Optical Neural Network Device
By combining modulated and pulsed light and amplifying it before inputting into a nonlinear medium, the optical neural network device overcomes power limitations, increasing neuron count and performance.
Patent Information
- Application Number
- JP2022157336
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2022-09-30
- Publication Date
- 2025-09-03
- Estimated Expiration
- 2042-09-30
AI Technical Summary
Existing optical neural network devices are limited by the inability to amplify the power of modulated light to the nonlinear medium, restricting the bandwidth and number of neurons.
The device generates signal light by combining modulated light with pulsed light and amplifying it before inputting it into a nonlinear medium, allowing for a strong nonlinear optical effect and increased neuron count.
This configuration enables a significant increase in the number of neurons in the optical neural network device, enhancing its performance and capabilities.
Smart Images

Figure 0007733629000001 
Figure 0007733629000002 
Figure 0007733629000003
Abstract
Description
[Technical Field]
[0001] The present disclosure relates to an optical neural network device that implements a neural network in hardware. [Background technology]
[0002] Patent Document 1 discloses an optical neural network (NN) device that implements a neural network (NN) in hardware. The following provides an overview of the configuration disclosed in Patent Document 1. The optical NN device includes a supercontinuum (SC) light source that generates SC light. SC light is a pulse signal with an extremely wide bandwidth. The optical NN device also includes a spectral modulator that changes the waveform of the SC light on the frequency axis in accordance with input data. The modulated light generated by the spectral modulator is input to a spectrometer via a nonlinear medium. Nonlinear optical effects in the nonlinear medium cause interference between the frequency components of the modulated light, thereby realizing connections between neurons. Each frequency component of the modulated light that has passed through the nonlinear medium corresponds to the output from each neuron. The spectrometer analyzes the frequency components of the modulated light that has passed through the nonlinear medium. The analysis results are used in applications of the optical NN device, such as classification. [Prior art documents] [Patent documents]
[0003] [Patent Document 1] US Patent Application Publication No. 2020 / 0311532 Summary of the Invention [Problem to be solved by the invention]
[0004] The stronger the power of the input light, the stronger the nonlinear optical effect in a nonlinear medium. However, even if a Raman amplifier is used, it is not possible to amplify the entire bandwidth of SC light, so the configuration of Patent Document 1 cannot increase the power of the modulated light input to the nonlinear medium. Therefore, it is not possible to generate a strong nonlinear optical effect in the nonlinear medium, and the bandwidth of the modulated light output by the nonlinear medium is also limited, which in turn limits the number of neurons.
[0005] The present disclosure provides a technique that can increase the number of neurons in an optical neural network device. [Means for solving the problem]
[0006] According to one aspect of the present disclosure, an optical neural network device includes a first generating means for generating carrier light, a second generating means for generating first pulse light, a modulating means for modulating the carrier light based on input data and outputting modulated light, a combining means for combining the modulated light and the first pulse light and outputting signal light, an amplifying means for amplifying the signal light, and a nonlinear medium for propagating the signal light after being amplified by the amplifying means. [Effects of the Invention]
[0007] According to the present disclosure, the number of neurons in an optical neural network device can be increased. [Brief explanation of the drawings]
[0008] [Figure 1] FIG. 1 is a configuration diagram of an optical neural network device according to an embodiment. [Figure 2] 3A and 3B are diagrams illustrating the configuration of a generation unit according to an embodiment and an explanatory diagram of signal light output by the generation unit. [Figure 3] FIG. 2 is a configuration diagram of a processing unit according to an embodiment. [Figure 4] FIG. 2 is a diagram illustrating the configuration of a generation unit according to an embodiment. [Figure 5] 3A and 3B are diagrams illustrating the configuration of a generating unit and modulated light according to an embodiment. DETAILED DESCRIPTION OF THE INVENTION
[0009] Hereinafter, the embodiments will be described in detail with reference to the accompanying drawings. Note that the following embodiments do not limit the scope of the invention as claimed, and not all combinations of features described in the embodiments are necessarily essential to the invention. Two or more features among the multiple features described in the embodiments may be arbitrarily combined. Furthermore, the same reference numerals are used for the same or similar components, and redundant explanations will be omitted.
