Optical Neural Network Device

Abstract

To provide an optical neural network apparatus which can provide diversification of integration among neurons.SOLUTION: An optical neural network apparatus has modulation means for modulating carrier light with input data to output modulated light, converting and coupling means through which the modulated light passes, and separating means for spatially separating the modulated light passing through the converting and coupling means. The converting and coupling means has coupling means for causing coupling between a plurality of propagation modes of the modulated light and converting means for converting polarization of at least one propagation mode of the plurality of propagation modes.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to an optical neural network device that implements a neural network in hardware. [Background technology]

[0002] Non-Patent Document 1 discloses an optical neural network device. According to Non-Patent Document 1, phase-modulated light, which is phase-modulated with input data, is propagated in multiple propagation modes in a multi-mode waveguide. Coupling between the propagation modes that occurs in the multi-mode waveguide realizes coupling between neurons. Furthermore, since the propagation time in the multi-mode waveguide differs depending on the propagation mode, memory is realized by coupling past and present neurons. [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] Satoshi Sunada,et.al,"Photonic neural field on a silicon chip: Large-scale, high-speed neuro-inspired computing and sensing," Optica Vol.8,Issue 11,pp.1388-1396,2021 Summary of the Invention [Problem to be solved by the invention]

[0004] To improve the performance of processing according to the output of an optical neural network device, for example, inference processing, it is important to diversify the coupling between neurons. Here, in a multi-mode waveguide, due to its structure, coupling between propagation modes of the same polarization is likely to occur, but coupling between propagation modes of orthogonal polarization is unlikely to occur. In other words, coupling between TE (Transverse Electric) modes and between TM (Transverse Magnetic) modes is likely to occur, but coupling between TE and TM modes is unlikely to occur. Therefore, the configuration described in Non-Patent Document 1 limits the diversification of coupling between neurons.

[0005] The present disclosure provides an optical neural network device that can diversify connections between neurons. [Means for solving the problem]

[0006] According to one aspect of the present disclosure, an optical neural network device comprises a modulation means for modulating carrier light with input data and outputting modulated light, a conversion / combination means through which the modulated light passes, and a separation means for spatially separating the modulated light that has passed through the conversion / combination means, wherein the conversion / combination means comprises a combination means for generating coupling between multiple propagation modes of the modulated light, and a conversion means for converting the polarization of at least one of the multiple propagation modes. [Effects of the Invention]

[0007] According to the present disclosure, the connections between neurons can be diversified. [Brief explanation of the drawings]

[0008] [Figure 1] FIG. 1 is a block diagram of an optical neural network device according to one embodiment. [Figure 2] FIG. 2 is a block diagram of a transform combiner according to one embodiment. [Figure 3] FIG. 2 is a block diagram of a conversion unit according to one embodiment. [Figure 4]FIG. 1 is a block diagram of an optical neural network device according to one 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 neural network device (hereinafter referred to as an optical NN device) according to this embodiment. Input data to the optical NN device is input to a modulator 2. The modulator 2 modulates carrier light from a light source 1 with the input data and outputs modulated light. The carrier light may be continuous light or a signal that is turned on and off in time, such as pulsed light. The modulation method used by the modulator 2 is arbitrary, and may be, for example, intensity modulation or angle modulation. Angle modulation includes phase modulation or frequency modulation. The modulator 2 outputs the modulated light to a converting and coupling unit 3.

[0011] As shown in Fig. 2, the converting and combining unit 3 according to this embodiment includes a converting unit 9 and two combining units 8. One of the two combining units 8 is disposed upstream of the converting unit 9, and the other is disposed downstream. Note that the configuration shown in Fig. 2 is merely an example, and a configuration may also be adopted in which multiple sets of a converting unit 9 and a combining unit 8 connected downstream of the converting unit 9 are connected in series downstream of the upstream combining unit 8 in Fig. 2.

[0012] The coupling section 8 mainly generates coupling between different propagation modes of the same polarization. The coupling section 8 can be, for example, a multimode waveguide including a curved region, as described in Non-Patent Document 1. The curved region of a multimode waveguide facilitates coupling between propagation modes, thereby realizing coupling between neurons. The configuration of the coupling section 8 is arbitrary as long as coupling between different propagation modes of the same polarization occurs. For example, providing a curved region facilitates coupling between propagation modes. However, if coupling between propagation modes occurs without the curved region, the curved region is not necessary. For example, using a waveguide with a grating structure facilitates coupling between different propagation modes of the same polarization. In this case, a grating structure with an irregular grating period can be used.

