Optical Neural Network Device

Abstract

To provide an optical neural network device capable of stably operating and of increasing the number of neurons.SOLUTION: An optical neural network device includes: first modulating means for modulating first carrier light by a first electric signal to output first modulated light; a non-linear medium that converts the first modulated light or second modulated light generated on the basis of the first modulated light into third modulated light; separating means for performing frequency separation on the third modulated light to output fourth modulated light; converting means for converting each of the plurality of fourth modulated lights into a second electric signal to output the plurality of second electric signals; and adding means for generating an output signal obtained by performing weighting addition on each of the plurality of second electric signals.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present disclosure relates to an optical neural network device for hardware-implementing a neural network.

Background Art

[0002] Non-Patent Document 1 and Non-Patent Document 2 disclose optical neural network devices. According to Non-Patent Document 1, light is spatially modulated with input data using a spatial light phase modulator, and the coupling between neurons is realized by passing the light through a scattering medium. Note that the light that has passed through the scattering medium is converted into an electrical signal by an imaging device. Further, according to Non-Patent Document 2, phase-modulated light phase-modulated with input data is propagated in a multimode waveguide in a plurality of propagation modes. In the multimode waveguide, coupling between propagation modes occurs, but since the propagation time in the multimode waveguide differs according to the propagation mode, the coupling between neurons is realized.

Prior Art Documents

Non-Patent Documents

[0003]

Non-Patent Document 1

Non-Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0004] Since the configuration described in Non-Patent Document 1 uses a spatial optical system, adjustment of the position and direction of each optical device is required. Furthermore, since the position and direction of each optical device can change with temperature changes, it becomes difficult to operate stably. Also, in the configuration described in Non-Patent Document 2, since the number of propagation modes corresponds to the number of neurons, it is necessary to increase the number of propagation modes in order to increase the complexity of the connections between neurons. However, the number of propagation modes is on the order of several tens, and it is difficult to increase the number of propagation modes.

[0005] The present disclosure provides an optical neural network device that can operate stably and increase the number of neurons.

Means for Solving the Problems

[0006] According to one aspect of the present disclosure, an optical neural network device includes: a first modulation means for modulating a first carrier light with a first electrical signal to output a first modulated light; a non-linear medium for converting the first modulated light or a second modulated light generated based on the first modulated light into a third modulated light; a separation means for frequency-separating the third modulated light to output a plurality of fourth modulated lights; a conversion means for converting each of the plurality of fourth modulated lights into a second electrical signal to output a plurality of second electrical signals; and an addition means for generating an output signal obtained by weighted addition of each of the plurality of second electrical signals. a branching means for branching the third modulated light and outputting one of the third modulated lights to the separating means; and a multiplexing means for multiplexing the other third modulated light branched and output by the branching means and the first modulated light to generate the second modulated light; It is provided with.

Effects of the Invention

[0007] According to the present disclosure, an optical neural network device that can operate stably and increase the number of neurons is provided.

Brief Description of the Drawings

[0008]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Mode for Carrying Out the Invention

[0009] Hereinafter, embodiments will be described in detail with reference to the accompanying drawings. Note that the following embodiments do not limit the invention according to the claims, and not all combinations of the features described in the embodiments are essential for the invention. Two or more of the plurality of features described in the embodiments may be arbitrarily combined. Also, the same or similar configurations are given the same reference numerals, and duplicate explanations are omitted.

[0010] <First Embodiment> FIG. 1 is a configuration diagram of an optical neural network device (hereinafter referred to as an optical NN device) according to the present embodiment. Input data to the optical NN device is input to the modulator 2. The modulator 2 modulates the carrier light from the light source 1 with the input data and outputs modulated light. In the following description, the carrier light is continuous light, but the carrier light may be a signal that turns on and off temporally such as pulsed light. The modulation method used by the modulator 2 is arbitrary, and can be, for example, intensity modulation or angle modulation. Angle modulation includes phase modulation or frequency modulation. The amplifier 3 amplifies the modulated light. The amplifier 3 is an arbitrary optical amplifier according to the frequency band of the modulated light, and can be, for example, an erbium-doped fiber amplifier (EDFA) or a Raman amplifier. The amplifier 3 outputs the amplified modulated light to the nonlinear medium 4.