[0010] First Embodiment FIG. 1 is a configuration diagram of an optical NN device according to this embodiment. A generator 1 generates signal light based on input data and outputs the signal light to an amplifier 2. FIG. 2(A) is a configuration diagram of the generator 1 according to this embodiment. A modulator 12 modulates carrier light from a light source 11 with the input data and outputs the modulated light to a coupler 13. In the following description, the carrier light is assumed to be continuous light. However, the carrier light may be pulsed light that is turned on and off in time. When pulsed light is used as the carrier light, the light source 11 is configured to output the carrier light at least during the period when the input data is being input. The modulator 12 may use any modulation method, such as intensity modulation or angle modulation. Angle modulation includes phase modulation or frequency modulation. The modulator 12 may also use a modulation method in which multiple subcarriers each carry data, such as optical orthogonal frequency division multiplexing (OFDM) modulation.
[0011] The pulsed light source 14 generates pulsed light that periodically turns on and off and outputs it to the coupler 13. The coupler 13 combines the modulated light from the modulator 12 with the pulsed light from the pulsed light source 14 and outputs the combined signal light. Note that any optical device that combines and outputs multiple input light beams can be used instead of the coupler 13. FIG. 2(B) shows the frequency components of the signal light. In FIG. 2(B), reference numeral 92 corresponds to pulsed light, and reference numeral 94 corresponds to modulated light. In FIG. 2(B), the frequency of the carrier light output by the light source 11 and the frequency of the pulsed light output by the pulsed light source 14 are different, but they may be the same. In FIG. 2(B), the frequency of the carrier light output by the light source 11 is lower than the frequency of the pulsed light output by the pulsed light source 14, but they may be higher.
[0012] Returning to FIG. 1, amplifier 2 amplifies the signal light. Amplifier 2 may be any optical amplifier capable of amplifying the entire frequency band of the signal light, such as an elliptical-beam-doped fiber amplifier (EDFA) or a Raman amplifier. The frequency band of the signal light is the sum of the frequency band of the modulated light and the frequency band of the pulsed light. Amplifier 2 outputs the amplified signal light to nonlinear medium 3.
[0013] The nonlinear medium 3 is a medium that generates a nonlinear optical effect on the signal light. Examples of nonlinear optical effects include self-phase modulation, cross-phase modulation, four-wave mixing, and Raman scattering. For example, a dispersion-flattened fiber (DFF), a dispersion-shifted fiber (DSF), or a dispersion-decreasing fiber (DDF) can be used as the nonlinear medium 3. Furthermore, a thin-wire waveguide formed on a silicon substrate can be used as the nonlinear medium 3. The spectrum of the signal light is significantly broadened by the influence of the nonlinear optical effect. Each frequency component (wavelength component) of the signal light output by the nonlinear medium 3 is a combination of each frequency component of the original signal light, and each frequency component corresponds to a neuron.
[0014] When the optional dispersion medium 4 is provided, the signal light that has passed through the nonlinear medium 3 is input to the dispersion medium 4. The dispersion medium 4 is a medium that imparts different delays depending on the frequency components of the signal light. For example, a single mode fiber (SMF), a dispersion compensation fiber (DCF), a fiber grating, a waveguide-type dispersion compensator, etc. can be used as the dispersion medium 4. By propagating through the dispersion medium 4, each frequency component of the signal light is mixed between bits depending on the chromatic dispersion. The signal light that has passed through the dispersion medium 4 is input to the processing unit 5. Note that when the dispersion medium 4 is not used, the signal light from the nonlinear medium 3 is input to the processing unit 5.
