JP7601818B2 - Speckle generating circuit and optical neural network device - Google Patents

Description

The present invention relates to a speckle generation circuit and an optical neural network device used in the field of artificial intelligence.

Artificial intelligence is a research field that aims to achieve advanced information processing approaching real intelligence by using computational models that mimic the neural circuits of the brain, and various research projects have been conducted in recent years. In artificial intelligence, computational performance can be improved by increasing the number of neurons used in calculations. However, because typical artificial intelligence models are realized by electrical calculations performed by computer programs, increasing the number of neurons poses the issue of increased power required for calculations and processing delays. To address this issue, there is a technology called "optical reservoir computer" that performs calculations using optical interference instead of electrical calculations. Optical reservoir computers enable information processing by weighting multiple output signals from optical circuits and performing addition calculations.

Non-Patent Document 1 discloses technology relating to an optical neural network device (optical reservoir computer) (Fig. 11A) that includes components including a single-wavelength light source, an optical phase modulator that phase-modulates the output light from the light source, a single-mode waveguide through which the output light from the optical phase modulator propagates, a multi-mode waveguide that is connected to the single-mode waveguide and has a curved region, a number of light-receiving elements that photoelectrically convert the output light from the multi-mode waveguide, a function for weighting the electrical signals from the multiple light-receiving elements, and a mechanism for adding the weighted electrical signals.

The operation of the optical neural network device is as follows: the data u(n), which is a time-dependent signal, is applied to a phase modulator to perform phase modulation, and the modulated light is then propagated through a single-mode waveguide. The modulated light is then converted into a number of spatial mode lights in a multi-mode waveguide having a curved region that functions to excite higher spatial modes, and a time-varying speckle (shade image) is generated at the end of the waveguide while providing a propagation time difference between the modes (Fig. 11B). The speckle obtained at the end of the waveguide also changes depending on the deviation of the center of each waveguide at the connection part between the single-mode waveguide and the multi-mode waveguide. In addition, in the multi-mode waveguide, a time-varying mixing is generated at the output end of the multi-mode waveguide by utilizing the difference in time that occurs when the higher spatial mode and the lower spatial mode propagate through the transmission path, thereby acting as a substitute for temporal memory. The speckle is converted into an electrical signal by a light-receiving element, and each electrical signal is multiplied by its respective weight, and then the output (inferred value) is obtained by summing all of them. Each weight is set so that it closely matches the target data y(n) during the learning phase.

In such optical neural network devices, the number of spatial modes in the multimode waveguide corresponds to the number of neurons in artificial intelligence, and increasing the number of neurons is important for improving inference performance.

In Non-Patent Documents 1 and 2, the number of spatial modes calculated from the structure of the multimode waveguide was 68, but because it is difficult to excite higher-order spatial modes, the number of spatial modes actually remained at 65. Therefore, as a method of increasing the number of neurons, multiple single-wavelength light sources are prepared, and after combining them, they are input to an optical phase modulator, and then split using the output from the optical phase modulator, thereby doubling the number of neurons by the number of wavelengths. In addition, the number of neurons is virtually quadrupled by time-dividing the input data at 1/4 of the time interval and performing photoelectric conversion, weighting, and summation.

Satoshi Sunada and Atsushi Uchida, "Photonic neural field on a silicon chip: large-scale, high-speed neuro-inspired computing and sensing", Optica Publishing Group, November 2021, pp. 1388-1396 Satoshi Sunada and Atsushi Uchida, “Photonic neural field on a silicon chip: large-scale, high-speed neuro-inspired computing and sensing: supplement,” Optica Publishing Group

However, in the optical neural network devices in Non-Patent Document 1 and Non-Patent Document 2, when the number of neurons is increased in order to further improve the inference performance, the number of spatial modes of the multimode waveguide cannot be further increased. In addition, although the number of neurons can be increased by increasing the number of wavelengths, this merely increases the number of neurons by picking up only the light at one measurement point of the speckle and splitting the wavelengths using a splitter element. In order to split the wavelengths at 65 measurement points, the number of wavelengths cannot be increased significantly due to the size of the splitter element placed in front of the light receiving element. Furthermore, the number of time divisions cannot be increased due to the limit of the operating speed of the electronic circuit and the reduced independence of the neurons. For these reasons, it was difficult to further increase the number of neurons.