[0011] The nonlinear medium 4 is a medium that causes a nonlinear optical effect on the modulated light. The nonlinear optical effects are, for example, self-phase modulation, cross-phase modulation, four-wave mixing, Raman scattering, etc. As the nonlinear medium 4, for example, a dispersion-flattened fiber (DFF), a dispersion-shifted fiber (DSF), a dispersion-decreasing fiber (DDF), etc. can be used. Further, as the nonlinear medium 4, a thin wire waveguide formed on a silicon substrate can be used. The spectrum of the modulated light greatly spreads due to the influence of the nonlinear optical effect. FIG. 2(A) shows the spectrum of the modulated light input to the nonlinear medium 4, and FIG. 2(B) shows the spectrum of the modulated light output from the nonlinear medium 4. Each frequency component (wavelength component) of the modulated light output from the nonlinear medium 4 is due to the combination of the components of the original modulated light, and each frequency component corresponds to a neuron.

[0012] The arrayed waveguide grating (AWG) 5 is a frequency separation unit (wavelength separation unit) that separates the modulated light output from the non-linear medium 4 for each frequency component. The photodiode (PD) 6 is a photoelectric conversion unit that performs photoelectric conversion on the modulated light of each frequency output from the AWG 5. The multiplier 7 multiplies a weight to the electrical signal output from each PD 6. Note that the weight by which each multiplier 7 multiplies the electrical signal is determined during the learning of the optical NN device and is set in each multiplier 7 in advance. The adder 8 adds the electrical signals output from each multiplier and outputs the output signal of the optical NN device.

[0013] As described above, in the optical NN device according to this embodiment, since a spatial optical system is not used, adjustment of the position and direction of each optical device is unnecessary and it can operate stably. Also, in the optical NN device according to this embodiment, the number of neurons depends on the frequency band of the modulated light output from the non-linear medium 4. That is, in the optical NN device according to this embodiment, the number of neurons corresponds to the number of modulated lights output from the AWG 5, and about several thousand are possible.

[0014] In this embodiment, frequency separation is performed using the AWG 5, but it is also possible to configure the frequency separation unit by combining the AWG 5 with other optical devices, for example, a band-pass filter or an interleaver. Also, in this embodiment, the output signal is generated by weighting and adding the analog signals output from each PD 6. However, a configuration may also be adopted in which the analog signals output from each PD 6 are converted into digital signals by an analog-to-digital converter, and then weighting and addition are performed in the digital domain.

[0015] <Second Embodiment> Subsequently, the differences between the second embodiment and the first embodiment will be mainly described. FIG. 3 is a configuration diagram of the optical NN device according to this embodiment. In this embodiment, the modulated light output from the non-linear medium 4 is input to the AWG 5 via the dispersion medium 9.

[0016] The dispersion medium 9 is a medium that provides different delays according to the frequency components of the modulated light. As the dispersion medium 9, for example, a single-mode fiber (SMF), a dispersion compensating fiber (DCF), a fiber grating, a waveguide type dispersion compensator, etc. can be used. By propagating the dispersion medium 9 to give different delays to each frequency component, the correlation between each frequency component can be reduced, that is, the correlation of the output of the neuron can be reduced.

[0017] <Third Embodiment> Subsequently, the third embodiment will be described focusing on the differences from the previous embodiments. FIG. 4 is a configuration diagram of the optical NN device according to this embodiment. In this embodiment, the modulated light output from the non-linear medium 4 is input to the optical deinterleaver 10. The optical deinterleaver 10 separates the modulated light output from the non-linear medium 4 for each frequency component, and outputs one of the adjacent frequency components to the dispersion medium 9-1 and the other to the dispersion medium 9-2. For example, when the spectrum of the modulated light output from the non-linear medium 4 is as shown in FIG. 2(B), the optical interleaver 10 outputs the frequency components of the white portion in FIG. 2(C) to the dispersion medium 9-1 and the frequency components of the shaded portion to the dispersion medium 9-2. The dispersion media 9-1 and 9-2 are the same as the dispersion medium 9 of the second embodiment. The modulated light output from each of the dispersion media 9-1 and 9-2 is separated for each frequency component by the AWGs 5-1 and 5-2, respectively. The subsequent processing is the same as that of the first and second embodiments.

