Optical neural network

Abstract

An optical neural network 1 comprises an input light generation unit 12 that generates an optical signal that has been modulated in accordance with an input d(n), a multicore optical fiber 11 that receives the optical signal as input and functions to convert the optical signal to a high-dimensional feature space that reflects the features of the input d(n), and a light reception unit 15 that receives an optical signal from the multicore optical fiber 11 and converts the optical signal to an electrical signal. The multicore optical fiber 11 has a plurality of cores 111, and the distance between nearest neighboring cores is a distance that promotes random coupling of optical signals.

Description

Optical neural network

[0001] The present disclosure relates to optical neural networks.

[0002] Optical calculation using optical elements is advantageous compared with electronic systems from the viewpoints of thermal efficiency, high speed, and low power consumption. Due to these advantages, research and study on optical calculation utilizing optical elements are being actively conducted. Non-Patent Document 1 reports an example in which a neuron learning layer of an optical neural network is formed by a multi-core optical fiber (MCF).

[0003] Eyal Cohen et al., “Neural networks within multi-core optic fibers,” Scientific reports 6.1 (2016)Hayashi, Tetsuya, et al. "Randomly-coupled multi-core fiber technology." Proceedings of the IEEE 110.11 (2022)Smith, R. G. et al., “Optical power handling capacity of low loss optical fibers as determined by stimulated Raman and Brillouin scattering,” Applied optics, 11(11), 2489-2494, (1972)

[0004] In the MCF disclosed in Non-Patent Document 1, the inter-core distance is approximately 9 μm, and light propagates while being linearly coupled as a supermode. However, due to the large core diameter and large number of cores, there are problems in connectivity and manufacturability. As a result, there have been problems that the manufacturing cost increases and the device configuration such as signal input / output connection becomes complicated.

[0005] The present disclosure has been made in view of the above, and an object of the present disclosure is to improve the connectivity and manufacturability of multi-core optical fibers used for optical neural networks.

[0006] An optical neural network in one aspect of the present disclosure comprises an input light generation unit that generates an optical signal modulated according to an input signal, an intermediate layer that receives the optical signal and converts it into a high-dimensional feature space that reflects the characteristics of the input signal, and a light receiving unit that receives the optical signal from the intermediate layer and converts it into an electrical signal, wherein the intermediate layer comprises a multicore optical fiber having a plurality of cores, and the distance between the nearest adjacent cores is a distance that promotes random coupling of the optical signals.

[0007] According to this disclosure, the connectivity and manufacturability of multicore optical fibers used in optical neural networks can be improved.

[0008] Figure 1 shows an example of the configuration of the optical neural network in Example 1. Figure 2 is a cross-sectional view showing an example of a radial cross-section of a multicore optical fiber. Figure 3 shows an example of the configuration of the optical neural network in Example 2. Figure 4 is a cross-sectional view showing an example of a radial cross-section of a multicore optical fiber. Figure 5 shows an example of the configuration of the optical neural network in Example 3. Figure 6 is a side view showing an example of the configuration of a multicore optical fiber. Figure 7 shows an example of the configuration of the optical neural network in Example 4. Figure 8 shows an example of the configuration of the optical neural network in Example 5. Figure 9 shows an example of the configuration of the optical neural network in Example 6. Figure 10 is a cross-sectional view showing an example of a radial cross-section of a multicore optical fiber. Figure 11 shows an example of the configuration of the optical neural network in Example 7.

[0009] [Example 1] Referring to Figure 1, an example of the configuration of the optical neural network 1 in Example 1 will be described. The optical neural network 1 shown in the figure comprises an input light generation unit 12, a fan-in unit 13, a randomly coupled MCF 11, a fan-out unit 14, and a light receiving unit 15. The fan-in unit 13, MCF 11, and fan-out unit 14 may be collectively referred to as the intermediate layer.

[0010] The input light generation unit 12 generates an optical signal modulated according to the input d(n). Modulation methods include intensity modulation, phase modulation, and frequency modulation.