The present invention has been made in view of the above circumstances, and aims to provide a speckle generation circuit and an optical neural network device that, when distributing a data signal into a multimode optical signal, can increase the number of neurons by branching the data signal using an optical branching circuit and converting the branched optical signals into higher-order modes using multiple multimode waveguides with different dimensions.

(1) In order to achieve the above object, the present invention takes the following measures. That is, the speckle generating circuit of the present invention is a speckle generating circuit used in an optical neural network device, and is characterized in that it includes a phase modulator that receives an optical signal from a light source and performs phase modulation, a single single-mode waveguide that receives and propagates the phase-modulated modulated light, an optical branching element that is connected to the single-mode waveguide and branches the propagated modulated light into multiple light beams, multiple single-mode waveguides that propagate each of the branched modulated lights, and multiple multi-mode waveguides that are connected to each of the branched single-mode waveguides and generate speckles from the propagated modulated light.

(2) In the speckle generating circuit of the present invention, each of the multimode waveguides has a different waveguide width.

(3) In the speckle generating circuit of the present invention, each of the multimode waveguides has a different thickness.

(4) In the speckle generating circuit of the present invention, each of the multimode waveguides has a different waveguide length.

(5) In the speckle generating circuit of the present invention, the multi-mode waveguides are connected at positions where the distance between the central axis of each branched single-mode waveguide and the central axis of each multi-mode waveguide connected to each branched single-mode waveguide is the same.

(6) In the speckle generating circuit of the present invention, the multi-mode waveguides are connected at positions where the distances between the central axis of each branched single-mode waveguide and the central axis of each multi-mode waveguide connected to each branched single-mode waveguide are all different.

(7) In the speckle generating circuit of the present invention, the optical branching element separates the input modulated light into TE mode (transverse electric modes) and TM mode (transverse magnetic modes), and outputs different signals to each of the branched single-mode waveguides.

(8) The optical neural network device of the present invention is an optical neural network device that generates a large number of neurons, and is characterized in that it includes a speckle generating circuit described in any one of (1) to (7) and a plurality of light receiving elements that photoelectrically convert the speckles generated by the speckle generating circuit into an electrical signal.

(9) In the optical neural network device of the present invention, the number of the light receiving elements corresponds to the number of modes output from the multimode waveguide to which each of the light receiving elements is connected.

According to the present invention, by increasing the number of neurons in an optical neural network device, it is possible to improve the inference performance of artificial intelligence.

1 is a diagram showing a schematic configuration of an optical neural network device according to an embodiment of the present invention; 2 is a cross-sectional view taken along line s-s in FIG. 2 is a cross-sectional view taken along line mm in FIG. FIG. 1 is a diagram showing locations corresponding to the length of a multimode waveguide (the length at the center of the waveguide). FIG. 1 is a diagram showing a schematic configuration of a multi-mode waveguide used in a simulation. FIG. 4 is a diagram showing the light intensity output from a multimode waveguide. FIG. 2 is a diagram showing a schematic configuration of a connection portion of a single-mode waveguide and a multi-mode waveguide. FIG. 11 is a diagram showing a schematic configuration of the length of a multi-mode waveguide in Modification 2. 13 is a diagram showing a connection portion of a single-mode waveguide and a multi-mode waveguide in Modification 3. FIG. FIG. 13 is a diagram showing a schematic configuration of an optical neural network system according to a fifth modified example. FIG. 1 illustrates a conventional optical neural network device. FIG. 1 is a diagram showing speckles generated by a conventional optical neural network device.

The inventors focused on the fact that the number of neurons cannot be increased due to the limit on the number of excitations of spatial modes in a multimode waveguide, and discovered that when distributing to a multimode optical signal, it is possible to generate different speckles and increase the number of neurons by branching the data signal using an optical branching circuit and converting the branched optical signals to higher modes using multiple multimode waveguides with different dimensions, which led to the invention.

