Reservoir Computing Device
By controlling propagation time differences through frequency adjustments in a reservoir computing device, the need for complex optical phase control is eliminated, allowing for a cost-effective implementation of reservoir computing devices.
Patent Information
- Application Number
- JP2022073639
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2022-04-27
- Publication Date
- 2025-12-10
- Estimated Expiration
- 2042-04-27
AI Technical Summary
The existing hardware implementation of reservoir computing devices, particularly those using optical fibers, requires complex optical phase control to adjust interference, increasing costs and complexity.
A reservoir computing device that includes an optical medium, a generating unit for generating a plurality of continuous beams of different frequencies, and a photoelectric conversion means for converting optical signals, where the propagation time difference between modulated lights is controlled by adjusting the frequency of continuous light beams, eliminating the need for complex optical phase control.
Enables the realization of a reservoir computing device at a lower cost by simplifying the optical phase control process, thereby reducing complexity and costs.
Smart Images

Figure 0007783778000001 
Figure 0007783778000002 
Figure 0007783778000003
Abstract
Description
[Technical Field]
[0001] The present invention relates to reservoir computing technology. [Background technology]
[0002] A recurrent neural network (RNN) consists of an input layer, a recurrent layer (reservoir layer), and an output layer. Generally, when an RNN is trained, the weights (coefficients) of the input layer, recurrent layer, and output layer are updated. Reservoir computing is a type of RNN that only updates the weights of the output layer during training, and does not update the weights of the input layer or recurrent layer. In reservoir computing, the weights of the input layer and recurrent layer are set to, for example, randomly generated fixed values. Instead of updating only the weights of the output layer, reservoir computing increases the number of neurons in the recurrent layer.
[0003] A hardware implementation of the recurrent layer of this reservoir computing is called "physical reservoir computing." Non-Patent Document 1 discloses physical reservoir computing that expands neurons on a time axis. For example, the input to the recurrent layer is a set of N pieces of input data (hereinafter, a set of inputs including N pieces of input data will be referred to as an "input dataset"), and this input dataset is continuously input to the recurrent layer. In this case, for example, to express random weights in the recurrent layer, a predetermined amplitude variation is applied to the N pieces of input data in the input dataset. This operation is called masking. Figure 1 is an explanatory diagram of the masking process. In Figure 1, D#pq indicates the qth input data among the N pieces of input data in the pth input dataset. The masking coefficient set consists of N masking coefficients F#1 to F#N. The input data D#pq is multiplied by the masking coefficient F#q.
[0004] The masked input data (hereinafter, masked data) is input to a feedback mechanism. The feedback mechanism is configured so that a certain masked data interferes with the masked data M after it. In other words, the feedback mechanism is configured so that M masked data can exist simultaneously within it. In the following description, this number M is also referred to as the "number of neurons in the loop." Figure 2 shows an example where N=6 and M=5. In Figure 2, the masked data obtained by processing data D#pq with masking coefficient F#q is represented as M#pq. As shown in Figure 2, masked data M#1-6 processed with masking coefficient F#6 interferes (combines) with masked data M#1-1, which was input earlier and processed with masking coefficient #1. Furthermore, masked data M#2-5 processed with masking coefficient #5 interferes with data (represented as M#ZZZ in Figure 2) generated by the interference between masked data #1-1 and masked data #6-6.
[0005] This corresponds to calculating the sum of the value of input data D#1-1 weighted by masking coefficient F#1, the value of input data D#1-6 weighted by masking coefficient F#6, and the value of input data D#2-5 weighted by masking coefficient F#5 in a neuron in a recurrent layer. In this way, by recurrently combining (interfering) data processed with different masking coefficients, a recurrent layer of an RNN in which many neurons are recurrently coupled can be realized in hardware. Note that, for example, if N and M are set to the same value, only interference occurs between input data processed with the same masking coefficient and in the same order in the input data set. For this reason, for example, N and M are set to different values.