[0011] The generated optical signals are input to each core of the MCF 11 via the fan-in 13. The fan-in 13 plays a role in appropriately coupling the optical signals from multiple single-core optical fibers to each core of the MCF 11.

[0012] As shown in the cross-sectional view of Figure 2, the MCF 11 has multiple cores 111 arranged within a cladding 112. The distance Λ between the nearest adjacent cores is set to 15 μm or more and 25 μm or less. By setting the inter-core distance Λ within this range, random coupling of optical signals is promoted, the interaction between neurons in the MCF 11 becomes more complex, and a reduction in the amount of computation required for learning is expected. With the reduction in computation, the number of required cores decreases, and improvements in connectivity and manufacturability can be expected. Non-patent document 2 describes that random coupling is promoted when the inter-core distance is 20 μm.

[0013] Furthermore, to promote random coupling, the optical fiber may be twisted in the longitudinal direction. The length of the MCF 11 can be set arbitrarily.

[0014] The optical signals that have passed through each core of the MCF 11 are input to the light receiving unit 15 via the fan-out 14. The fan-out 14 has the role of separating the output optical signals from each core of the MCF 11 into multiple single-core optical fibers.

[0015] The light receiving unit 15 converts the optical signals output from each core into electrical signals and outputs the result y(n).

[0016] During the learning process, for example, the weights for the electrical signals converted from the optical signals by the light receiving unit 15 may be learned and optimized so that the output y(n) approaches the target value.

[0017] In optical neural networks using randomly coupled MCF, the features of the input d(n) are transformed into a high-dimensional feature space that reflects the characteristics of the input d(n) through inter-core connections and interference effects that occur between multiple cores.

[0018] [Example 2] Referring to Figure 3, an example of the configuration of the optical neural network 1 in Example 2 will be described. The optical neural network 1 shown in the figure comprises an input light generation unit 12, an excitation light generation unit 16, a fan-in unit 13, an MCF 11, a fan-out unit 14, and a light receiving unit 15. Example 2 differs from Example 1 in that some of the cores of the MCF 11 are rare-earth doped cores. The description of the same configuration as Example 1 will be omitted.

[0019] As shown in the cross-sectional view of Figure 3, the MCF 11 is composed of randomly coupled MCFs in which multiple cores 111, including rare-earth-doped cores 113, are arranged within a cladding 112. Excitation light excites the rare-earth elements, and the intensity of the optical signal propagating through the rare-earth-doped cores 113 is amplified by stimulated emission. If the total number of cores in the MCF 11 is M and the number of rare-earth-doped cores is P, then it is sufficient that M > P.

[0020] The excitation light generation unit 16 emits excitation light. The excitation light is combined with an optical signal and incident on the rare-earth-doped core. In the case of cladding excitation, the excitation light is incident on the cladding.

[0021] By making a rare-earth doped core part of the MCF11 core, the optical signal input to the rare-earth doped core of the MCF11 propagates while being amplified by the excitation light, expanding the randomness of the coupling, and thus making the intermediate layer more complex.

[0022] [Example 3] Referring to Figure 5, an example of the configuration of the optical neural network 1 in Example 3 will be described. The optical neural network 1 shown in the figure comprises an input light generation unit 12, an excitation light generation unit 16, a fan-in unit 13, an MCF 11, a fan-out unit 14, and a light receiving unit 15. Example 3 differs from Example 1 in that, like Example 2, a part of the core of the MCF 11 is a rare-earth doped core. The description of the same configuration as Example 1 will be omitted.

[0023] As shown in the side view of Figure 6, the MCF 11 has a configuration in which an MCF 11A having a core without rare earth elements (a normal core) 111 and an MCF 11B having a core with rare earth elements added 113 are connected. In other words, the MCF 11 of Example 3 can be said to have a rare earth element added to a part of the longitudinal direction of a randomly bonded MCF. In Figure 6, the MCF 11B is connected between the MCF 11A, but this is not the only configuration, and the MCF 11A and MCF 11B may be connected alternately in a vertical column.