[0013] The converter 9 converts between propagation modes with different polarizations, such as from TM mode to TE mode or from TE mode to TM mode. FIG. 3 shows an example of the configuration of the converter 9. The converter 9 has a substrate 91 made of silicon dioxide (SiO2) and a waveguide 92 made of silicon (Si) on the substrate 91. The substrate 91 functions as a lower cladding. The upper cladding may be air. In other words, the refractive index of the cladding around the waveguide 92 is not uniform.

[0014] FIG. 3A is a top view of the conversion unit 9, in which light propagates, for example, from the first port 921 to the second port 922, i.e., along the −Y direction in FIG. 3A. FIG. 3B is a view from above in FIG. 3A along the X direction, and FIG. 3C is a view from the right side in FIG. 3A along the Y direction. As shown in FIG. 3A, the waveguide 92 has a tapered region 923, and the width (length) of the waveguide 92 in the direction perpendicular to the light propagation direction (−Y direction) and the normal direction of the substrate 91 (X direction) gradually increases along the light propagation direction. As shown in FIG. 3A, the width of the waveguide 92 at the first port 921 of the conversion unit 9 is A, while the width at the second port 922 is B, which is larger than A.

[0015] Since the refractive indexes of the tapered region 923 and the cladding surrounding the waveguide 92 are not constant, propagation mode conversion accompanied by polarization conversion is likely to occur. Specifically, for example, if A is 2.5 μm and B is 4 μm, light in TM0, TM1, TM2, TM3, TE6, and TE7 modes input to the input port 921 is converted to light in TE6, TE8, TE9, TE10, TE7, and TM0 modes, respectively, and output from the output port 922. Note that how the propagation mode is converted differs depending on the values ​​of A and B.

[0016] 3, the light input port is the first port 921 and the light output port is the second port 922, but the light input port can also be the second port 922 and the light output port can also be the first port 921. In this case, if A is 2.5 μm and B is 4 μm, light in TE6, TE8, TE9, TE10, TE7, and TM0 modes input to the second port 921 is converted to light in TM0, TM1, TM2, TM3, TE6, and TE7 modes, respectively, and is output from the first port 921.

[0017] As shown in FIG. 2 , modulated light from the modulator 2 is first input to the upstream coupling unit 8. The upstream coupling unit 8 may be configured to input modulated light in TE mode, modulated light in TM mode, or modulated light in both TE and TM modes. In the upstream coupling unit 8, higher-order propagation modes are induced in addition to the fundamental mode, and coupling between propagation modes occurs. However, coupling between propagation modes occurs only between the same polarization, i.e., between TE modes or between TM modes, and coupling between different polarizations is unlikely. For this reason, in this embodiment, a converter 9 is provided downstream of the coupling unit 8. The converter 9 converts the propagation mode, as described above, by converting the polarization. Another coupling unit 8 is then provided downstream of the converter 9. As a result, in the downstream coupling unit 8, coupling occurs between modulated light that was in TM mode at the upstream coupling unit 8 and modulated light that was in TE mode, and between modulated light that was in TE mode at the upstream coupling unit 8 and modulated light that was in TM mode at the upstream coupling unit 8. The above configuration makes it possible to diversify the connections between neurons.

[0018] Returning to FIG. 1, the separator 4 spatially divides the light output by the converter / combiner 4 and outputs modulated light at each spatial position. Each of the multiple modulated lights output by the separator 4 corresponds to a neuron at the most downstream side of the optical NN device. Each modulated light output by the separator 4 is photoelectrically converted and weighted and added in the processor 5. The processor 5 outputs the signal after weighted addition as output data of the optical NN device. The output data of the optical NN device is used for applications of the optical NN device, such as inference.