[0006] In Non-Patent Document 1, the above feedback mechanism is realized using optical fibers. Specifically, as shown in FIG. 3, a modulator 92 modulates continuous light from a light source 91 with masking data to generate modulated light and outputs the generated modulated light to an optical fiber 94. The feedback mechanism is composed of optical fibers 94 to 96. The optical fibers 94 to 96 are interconnected at an interference region 97. A portion of the modulated light that reaches the interference region 97 from the optical fiber 94 enters the optical fiber 95, and the remaining portion enters the optical fiber 96. The modulated light that enters the optical fiber 95 propagates through the optical fiber 95 and reaches the interference region 97 again, where it interferes with the modulated light from the optical fiber 94. The modulated light that enters the optical fiber 96 is demodulated (photoelectrically converted) by a photodiode (PD) 93 and becomes the output of the recurrent layer. The length of the optical fiber 95 is adjusted so that a symbol of modulated light corresponding to a certain masking data interferes (combines) with a symbol corresponding to the masking data M times later. [Prior art documents] [Non-patent literature]
[0007] [Non-Patent Document 1] Quentin Vinckier,et.al.,"High-performance photonic reservoir computer based on a coherently driven passive cavity",Optica 2,438-446,2015 Summary of the Invention [Problem to be solved by the invention]
[0008] In the configuration of Non-Patent Document 1, it is necessary to adjust the phase of the modulated light in interference region 97 so that the phase of the modulated light arriving from optical fiber 95 and the phase of the modulated light arriving from optical fiber 94 have a predetermined relationship. The phase of the modulated light propagating through the optical fiber also changes depending on temperature, etc. For this reason, Non-Patent Document 1 discloses providing a configuration for controlling the optical phase within the feedback mechanism. However, providing a configuration for controlling the optical phase makes the control more complex and increases costs.
[0009] The present invention provides a technology for realizing a reservoir computing device at low cost. [Means for solving the problem]
[0010] According to one aspect of the present invention, a reservoir computing device includes an optical medium having different propagation times depending on the frequency, a generating unit for generating a plurality of continuous beams of different frequencies, and a generating unit for generating each of the plurality of continuous beams of different frequencies. The same at the same time a modulation means for modulating the optical signal with data to generate a plurality of modulated lights and outputting a signal light including the plurality of modulated lights to the optical medium; Before and a photoelectric conversion means for photoelectrically converting the signal light. a difference in propagation time between the first modulated light and the second modulated light in the optical medium is equal to or greater than the period of one symbol of the plurality of modulated light. . [Effects of the Invention]
[0011] The present invention allows for a reservoir computing device to be implemented at low cost. [Brief explanation of the drawings]
[0012] [Figure 1] 1 is an explanatory diagram of physical reservoir computing according to the background art. [Figure 2] 1 is an explanatory diagram of physical reservoir computing according to the background art. [Figure 3] 1 is an explanatory diagram of physical reservoir computing according to the background art. [Figure 4] FIG. 1 is a block diagram of a reservoir computing device according to one embodiment. [Figure 5] FIG. 3 is a diagram showing frequency components of light output from a light source and a modulator. [Figure 6] 10 is a diagram showing the positional relationship on the time axis of the symbols of modulated light output by a light source and the positional relationship on the time axis of the symbols of modulated light received by a PD. FIG. [Figure 7] FIG. 2 is a diagram showing an example of the configuration of a modulator. DETAILED DESCRIPTION OF THE INVENTION
[0013] 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.
