Optical neural network device

Abstract

To provide a technique capable of increasing the amount of input data of an optical neural network system.SOLUTION: An optical neural network system comprises: a light source generating carrier light; modulation means of subjecting the carrier light to orthogonal frequency multiplex modulation based on input data to generate modulation light; amplification means of amplifying the modulation light; and a nonlinear medium propagating the modulation light amplified by the amplification means.SELECTED DRAWING: Figure 1

Description

Technical Field

[0004] ,

[0002] , , , , , , , , , , , , ,

[0003] , ,

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

Background Art

[0002] Patent Document 1 discloses an optical NN device that hardware - realizes a neural network (hereinafter referred to as NN). The following describes an outline of the configuration disclosed in Patent Document 1. The optical NN device includes a SC light source that generates super - continuum (SC) light. SC light is a pulse signal with a very wide bandwidth. The optical NN device also includes a spectral modulator. The spectral modulator generates modulated light by changing the level of each frequency component of the input light (SC light in Patent Document 1) according to the input data. The modulated light generated by the spectral modulator is input to a spectrometer through a non - linear medium. Due to the non - linear optical effect in the non - linear medium, interference occurs between the frequency components of the modulated light, thereby realizing the connection between neurons. Each frequency component of the modulated light that has passed through the non - linear medium corresponds to the output from each neuron. The spectrometer analyzes the frequency components of the modulated light that has passed through the non - linear medium. This analysis result is used for applications of the optical NN device, such as "classification" and the like.

[0003] The spectral modulator can be realized, for example, by spatially separating the input light according to the frequency components using a diffraction grating or the like, individually attenuating the light of each frequency component using a liquid crystal device such as LCoS (Liquid crystal on silicon), and then multiplexing the light of each frequency component.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

[0005] In the configuration of Patent Document 1, the amount of data input to the optical NN device, that is, the amount of data that can be carried by one symbol of modulated light generated in the optical NN device, depends on the number of resolutions of light in the spectral modulator, that is, the number of frequency components separated in the spectral modulator. For example, if the number of resolutions of light is X, and Y bits of information can be carried by one frequency component of light, then the input data can be up to (X × Y) bits. Therefore, in order to increase the amount of data input to the optical NN device, the number of resolutions of the spectral modulator must be increased.

[0006] However, because spectral modulators have devices that process spatial light, the maximum number of resolutions for light is limited when considering the realistic size of spectral modulators. For example, the number of resolutions X for currently available spectral modulators is around 500.

[0007] This disclosure provides a technique for increasing the amount of data in the input data of an optical neural network device. [Means for solving the problem]

[0008] According to one aspect of the present disclosure, an optical neural network device comprises a light source that generates carrier light, a modulation means that generates modulated light by orthogonal frequency multiplexing modulating the carrier light based on input data, an amplification means that amplifies the modulated light, and a nonlinear medium that propagates the modulated light after amplification by the amplification means. [Effects of the Invention]

[0009] According to this disclosure, the amount of data input to an optical neural network device can be increased. [Brief explanation of the drawing]

[0010] [Figure 1] A diagram illustrating the configuration of an optical neural network device according to one embodiment. [Figure 2] A diagram illustrating modulated light according to one embodiment. [Figure 3] A diagram illustrating the configuration of an optical neural network device according to one embodiment. [Modes for carrying out the invention]

[0011] The embodiments will be described in detail below with reference to the attached drawings. Note that the following embodiments do not limit the invention as defined in the claims, and not all combinations of features described in the embodiments are essential to the invention. Two or more of the features described in the embodiments may be combined arbitrarily. Furthermore, identical or similar configurations will be given the same reference numeral, and redundant descriptions will be omitted.

[0012] <First Embodiment> Figure 1 is a diagram of the configuration of the optical neural network (NN) device according to this embodiment. Light source 1 generates carrier light and outputs it to quadrature modulator (IQ modulator) 2. In the following description, the carrier light will be assumed to be continuous light. However, the carrier light may also be pulsed light obtained by temporally switching the continuous light on and off. When pulsed light is used as the carrier light, light source 1 is configured to output carrier light at least during the period when input data is being input.

[0013] The arithmetic unit 3 generates a DMT (Discrete Multitone) signal based on the input data. The DMT signal is obtained by setting a complex value corresponding to the data value in the positive frequency subcarrier (bin), and setting the complex conjugate of the complex value set in the corresponding positive frequency bin in each negative frequency bin, and then performing a discrete inverse Fourier transform. In this embodiment, for example, a reference amplitude value (reference value) is determined in advance, and the deviation between the amplitude of each bin and the reference value is made to correspond to the data value. However, the amplitude value of each bin may also be made to correspond to the data value. In any case, each bin is intensity modulated.

[0014] The arithmetic unit 3 outputs a DMT signal to the in-phase (I) port of the IQ modulator 2 and a Hilbert-transformed signal of the DMT signal to the quadrature-phase (Q) port. The IQ modulator 2 performs quadrature modulation (IQ modulation) based on the signals input to the I port and Q port and outputs modulated light. More specifically, the IQ modulator 2 splits the carrier light from the light source 1 into two branches and modulates the intensity of each branch with the signal input to the I port and the signal input to the Q port. The IQ modulator 2 generates modulated light by combining the two intensity-modulated lights with a phase difference of π / 2. By performing IQ modulation based on the DMT signal and the Hilbert-transformed signal of the DMT signal, the negative frequency components of the DMT signal are canceled out, and OFDM modulated light with multiple subcarriers carrying complex values ​​set in the positive frequency bins is obtained. Figure 2 shows an example of OFDM modulated light according to this embodiment. Each vertical line on the frequency axis corresponds to a subcarrier, and the deviation between the amplitude of the subcarrier and the reference value corresponds to the data value. Note that the configuration for generating OFDM-modulated light (hereinafter simply referred to as modulated light) is not limited to the configuration shown in Figure 1, and any configuration can be used.