[0015] The processing unit 5 processes the signal light that has been subjected to the nonlinear optical effect in the nonlinear medium 3 and that has optionally been given different delays for each frequency component, and outputs output data. Figure 3 shows an example of the processing unit 5. The arrayed waveguide grating (AWG) 51 is a frequency separator (wavelength separator) that separates the signal light into frequency components. As described above, the light of each frequency corresponds to the output from the neuron. The photodiode (PD) 52 is an optoelectronic converter that performs optoelectronic conversion of the light of each frequency output by the AWG 51. The calculation unit 53 performs a predetermined calculation on the electrical signal output by each PD 52. Note that multiple coefficients (weights) used in the predetermined calculation are determined during training of the optical neural network device. For example, the calculation unit 53 converts the electrical signal output by each PD 52 into a digital value, multiplies the digital value by the weight determined during training, adds the multiplied values, and outputs the resulting data value as output data. The output data output by the processing unit 5 is used for applications in the optical neural network device. The application of the optical NN device is classification etc. when the dispersion medium 4 is not used, and prediction etc. when the dispersion medium 4 is used.
[0016] The configuration of the processing unit 5 is not limited to that shown in Fig. 3. For example, the processing unit 5 may analyze the frequency components of the input signal light with a spectrometer (spectrum analyzer) and output a value based on the analysis result. Alternatively, the processing unit 5 may analyze the electrical signals output from each PD 52 in Fig. 3 with a spectrometer and output a value based on the analysis result of the electrical signals output from each PD 52.
[0017] As described above, the optical NN device according to this embodiment does not modulate SC light, but rather modulates carrier light with a narrower bandwidth than the SC light. Then, signal light is generated by combining pulsed light and modulated light, and the signal light is amplified by amplifier 2 before being input to nonlinear medium 3. The frequency band of the signal light is sufficiently narrow compared to SC light, allowing it to be amplified by amplifier 2. Therefore, a strong nonlinear optical effect can be generated in nonlinear medium 3, thereby increasing the number of neurons.
[0018] Furthermore, adjusting the power of the pulsed light relative to the power of the modulated light can prevent the correlation between the input data and data corresponding to each frequency component of the signal light output by the nonlinear medium 3 from becoming too low. For example, when the output data is used for prediction, the data corresponding to each frequency component of the signal light output by the nonlinear medium 3 must have a certain degree of correlation with the input data that serves as the basis for the prediction. However, depending on the modulation depth of the modulated light, the phase rotation of each frequency component in the nonlinear medium 3 can cause the correlation between the input data and the data corresponding to each frequency component to become too low. In this embodiment, the modulated light and the pulsed light are combined to generate the signal light, so the modulation depth of the entire signal light is lower than that of the modulated light. In other words, the modulation depth of the signal light can be adjusted by adjusting the power of the pulsed light. This prevents the correlation between the input data and data corresponding to each frequency component of the signal light output by the nonlinear medium 3 from becoming too low. As an example, the generator 1 is configured so that the power of the pulsed light can be changed from approximately 1 to 100 times the power of the modulated light.
[0019] In this embodiment, one set of light source 11 and modulator 12 generates one modulated light, which is then combined with one pulsed light to generate signal light. However, a configuration may be adopted in which multiple sets of light source 11 and modulator 12 are provided to generate multiple modulated light, and the multiple modulated light is combined with one pulsed light to generate signal light. In this case, the frequency of the carrier light generated by the light source 11 in each set is made different. Using multiple modulated light can increase the amount of input data. Even in this case, the frequency band of the signal light must be within the frequency band that can be amplified by amplifier 2.
[0020] Although not shown in FIG. 1, the output data of the processing unit 5 can also be fed back to the generation unit 1. In this case, the generation unit 1 generates signal light based on, for example, input data and output data. Furthermore, when an optical NN device is used for prediction, the input data that serves as the basis for prediction can be input to the optical NN device to obtain output data, and then this output data can be input as input data to the generation unit 1. This can be done to obtain prediction data that is a prediction result at a first time point based on the input data that serves as the basis for prediction, and then obtain a prediction result at a second time point that is later in time than the first time point based on this prediction data at the first time point.