[0018] In the modulated light output from the non-linear medium 4, the correlation between adjacent frequency components becomes high. In the present embodiment, in the modulated light output from the non-linear medium 4, the optical interleaver 10 separates the components of the modulated light such that adjacent frequency components are different modulated light components, and a delay corresponding to the frequency is given to each of the two modulated lights output from the optical interleaver 10. With this configuration, the correlation between adjacent frequency components in the modulated light output from the non-linear medium 4 can be reduced. Note that, by using media having different dispersion characteristics (relationship between frequency (wavelength) and delay) as the dispersion medium 9-1 and the dispersion medium 9-2, the correlation between adjacent frequency components in the modulated light output from the non-linear medium 4 can be made even lower.

[0019] <Fourth Embodiment> Subsequently, the fourth embodiment will be described focusing on the differences from the previous embodiments. FIG. 5 is a configuration diagram of an optical NN device according to the present embodiment. In the present embodiment, the output of the modulator 2 is input to the amplifier 3 via the coupler 11. Also, the modulated light output from the non-linear medium 4 is branched into two by the branching unit 12, one is input to the AWG 5, and the other is input to the coupler 11. That is, the coupler 11 combines the modulated light from the modulator 2 and the modulated light from the branching unit 12 and outputs it to the amplifier 3.

[0020] In the present embodiment, at the coupler 11, the output of the branching unit 12 and the output of the modulator 2 are made to interfere with each other. With this configuration, the complexity of the coupling between neurons can be increased. Note that, although FIG. 5 is obtained by adding a feedback configuration from the branching unit 12 to the coupler 11 to the configuration of the first embodiment, a feedback configuration can also be added to the configuration of the second embodiment or the third embodiment.

[0021] <Fifth Embodiment> Next, the fifth embodiment will be described focusing on the differences from the fourth embodiment. FIG. 6 is a configuration diagram of the optical NN device according to this embodiment. In the fourth embodiment, a feedback configuration was provided in the optical domain. In this embodiment, a feedback configuration is provided in the electrical domain. Specifically, as shown in FIG. 6, in this embodiment, the output of the adder 8 is branched into two at the branching unit 13. One is used as the output signal of the optical NN device, and the other is input to the combining unit 14. The combining unit 14 adds or subtracts the output signal from the branching unit 13 to the input signal of the optical NN device and inputs it to the modulator 2. Note that the addition or subtraction may be weighted addition or subtraction.

[0022] <Sixth Embodiment> Next, the sixth embodiment will be described focusing on the differences from the fifth embodiment. Also in this embodiment, a feedback configuration is provided in the electrical domain as in the fifth embodiment. FIG. 7 is a configuration diagram of the optical NN device according to this embodiment. In this embodiment, two modulators 2-1 and 2-2 are provided. Note that the modulation methods in the modulators 2-1 and 2-2 are different from each other. For example, when the modulator 2-1 uses intensity modulation, the modulator 2-2 uses phase modulation, and when the modulator 2-1 uses phase modulation, the modulator 2-2 uses intensity modulation. The modulator 2-1 modulates the carrier light from the light source 1 with the input signal to the optical NN device and outputs the modulated light to the modulator 2-2. The modulator 2-2 modulates the modulated light from the modulator 2-1 with the output signal from the branching unit 13. Note that when the modulation methods in the modulators 2-1 and 2-2 are the same as each other, the configuration is equivalent to that of the fifth embodiment.

[0023] <Seventh Embodiment> Next, regarding the seventh embodiment, the differences from the sixth embodiment will be mainly described. Similar to the sixth embodiment, this embodiment also provides a feedback configuration in the electrical domain. FIG. 8 is a configuration diagram of the optical NN device according to this embodiment. Also in this embodiment, similar to the sixth embodiment, two modulators 2-1 and 2-2 are provided. However, the modulator 2-1 modulates the carrier light from the light source 1-1 with the input signal to the optical NN device, and the modulator 2-2 modulates the carrier light from the light source 1-2 with the output signal of the optical NN device. The multiplexer 14 multiplexes the modulated lights output from the modulator 2-1 and the modulator 2-2, and then inputs them to the amplifier 3.

[0024] With the above configuration, an optical neural network device that can operate stably and increase the number of neurons can be realized. Therefore, it becomes possible to contribute to Goal 9 of the Sustainable Development Goals (SDGs) led by the United Nations, which is to "build resilient infrastructure, promote sustainable industrialization, and foster innovation."