[0015] Returning to Figure 1, amplifier 4 amplifies the modulated light. Amplifier 4 is any optical amplifier capable of amplifying the entire frequency band of the modulated light, and could be, for example, an L-beam doped fiber amplifier (EDFA) or a Raman amplifier. Amplifier 4 outputs the amplified modulated light to the nonlinear medium 5.

[0016] The nonlinear medium 5 is a medium that induces nonlinear optical effects on modulated light. Examples of nonlinear optical effects include self-phase modulation, cross-phase modulation, four-wave mixing, and Raman scattering. Examples of nonlinear medium 5 include dispersion planarization fibers (DFFs), dispersion shift fibers (DSFs), and dispersion reduction fibers (DDFs). Furthermore, a nanowire waveguide formed on a silicon substrate can be used as the nonlinear medium 5. The spectrum of the modulated light broadens significantly due to the influence of nonlinear optical effects. Each frequency component (wavelength component) of the modulated light output by the nonlinear medium 5 is a result of the coupling of each frequency component of the original modulated light, with each frequency component corresponding to a neuron.

[0017] The processing unit 6 processes the modulated light that has undergone a nonlinear optical effect in the nonlinear medium 5 and outputs output data according to the application of the optical NN device. Applications of the optical NN device include, for example, prediction and classification. The coefficients (weights) used in the processing performed by the processing unit 6 are determined during the training of the optical NN device. For example, the processing unit 6 may output output data by analyzing the levels of the frequency components of the input modulated light with a spectrometer (spectrum analyzer) and processing the levels of each frequency component. Alternatively, for example, the processing unit 6 may photoelectrically convert the input modulated light into an electrical signal, analyze the levels of the frequency components of the electrical signal with a spectrometer (spectrum analyzer), and output output data by processing the levels of each frequency component.

[0018] An optional dispersion medium can be provided between the nonlinear medium 5 and the processing unit 6. The dispersion medium is a medium that provides different delays depending on the frequency components of the modulated light. Examples of dispersion mediums 4 include single-mode fiber (SMF), dispersion-compensated fiber (DCF), fiber grating, waveguide-type dispersion compensator, etc. By propagating through the dispersion medium, each frequency component of the modulated light is mixed according to the wavelength dispersion.

[0019] As described above, the optical neural network device according to this embodiment uses optical OFDM modulation. For example, in DMTs, devices that transmit data using approximately 8000 bins have been put into practical use. Here, the number of bins corresponds to the number of resolutions X mentioned above when using a spectral modulator. Therefore, with the configuration of this embodiment, the amount of input data for the optical neural network device can be increased.

[0020] <Second Embodiment> Next, 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 processing unit 6 also outputs output data to the calculation unit 3. The calculation unit 3 modulates a part of all the bins of the DMT signal, for example, one bin, with the output data, and modulates the remaining bins with the input data. By recursively feeding back the output data to the input side, the optical NN device can be used for prediction and the like.

[0021] With the above configuration, the data amount of the input data of the optical neural network device can be increased. Therefore, it becomes possible to contribute to Goal 9 of the Sustainable Development Goals (SDGs) led by the United Nations, "Build resilient infrastructure, promote sustainable industrialization, and foster innovation."

Description of Reference Numerals

[0022] 1: Light source, 2: Quadrature modulator, 3: Calculation unit, 4: Amplifier, 5: Nonlinear medium

Claims

1. A light source that generates carrier light, A modulation means that generates modulated light by orthogonal frequency multiplexing the carrier light based on input data, Amplification means for amplifying the modulated light, A nonlinear medium that propagates the modulated light after amplification by the amplification means, An optical neural network device equipped with [a specific feature].

2. The optical neural network apparatus according to claim 1, wherein the nonlinear medium includes at least one of a dispersion planarizing fiber, a dispersion shifting fiber, a dispersion reduction fiber, and a nanowire waveguide formed on a silicon substrate.

3. The optical neural network device according to claim 1, further comprising delay means for giving different propagation delays to each frequency component of the modulated light propagated through the nonlinear medium.

4. The optical neural network apparatus according to claim 3, 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.

5. The modulation means is A generation means for generating discrete multitone signals based on the aforementioned input data, A quadrature modulation means that generates the modulated light by quadrature modulating the carrier light based on the discrete multitone signal, An optical neural network device according to claim 1, comprising the following:

6. The generation means generates a converted signal by performing a Hilbert transform on the discrete multitone signal, The optical neural network device according to claim 5, wherein the orthogonal modulation means generates the modulated light by combining a first intensity-modulated light obtained by intensity-modulating the carrier light based on the discrete multitone signal and a second intensity-modulated light obtained by intensity-modulating the carrier light based on the conversion signal.

7. The optical neural network device according to any one of claims 1 to 6, further comprising processing means for processing the modulated light propagated through the nonlinear medium and outputting output data.

8. The optical neural network device according to claim 7, wherein the processing means outputs the output data by analyzing the frequency components of the modulated light or the frequency components of the electrical signal obtained by photoelectric conversion of the modulated light.

9. The optical neural network device according to claim 7, wherein the modulation means further modulates the carrier light using orthogonal frequency multiplexing based on the output data.

Images