[0014] FIG. 4 is a configuration diagram of a recurrent layer of a reservoir computing device according to this embodiment. A masking unit (not shown) performs masking processing on N pieces of input data in an input dataset from the input layer to generate N pieces of masked data. The masking unit sequentially outputs the masked data to the modulator 12. The light source 11 generates continuous light beams of multiple different frequencies (wavelengths) and outputs them to the modulator 12. FIG. 5(A) shows continuous light beams 21 to 23 output by the light source 11 according to this embodiment. Note that, although the number of continuous light beams is three in this embodiment, the number of continuous light beams may be any number greater than or equal to two. In this embodiment, the frequencies of the continuous light beams 21 to 23 are f1, f2, and f3, respectively. Note that the frequency of the continuous light beams is arbitrary. For example, when generating three or more continuous light beams, the frequency difference between two adjacent continuous light beams on the frequency axis does not need to be constant. In this example, the frequency difference between frequency f1 and frequency f2 is smaller than the frequency difference between frequency f2 and frequency f3. Light source 11 may include a plurality of individual light emitting elements each generating continuous light of one frequency, or may be a comb light source generating continuous light of multiple frequencies.
[0015] The modulator 12 collectively intensity-modulates the three continuous light beams 21-23 from the light source 11 with masking data. FIG. 5B shows modulated light beams 31-33 generated by the modulator 12 collectively intensity-modulating the three continuous light beams 21-23 with masking data. FIG. 6A shows the relationship of each symbol of the modulated light beams 31-33 output by the modulator 12 on the time axis. Note that a square in FIG. 6 represents one symbol, and the characters within the square representing a symbol represent masking data corresponding to that symbol. Note that the notation of the masking data is the same as in FIG. 2. As shown in FIG. 6A, the modulator 12 collectively intensity-modulates the three continuous light beams 21-23 from the light source 11 with masking data, and therefore the data carried by the symbols of the modulated light beams 31-33 at the same time timing is the same masking data.
[0016] Returning to FIG. 4, the modulator 12 outputs signal light including a plurality of modulated lights to the wavelength dispersion medium 14. The wavelength dispersion medium 14 is a medium in which the propagation time varies depending on the wavelength. For example, an optical fiber can be used as the wavelength dispersion medium 14. Alternatively, for example, an FBG (Fiber Bragg Grating) that can vary the propagation time depending on the wavelength can be used as the wavelength dispersion medium 14. Alternatively, any optical device in which the propagation time varies depending on the wavelength can be used as the wavelength dispersion medium 14.
[0017] PD13 receives the signal light propagated through the wavelength dispersion medium 14 and performs photoelectric conversion, i.e., demodulation. FIG. 6B shows the relationship on the time axis of each symbol of modulated light 31 to 33 received by PD13. Because the propagation delay of each modulated light 31 to 33 differs due to the wavelength dispersion medium 14, the positional relationship of each symbol on the time axis differs from that shown in FIG. 6A. Specifically, at the timing when PD13 receives the symbol carrying masking data M#1-1 of modulated light 33, PD13 also receives the symbol carrying masking data M#1-4 of modulated light 31 and the symbol carrying masking data M#1-3 of modulated light 32. Therefore, at this timing, PD13 outputs data obtained by adding together masking data M#1-1, M#1-3, and M#1-4 as output data of the recurrent layer.
[0018] For example, suppose a 62.5 km long optical fiber with a chromatic dispersion value of 16 ps / nm / km is used as the chromatic dispersion medium 14, and the frequency difference between the two modulated lights is 10 nm. In this case, the propagation delay difference between the two modulated lights is 10,000 ps. If the symbol rate of the two modulated lights is 10 G symbols / s, one symbol period is 100 ps. Therefore, in this case, the masking data carried by a symbol of one modulated light is added at PD 13 to the masking data carried by the symbol 100 symbols after the other modulated light.
[0019] As described above, in this embodiment, the modulated light is not combined at the optical level as in the configuration described in Non-Patent Document 1, but is combined through photoelectric conversion by the PD 13. Therefore, complex optical phase control is not required, and a reservoir computing device can be realized at low cost. Furthermore, in the configuration described in Non-Patent Document 1, the number of neurons in the loop, M, is determined by the length of the optical fiber 95. Therefore, changing the value of M required replacing the optical fiber 95 with one of a different length. In the configuration of this embodiment, the time difference between the masking data to be interfered with can be controlled by adjusting the frequency difference between the continuous light beams. In other words, by using a light source 11 that can control the frequency of each of multiple continuous light beams, the time difference between the masking data to be interfered with can be controlled by controlling the light source 11.