Explanation of Reference Numerals

[0025] 2: Modulator, 4: Nonlinear medium, 5: AWG, 6: PD, 7: Adder

Claims

1. A first modulation means for modulating a first carrier light with a first electrical signal to output a first modulated light; A non-linear medium for converting a second modulated light generated based on the first modulated light into a third modulated light; A separation means for separating the third modulated light into a plurality of fourth modulated lights and outputting them; A conversion means for converting each of the plurality of fourth modulated lights into a second electrical signal to output a plurality of second electrical signals; An addition means for generating an output signal obtained by weighted addition of each of the plurality of second electrical signals; A branching means for branching the third modulated light and outputting one of the third modulated lights to the separation means; A multiplexing means for generating the second modulated light by multiplexing the other third modulated light branched and output by the branching means and the first modulated light; An optical neural network device comprising the above.

2. A first modulation means for modulating a first carrier light with a first electrical signal to output a first modulated light; A non-linear medium for converting a second modulated light generated based on the first modulated light into a third modulated light; A separation means for separating the third modulated light into a plurality of fourth modulated lights and outputting them; A conversion means for converting each of the plurality of fourth modulated lights into a second electrical signal to output a plurality of second electrical signals; An addition means for generating an output signal obtained by weighted addition of each of the plurality of second electrical signals; A second modulation means for generating the second modulated light by modulating the first modulated light with the output signal; Comprising: An optical neural network device, wherein the first modulation means and the second modulation means use different modulation methods.

3. A first modulation means for modulating a first carrier light with a first electrical signal to output a first modulated light; A non-linear medium for converting a second modulated light generated based on the first modulated light into a third modulated light; A separation means for separating the third modulated light into a plurality of fourth modulated lights and outputting them; A conversion means for converting each of the plurality of fourth modulated lights into a second electrical signal to output a plurality of second electrical signals; An addition means for generating an output signal obtained by weighted addition of each of the plurality of second electrical signals; A second modulation means for modulating a second carrier light with the output signal to output a fifth modulated light; A multiplexing means for multiplexing the first modulated light and the fifth modulated light to generate the second modulated light; Comprising: An optical neural network device, wherein the first modulation means and the second modulation means use different modulation methods.

4. The optical neural network device according to any one of claims 1 to 3, wherein the non-linear medium converts the second modulated light into the third modulated light by causing a non-linear optical effect on the second modulated light.

5. The optical neural network device according to any one of claims 1 to 3, wherein a frequency band of the third modulated light is wider than a frequency band of the second modulated light.

6. The optical neural network device according to any one of claims 1 to 3, wherein the non-linear medium includes at least one of a dispersion-flattened fiber, a dispersion-shifted fiber, a dispersion-reducing fiber, and a thin wire waveguide formed on a silicon substrate.

7. Further comprising delay means for providing different propagation delays according to frequency, The optical neural network device according to any one of claims 1 to 3, wherein the separation means frequency-separates the third modulated light that has propagated through the delay means.

8. The optical neural network device according to claim 7, wherein the delay means includes at least one of a single-mode fiber, a dispersion-compensating fiber, a fiber grating, and a waveguide-type dispersion compensator.

9. The optical neural network device according to any one of claims 1 to 3, wherein the first electrical signal is an input signal of the optical neural network device.

10. A first modulation means for modulating a first carrier light with a first electrical signal to output a first modulated light; A non-linear medium for converting the first modulated light or a second modulated light generated based on the first modulated light into a third modulated light; An optical deinterleaver means for outputting a fourth modulated light and a fifth modulated light by deinterleaving the third modulated light in a frequency domain; First delay means for providing a propagation delay according to frequency to the fourth modulated light; Second delay means for providing a propagation delay according to frequency to the fifth modulated light; First separation means for frequency-separating the fourth modulated light that has passed through the first delay means to output a plurality of sixth modulated lights; Second separation means for frequency-separating the fifth modulated light that has passed through the second delay means to output a plurality of seventh modulated lights; Conversion means for outputting a plurality of second electrical signals by converting each of the plurality of sixth modulated lights and the plurality of seventh modulated lights into a second electrical signal; Adding means for generating an output signal obtained by weighted addition of each of the plurality of second electrical signals; An optical neural network device comprising the above.

11. The optical neural network device according to claim 10, wherein the relationship between the delay and the frequency provided by the first delay means and the second delay means is different.

Images