[0021] Second Embodiment Next, the second embodiment will be described, focusing on the differences from the first embodiment. FIG. 4(A) is a configuration diagram of the generator 1 according to this embodiment. In this embodiment, pulsed light generated by a pulsed light source 14 is amplified by an amplifier 15 and input to a coupler 13 via a nonlinear medium 16. The nonlinear optical effect of the nonlinear medium 16 makes the bandwidth of the pulsed light wider than the original pulsed light, and the resulting pulsed light can then be combined with the modulated light. This makes it possible to further widen the bandwidth of the signal light output by the nonlinear medium 3. Note that, in this embodiment as well, each device is configured so that the frequency band of the signal light falls within the frequency band that can be amplified by the amplifier 2. For example, the gain of the amplifier 15 is adjusted so that the frequency band of the pulsed light output by the nonlinear medium 16 falls within the range that can be amplified by the amplifier 2.
[0022] Third Embodiment Next, the third embodiment will be described, focusing on the differences from the first embodiment. FIG. 4(B) is a configuration diagram of the generator 1 according to this embodiment. In this embodiment, carrier light generated by a light source 11 is branched into two by a coupler 17, one of which is modulated by a modulator 12, and the other is input to a pulser 18. The modulator 12 modulates the carrier light with input data and outputs the modulated light to the coupler 13. The pulser 18 also generates pulsed light by periodically turning the carrier light on and off, and outputs the generated pulsed light to the coupler 13. In this embodiment, both carrier light and pulsed light are generated using a single light source 11, thereby reducing the number of light sources required.
[0023] The polarization relationship between the modulated light and the pulsed light is arbitrary. That is, the polarization of the modulated light may be parallel to the polarization of the pulsed light, or may be perpendicular to the polarization of the pulsed light, or at some other angle. When the polarizations of the modulated light and the pulsed light are made orthogonal, a polarizing beam splitter (PBS) is used instead of the coupler 13 in Figure 4(B). The PBS combines the modulated light and the pulsed light so that the polarizations are orthogonal. By making the polarizations of the modulated light and the pulsed light orthogonal, the effect of cross phase modulation in the nonlinear medium 3 is reduced compared to when the polarizations of the modulated light and the pulsed light are parallel. However, the frequency band of the signal light is still broadened due to the nonlinear optical effect in the nonlinear medium 3.
[0024] <Fourth embodiment> Next, the fourth embodiment will be described, focusing on the differences from the first embodiment. FIG. 5(A) is a configuration diagram of a generator 1 according to this embodiment. This embodiment uses modulated light that has been optically OFDM-modulated, with each of multiple subcarriers carrying input data. A calculator 20 generates a DMT (Discrete Multitone) signal based on the input data. The DMT signal is obtained by assigning complex values corresponding to data values to positive frequency subcarriers (bins), and assigning conjugate complex values of the set complex values to positive frequency bins corresponding to 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 indicates the data value. However, the amplitude value of each bin may also indicate the data value.
[0025] The calculation unit 20 outputs a DMT signal to the in-phase (I) port of the quadrature modulator (IQ modulator) 21, and outputs a signal obtained by subjecting the DMT signal to a Hilbert transform to a quadrature (Q) port. The IQ modulator 21 performs quadrature modulation (IQ modulation) based on signals input to the I and Q ports, and outputs modulated light. More specifically, the IQ modulator 21 splits the carrier light from the light source 11 into two beams and intensity-modulates each beam with the signal input to the I and Q ports. The IQ modulator 21 generates modulated light by combining the two intensity-modulated beams with a phase difference of π / 2. By performing IQ modulation based on the DMT signal and a signal obtained by subjecting the DMT signal to a Hilbert transform, the negative frequency components of the DMT signal are canceled out, resulting in OFDM-modulated light having multiple subcarriers carrying complex values set in positive frequency bins. Figure 5(B) shows the frequency components of the OFDM-modulated light. Each vertical line on the frequency axis corresponds to a subcarrier, and the deviation of the subcarrier amplitude from a reference value corresponds to a data value. Note that the configuration for generating OFDM modulated light is not limited to the configuration in Figure 5(B), and any configuration may be used.
