output from the activation function calculation unit and outputs the classification result of the target to be classified.

2. The classification device according to claim 1, further comprising a learning unit that performs an update processing of weight coefficients for the M feature signals, which is used when the optical calculation unit performs the weight matrix calculation based on the calculation results of K activation functions output from the activation function calculation unit.

3. The classification device according to claim 2, characterized in that the learning unit calculates a loss function for the K calculation results output from the activation function calculation unit using the true classification result of the object to be classified, and updates the weight coefficients for the M feature signals based on the calculation result of the loss function.

4. The classification device according to any one of claims 1 to 3, wherein N is 2, the K activation functions of the activation function calculation unit are two-dimensional activation functions, and the two-dimensional activation function is a Gaussian distribution function that shows a Gaussian distribution for either the argument representing the ratio of two weighting matrix calculation results output from the optical calculation unit, or the mapping position of the two weighting matrix calculation results in two-dimensional space.

5. The classification device according to any one of claims 1 to 3, wherein N is 2, the K activation functions of the activation function calculation unit are two-dimensional activation functions, and the two-dimensional activation function is a linear distribution function that shows a linear distribution with respect to either the argument representing the ratio of two weighting matrix calculation results output from the optical calculation unit, or the mapping position of the two weighting matrix calculation results in two-dimensional space.

6. The classification device according to any one of claims 1 to 3, wherein N is 2, and the K activation functions in the activation function calculation unit are two-dimensional activation functions such that the angles at which the calculation results of each of the K activation functions are maximized are equally spaced.

7. The classification device according to any one of claims 1 to 6, characterized in that the classification unit compares K calculation results output from the activation function calculation unit and classifies the items to be classified based on the comparison result of the K calculation results.

8. A classification device comprising: an optical processing unit that, given M (M is an integer of 2 or more) feature signals of different types representing features of a target to be classified, performs weighting matrix operations on the M feature signals and outputs N (N is an integer less than the number of classifications of the target to be classified) weighting matrix operation results; and a classification unit that maps the N weighting matrix operation results output from the optical processing unit to an N-dimensional space, performs classification processing on the target to be classified based on the mapping positions of the weighting matrix operation results in the N-dimensional space, and outputs the classification results of the target to be classified.

9. The classification device according to claim 8, comprising: an activation function calculation unit which has K (where K is the same number as the number of classification targets) distinct activation functions as input variables for N weighting matrix calculation results output from the optical calculation unit, calculates each activation function based on the N weighting matrix calculation results and outputs the calculation results of the K activation functions; and a learning unit which performs update processing of the weight coefficients for the M feature signals used when the optical calculation unit performs the weighting matrix calculation based on the calculation results of the K activation functions output from the activation function calculation unit.

10. The classification device according to claim 9, wherein the learning unit calculates a loss function for the K calculation results output from the activation function calculation unit using the true classification result of the object to be classified, and updates the weight coefficients for the M feature signals based on the calculation result of the loss function.

11. The classification device according to any one of claims 8 to 10, wherein N is 2, and the classification unit maps two weighting matrix calculation results output from the optical calculation unit to a two-dimensional space, and classifies the objects to be classified based on the mapping position of the weighting matrix calculation results to the two-dimensional space.

12. The classification device according to claim 11, characterized in that the classification unit divides the two-dimensional space into a plurality of regions and classifies the objects to be classified based on the region to which the mapping position belongs.

13. The classification device according to claim 12, characterized in that the classification unit divides the two-dimensional space into a plurality of regions based on the deviation angle which indicates the ratio of two weighting matrix calculation results output from the optical calculation unit.

14. The classification device according to claim 12, characterized in that the classification unit divides the two-dimensional space into grid-like regions.

15. The classification device according to any one of claims 1 to 14, wherein the optical calculation unit is provided with an optical signal in which the feature quantities of the object to be classified are reflected in either the intensity or amplitude of the light as the feature quantity signal, performs an N × M matrix operation as a weighting matrix operation on the M feature quantity signals, converts the optical signals representing each of the N calculation results of the N × M matrix operation into electrical signals, converts each electrical signal into digital data, and outputs the respective digital data as N weighting matrix operation results.

16. The classification device according to any one of claims 1 to 15, characterized in that the optical processing unit has a processing layer that performs unitary matrix operations as a processing layer that performs weighting matrix operations on the M feature signals.

17. The classification device according to any one of claims 1 to 15, characterized in that the optical processing unit has a processing layer that performs unitary matrix operations and a processing layer that performs diagonal matrix operations as a processing layer that performs weighting matrix operations on the M feature signals.

18. The classification apparatus according to claim 16 or 17, characterized in that the optical processing unit has a processing layer including a Mach-Zehnder interferometer as a processing layer for performing the unitary matrix calculation.

19. The classification apparatus according to claim 18, characterized in that the arrangement of the Mach-Zehnder interferometers included in the calculation layer is either Clements type or Reck type.

20. A classification method comprising: an optical calculation unit, given M (M is an integer of 2 or more) feature signals of different types, which represent the features of the object to be classified, performs a weighting matrix operation on the M feature signals and outputs N (N is an integer less than the number of classifications of the object to be classified) weighting matrix operation results; an activation function calculation unit, which has K (K is the same number as the number of classifications of the object to be classified) different activation functions as functions that take the N weighting matrix operation results output from the optical calculation unit as input variables, calculates each activation function based on the N weighting matrix operation results and outputs the calculation results of the K activation functions; and a classification unit, which performs a classification process on the object to be classified based on the calculation results of the K activation functions output from the activation function calculation unit and outputs the classification result of the object to be classified.

21. A classification method comprising: an optical processing unit, upon receiving M (where M is an integer of 2 or more) feature signals of different types, which represent the features of the object to be classified, performs a weighting matrix operation on the M feature signals and outputs N (where N is an integer less than the number of classifications of the object to be classified) weighting matrix operation results; a classification unit maps the N weighting matrix operation results output from the optical processing unit into an N-dimensional space, performs a classification process on the object to be classified based on the mapping positions of the weighting matrix operation results in the N-dimensional space, and outputs a classification result for the object to be classified.

Images