[0019] In order to efficiently perform polarization conversion in the conversion unit 9, in this embodiment, the refractive index of the clad around the waveguide 92 is made different. Specifically, in this embodiment, the lower clad below the waveguide 92 is made of SiO2, and the upper clad above the waveguide 92 is made of air, thereby making the refractive index of the clad around the waveguide 92 different. However, instead of making the refractive index different between the top and bottom of the waveguide 92, it is also possible to make the refractive index different between the left and right. Furthermore, instead of making the refractive index of the clad around the waveguide 92 constant, by adjusting the shape of the waveguide 92, it is possible to achieve a configuration that efficiently performs polarization conversion in the conversion unit 9. Specifically, by making the cross-sectional shape of the waveguide 92 orthogonal to the propagation direction of light asymmetric with respect to a line in the width direction of the waveguide 92 or a line normal to the substrate 91, it is possible to efficiently perform polarization conversion.

[0020] In this embodiment, the converter 9 is configured using a tapered waveguide 92. However, as long as conversion between propagation modes of different polarizations is performed, the converter 9 is not limited to a tapered type. For example, a periodic asymmetric rib waveguide, a sloped cross-section waveguide, a mode evolution type, an RIE-lag type, a periodic sloped slot type, a half-ridge type, an antisymmetric grating type, or the like can be used. Furthermore, by adjusting the shape of the waveguide 92, it is possible to configure the waveguide 92 to perform both coupling between propagation modes and polarization conversion. Furthermore, when multiple sets of converters 9 and coupling units 8 are provided downstream of the upstream coupling unit 8 in FIG. 2, the configuration of the converters 9 can be different for each set. For example, the width A of the first port 921 and the width B of the second port 922 can be different for each set, or a tapered converter 9 can be used in one set and an antisymmetric grating type can be used in another set.

[0021] Second Embodiment Next, the second embodiment will be described, focusing on the differences from the first embodiment. Figure 4 is a configuration diagram of the optical NN device according to this embodiment. In this embodiment, a polarization splitter 6 corresponding to each of the multiple modulated light beams output by the splitter 4 is provided between the splitter 4 and the processor 5. The polarization splitter 6 splits the input modulated light into two orthogonal modulated light beams and outputs each beam individually to the processor 5. With this configuration, the number of neurons at the output terminal is double that of the first embodiment, thereby improving the performance of processing performed using the output of the optical NN device.

[0022] The above configuration makes it possible to diversify the connections between 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]

[0023] 2: Modulator, 3: Conversion and Combination Section, 4: Separation Section

Claims

1. a modulation means for modulating a carrier light with input data and outputting modulated light; a conversion / combination means through which the modulated light passes; a separating means for spatially separating the modulated light that has passed through the converting and combining means; Equipped with The conversion and combination means a coupling means for generating coupling between a plurality of propagation modes of the modulated light; a conversion means for converting the polarization of at least one of the plurality of propagation modes; An optical neural network device comprising:

2. 2. The optical neural network device according to claim 1, wherein the conversion means has a waveguide whose input port has a width different from that of its output port.

3. 3. The optical neural network device according to claim 2, wherein the waveguide includes a tapered region in which the width of the waveguide changes from the input port to the output port.

4. 3. The optical neural network device of claim 2, wherein the refractive index around the waveguide is not constant.

5. 3. The optical neural network device of claim 2, wherein the waveguide is provided on a substrate, and a cross section of the waveguide perpendicular to the light propagation direction in the waveguide is not symmetrical with respect to a line normal to the substrate, or is not symmetrical with respect to a line perpendicular to both the propagation direction and the normal direction.

6. 2. The optical neural network device according to claim 1, further comprising polarization separation means for separating, by polarization, each of the spatially separated modulated lights output by said separation means.

7. 2. The optical neural network device according to claim 1, further comprising processing means for photoelectrically converting and weighting each of the spatially separated modulated lights output by said separating means and adding them together.

8. 8. The optical neural network device according to claim 1, wherein the conversion / combining means has the combining means on both the upstream and downstream sides of the conversion means.

9. 9. The optical neural network device according to claim 8, wherein the conversion means converts from TE mode to TM mode or from TM mode to TE mode.

10. 10. The optical neural network device according to claim 9, wherein the coupling means generates coupling between a plurality of propagation modes in the TE mode and coupling between a plurality of propagation modes in the TM mode.

Images