[0026] In this embodiment, optical OFDM modulation is used, so the amount of input data can be increased. Note that, as described in the first embodiment, this embodiment can also generate signal light by multiplexing a plurality of modulated lights with one pulsed light. Also, as described in the second embodiment, this embodiment can also be configured to convert pulsed light generated by the pulsed light source 14 into pulsed light with a wider bandwidth and then multiplex it with the modulated light. Also, in this embodiment, carrier light and pulsed light can be generated based on one light source 11, as in the third embodiment.
[0027] This configuration makes it possible to increase the number of neurons in optical neural network devices, thereby contributing to Goal 9 of the United Nations' Sustainable Development Goals (SDGs), which is to "Build resilient infrastructure, promote sustainable industrialization, and foster innovation." [Explanation of symbols]
[0028] 11: light source, 14: pulsed light source, 12: modulator, 13: coupler, 2: amplifier, 3: nonlinear medium
Claims
1. a first generating means for generating carrier light; a second generating means for generating a first pulsed light; a modulation means for modulating the carrier light based on input data and outputting modulated light; a combining means for combining the modulated light and the first pulsed light and outputting a signal light; an amplifier for amplifying the signal light; a nonlinear medium that propagates the signal light amplified by the amplifier; An optical neural network device comprising:
2. 2. The optical neural network device according to claim 1, wherein the nonlinear medium includes at least one of a dispersion-flattened fiber, a dispersion-shifted fiber, a dispersion-decreasing fiber, and a wire waveguide formed in a silicon substrate.
3. 2. The optical neural network device according to claim 1, further comprising delay means for imparting different propagation delays to the respective frequency components of said signal light that has propagated through said nonlinear medium.
4. 4. The optical neural network device according to claim 3, wherein the delay means includes at least one of a single mode fiber, a dispersion compensation fiber, a fiber grating, and a waveguide type dispersion compensator.
5. the first generating means generates a plurality of carrier lights having different frequencies; the modulation means modulates the plurality of carrier lights based on input data corresponding to each of the plurality of carrier lights and outputs a plurality of modulated lights; 2. The optical neural network device according to claim 1, wherein the multiplexing means multiplexes the plurality of modulated lights and the first pulsed light to output the signal light.
6. The second generating means a third generating means for generating a second pulsed light; a conversion means for converting the second pulsed light into the first pulsed light having a bandwidth wider than that of the second pulsed light; 10. The optical neural network device of claim 1, comprising:
7. 2. The optical neural network device according to claim 1, wherein the second generating means generates the first pulsed light from the carrier light.
8. 8. The optical neural network device according to claim 7, wherein the multiplexing means multiplexes the modulated light and the first pulsed light so that the polarization of the modulated light and the polarization of the first pulsed light are orthogonal to each other.
9. The optical neural network device according to claim 1 , wherein the modulated light has a plurality of subcarriers.
10. The modulation means a third generating means for generating a discrete multi-tone signal based on the input data; a quadrature modulation unit that generates the modulated light by quadrature-modulating the carrier light based on the discrete multi-tone signal; 10. The optical neural network device of claim 1, comprising:
11. the third generating means generates a converted signal by subjecting the discrete multi-tone signal to a Hilbert transform; 11. The optical neural network device according to claim 10, wherein the quadrature modulation means generates the modulated light by combining first intensity-modulated light obtained by intensity-modulating the carrier light based on the discrete multi-tone signal and second intensity-modulated light obtained by intensity-modulating the carrier light based on the converted signal.
Citation Information
Patent Citations
Photonic neural network
CN112232504A
Reinforcement learning method for coherent combination
CN112446470A
Optical fiber pulse neuron construction scheme
CN113642718A
Information processing system and information processing method
JP2020067984A
Lambda-reservoir computing
US20200311532A1