[0024] The excitation light generation unit 16 emits excitation light. As shown in Figure 5, the excitation light may be input from the side of the MCF 11B. For example, an optical waveguide may be formed in the cladding region of the MCF 11B, and the excitation light may be input from the side of the MCF 11B. Inputting the excitation light from the side of the MCF 11B is expected to suppress connection loss between the MCF 11A and the MCF 11B. Alternatively, the excitation light and the optical signal may be combined and input to each core from the end face of the MCF 11.

[0025] The MCF11s from Example 2 and Example 3 may be combined.

[0026] [Example 4] Referring to Figure 7, an example of the configuration of the optical neural network 1 in Example 4 will be described. The optical neural network 1 shown in the figure comprises an input light generation unit 12, a fan-in 13, an MCF 11, a fan-out 14, a light receiving unit 15, an optical power coupler 17, and a single-mode optical fiber (SMF) 18. In Example 4, a portion of the optical signal from the intermediate layer is looped back to the input side of the intermediate layer. Any of the MCFs 11 from Examples 1 to 3 may be used for the MCF 11. The description of the same configuration as in Example 1 will be omitted.

[0027] A portion of the optical signal is branched from at least one channel of the fan-out 14 using an optical power coupler 17 and looped back to the input side via the SMF 18. On the input side, the branched optical signal is combined with the channel of the fan-in 13 using the optical power coupler 17. This makes it possible to extend the network into a recurrent neural network (feedback neural network) that is effective for learning time-series data.

[0028] All channels of optical signals may be looped back to the input side. The destination to which the branched optical signals are looped back may be a different channel from the channel from which a portion of the optical signal was branched. The SMF18 used for loopback may be a multicore optical fiber.

[0029] [Example 5] Referring to Figure 8, an example of the configuration of the optical neural network 1 in Example 5 will be described. The optical neural network 1 shown in the figure comprises an input light generation unit 12, a fan-in unit 13, an MCF 11, a fan-out unit 14, a light receiving unit 15, and an SMF 18. In Example 5, a portion of the optical signal from the intermediate layer is looped back to the input side of the intermediate layer. Any of the MCF 11s from Examples 1 to 3 may be used for the MCF 11. The description of the same configuration as in Example 1 will be omitted.

[0030] The optical signal is looped back to the input side via the SMF 18 from at least one channel of the fan-out 14. On the input side, the looped-back optical signal is input to the channel of the fan-in 13. This makes it possible to extend to a recurrent neural network while suppressing the loss to the optical signal compared to when an optical power coupler is used.

[0031] The destination channel to which the optical signal is returned via loopback may be a different channel from the original channel. The SMF18 used for loopback may also be a multicore optical fiber.

[0032] [Example 6] Referring to Figure 9, an example of the configuration of the optical neural network 1 in Example 6 will be described. The optical neural network 1 shown in the figure comprises an input light generation unit 12, a fan-in unit 13, an MCF 11, a fan-out unit 14, a light receiving unit 15, and an MCF 19 for loopback. In Example 6, a portion of the optical signal from the MCF 11 is looped back to the input side of the MCF 11. A description of the same configuration as in Example 1 will be omitted.

[0033] As shown in the cross-sectional view of Figure 10, the MCF 11 comprises a core 111 for inputting and outputting optical signals and a loopback core 114. The loopback MCF 19 is connected to the output core 114, and the other end of the MCF 19 is connected to the input core 114. This makes it possible to extend to a recurrent neural network while suppressing the loss to the optical signal compared to when an optical power coupler is used. As in Examples 2 and 3, a part of the core 111 may be a rare-earth doped core.

[0034] The MCF 11 and loop-shaped MCF 19 from Examples 1 to 3 may be bundled in parallel to each other, and configured to create a core-to-core coupling between the MCF 11 and MCF 19. The MCF 11 and MCF 19 may also be twisted together.