[0020] In FIG. 5, the modulated light 31-33 has both sidebands, i.e., double-sideband (DSB) signals. Here, in the chromatic dispersion medium 14, the upper and lower sidebands have different propagation delays, which can cause chromatic dispersion fading, which causes the signal level to fluctuate at the PD 13. Therefore, the modulated light 31-33 can be configured to have only one sideband, i.e., single-sideband (SSB) signals. For example, after generating DSB modulated light, a single-sideband signal can be generated by suppressing one sideband of each modulated light with a filter. Alternatively, a single-sideband signal can be generated directly by using an IQ modulator 122 as shown in FIG. 7. Specifically, masking data is input to the I-side terminal of the IQ modulator 122, and the masking data is Hilbert-transformed by a Hilbert transformer 121 and input to the Q-side terminal of the IQ modulator 122. By combining DSB modulated light generated by modulating continuous light based on data input to the I-side terminal and DSB modulated light generated by modulating the continuous light based on data input to the Q-side terminal, one sideband is canceled out and a single-sideband signal is generated.
[0021] Note that differences in wavelength within a single sideband cause distortion in each symbol, which causes inter-symbol interference. Therefore, to prevent inter-symbol interference, a configuration can be adopted in which the propagation delay differences within each sideband (modulated light) caused by the wavelength dispersion medium 14 are compensated for before each modulated light is incident on the wavelength dispersion medium 14. Note that because inter-symbol interference is also the combination of two temporally consecutive masking data, a configuration can also be adopted in which the signal light is incident on the wavelength dispersion medium 14 without being equalized in advance. Also, in Figure 6(B), the symbol switching timing of each modulated light is synchronized when received by the PD 13, but this is an example and the symbol switching timing of each modulated light does not need to be synchronized.
[0022] In this embodiment, multiple continuous light beams are modulated collectively in the modulator 12. However, the modulator may be configured to individually modulate each continuous light beam to generate multiple modulated light beams, and then wavelength-multiplex the multiple modulated light beams to generate signal light including the multiple modulated light beams. Although individually modulating each continuous light beam may result in the timing of symbol switching of each modulated light beam not being synchronized, as shown in FIG. 6(A), the time relationship of the symbols of each modulated light beam when output from the modulator is arbitrary in the present invention.
[0023] The above configuration makes it possible to realize a reservoir computing device at low cost, which will contribute 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]
[0024] 11: light source, 12: modulator, 14: wavelength dispersion medium, 13: PD
Claims
1. An optical medium in which propagation times vary depending on frequency, generating means for generating a plurality of continuous lights of different frequencies; a modulation means for generating a plurality of modulated lights including a first modulated light and a second modulated light by modulating each of the plurality of continuous light beams with the same data at the same timing, and outputting a signal light including the plurality of modulated lights to the optical medium; a photoelectric conversion means for photoelectrically converting the signal light received via the optical medium; Equipped with A reservoir computing device, wherein a difference in propagation time between the first modulated light and the second modulated light in the optical medium is equal to or greater than the period of one symbol of the plurality of modulated light.
2. The reservoir computing device of claim 1 , wherein the plurality of modulated lights have an upper sideband and a lower sideband.
3. The reservoir computing device of claim 1 , wherein the plurality of modulated lights have only one sideband among an upper sideband and a lower sideband.
4. 4. The reservoir computing device of claim 3, further comprising compensation means for compensating for propagation time differences within said one sideband imparted by said optical medium.
5. The reservoir computing device of claim 1 , wherein the generating means is configured to be able to control the frequencies of the plurality of continuous lights.
6. The reservoir computing device of claim 1 , wherein the optical medium comprises an optical fiber.
7. The reservoir computing device of claim 1 , wherein the optical medium comprises a fiber Bragg grating.
Citation Information
Patent Citations
Optical transmission and reception system
WO2020261401A1