[0035] [Example 7] Referring to Figure 11, an example of the configuration of the optical neural network 1 in Example 7 will be described. The optical neural network 1 shown in the figure comprises an input light generation unit 12, a fan-in unit 13, an MCF 11, a fan-out unit 14, a light receiving unit 15, and an optical amplifier 20. In Example 7, the optical amplifier 20 amplifies the optical signal input to the intermediate layer. Example 7 can be applied to any of Examples 1 to 6. A description of the same configuration as in Example 1 will be omitted.

[0036] In the optical neural network 1 of Example 7, the optical amplifier 20 amplifies the optical signal, and the optical power above the threshold at which nonlinear optical effects occur is input to the MCF 11. As a result, the optical signal is modulated by nonlinear optical effects such as self-phase modulation (SPM) and stimulated Raman scattering (SRS), making it possible to make the intermediate layer more complex. Non-patent document 3 describes the threshold of optical power at which stimulated Raman scattering occurs.

[0037] As described above, the optical neural network 1 of this embodiment includes an input light generation unit 12 that generates an optical signal modulated according to the input d(n), a unit 11 that receives the optical signal and converts it into a high-dimensional feature space that reflects the features of the input d(n), and a light receiving unit 15 that receives the optical signal from the MCF 11 and converts it into an electrical signal. The MCF 11 has multiple cores 111, and the distance between the closest adjacent cores is the distance at which random coupling of optical signals is promoted. As a result, random coupling of optical signals is promoted, the interaction between neurons in the hidden layer becomes more complex, and a reduction in the amount of computation required for learning is expected. Consequently, the number of required cores decreases, and improvements in connectivity and manufacturability can be expected.

[0038] Furthermore, by making some of the multiple cores rare-earth doped cores, the optical signal input to the rare-earth doped core propagates while being amplified by the excitation light, expanding the randomness of the coupling, which makes it possible to make the intermediate layer more complex.

[0039] Furthermore, by looping the optical signal back from the output side to the input side of the intermediate layer, it becomes possible to extend the optical neural network 1 into a recurrent neural network (feedback neural network) that is effective for learning time-series data.

[0040] Furthermore, by providing an optical amplifier 20 that amplifies the intensity of the optical signal so that a nonlinear optical effect occurs in the MCF 11, it becomes possible to make the intermediate layer more complex.

[0041] 1 Optical neural network 11 Multicore optical fiber 111 Core 112 Cladding 113 Rare earth doped core 12 Input light generation unit 13 Fan-in 14 Fan-out 15 Light receiving unit 16 Excitation light generation unit 17 Optical power coupler 18 Single-mode optical fiber 19 Multicore optical fiber 20 Optical amplifier

Claims

1. An optical neural network comprising: an input light generation unit that generates an optical signal modulated according to an input signal; an intermediate layer that receives the optical signal and converts it into a high-dimensional feature space that reflects the characteristics of the input signal; and a light receiving unit that receives the optical signal from the intermediate layer and converts it into an electrical signal, wherein the intermediate layer comprises a multicore optical fiber having multiple cores, and the distance between the nearest adjacent cores is a distance that promotes random coupling of the optical signals.

2. An optical neural network according to claim 1, wherein some of the plurality of cores are rare-earth doped cores.

3. An optical neural network according to claim 1 or 2, wherein an optical signal is looped back from the output side to the input side of the intermediate layer.

4. An optical neural network according to claim 1, comprising an optical amplifier that amplifies the intensity of the optical signal such that a nonlinear optical effect occurs in the intermediate layer.

5. The optical neural network according to claim 2, wherein at least one of the plurality of cores is a rare-earth-doped core.

6. An optical neural network according to claim 2, wherein a portion of the longitudinal direction of the plurality of cores is a rare earth-doped core.

7. An optical neural network according to claim 3, wherein at least a portion of any optical signal is branched from the output side of the intermediate layer and looped back to the input side.

8. An optical neural network according to claim 3, wherein at least one of the cores on the output side of the intermediate layer is connected to the core on the input side.

Images