That is, the present invention is a speckle generation circuit for use in an optical neural network device, characterized in that it comprises: a phase modulator that receives an optical signal from a light source and performs phase modulation; a single single-mode waveguide that receives and propagates the phase-modulated modulated light; an optical branching element that is connected to the single-mode waveguide and branches the propagated modulated light into multiple light beams; multiple single-mode waveguides that propagate each of the branched modulated lights; and multiple multi-mode waveguides that are respectively connected to each of the branched single-mode waveguides and generate speckles from the propagated modulated light.

This has enabled the inventors to significantly increase the number of neurons required for an optical reservoir computer. Below, an embodiment of the present invention will be specifically described with reference to the drawings.

Figure 1 is a diagram showing a schematic configuration of an optical neural network device according to this embodiment. The optical neural network device 1 according to this embodiment is composed of a single-wavelength light source 111, a phase modulator 113, an optical branching element 117, a first single-mode waveguide 121, a second single-mode waveguide 123, a third single-mode waveguide 223, a first multimode waveguide 125, a second multimode waveguide 225, a first light-receiving element 119, a second light-receiving element 219, an A/D converter 131, and a digital circuit 133 consisting of a weighting electrical circuit, an electrical circuit for adding weighted electrical signals, and a D/A converter.

In this embodiment, the optical branching element 117, the first single-mode waveguide 121, the second single-mode waveguide 123, the third single-mode waveguide 223, the first multimode waveguide 125, and the second multimode waveguide 225 are formed on one silicon substrate (chip) 310, but this is not limited to this. For example, the circuits branched into multiple circuits by the branching element may be formed on the silicon substrate (chip) for each branched circuit.

Laser light is output from the single-wavelength light source 111, and data u(n), which is a time-dependent signal, is applied to the phase modulator 113 to phase-modulate the output light from the single-wavelength light source 111. In this embodiment, the phase modulator 113 uses a lithium niobate phase modulator.

The first single mode waveguide 121 is connected to the phase modulator 113 and propagates modulated light from the phase modulator 113. The first single mode waveguide 121 is then branched into two, the second single mode waveguide 123 and the third single mode waveguide 223, in the optical branching element 117 of the multimode interference waveguide. As an example, the branching element of this embodiment uses an element with two branches, but an element with four branches, eight branches, or more branches may also be used. In addition, the input polarization to the second single mode waveguide 123 and the third single mode waveguide 223 is set to TE mode (transverse electric modes), and a branching element of the multimode interference waveguide is applied to the optical branching element 117.

The second single-mode waveguide 123 and the third single-mode waveguide 223 branched by the optical branching element 117 are connected to the first multi-mode waveguide 125 and the second multi-mode waveguide 225, respectively.

Figure 2 is a cross-sectional view taken along line s-s in Figure 1, showing the cross sections of the second single-mode waveguide and the third single-mode waveguide. Figure 3 is a cross-sectional view taken along line m-m in Figure 1, showing the cross sections of the first multimode waveguide and the second multimode waveguide. Figure 4 is a diagram showing the location corresponding to the length of the multimode waveguide (the length of the center of the waveguide).

The first single mode waveguide 121, the second single mode waveguide 123, the third single mode waveguide 223, the first multimode waveguide 125 and the second multimode waveguide 225 are formed on a silicon substrate (chip) 310, the cores of the waveguides 121, 123, 125, 223 and 225 are made of silicon, and the clads of the waveguides 121, 123, 125, 223 and 225 are made of silicon dioxide and have a thickness of 0.22 μm. The widths of the first single mode waveguide 121, the second single mode waveguide 123 and the third single mode waveguide 223 are 0.5 μm. The first multimode waveguide 125 and the second multimode waveguide 225 have a width of 23.9 μm and 18.8 μm, respectively, and a total length of 39 mm. The first multimode waveguide 125 and the second multimode waveguide 225 each have a spiral shape with an S-shaped curved region in the center, but are not limited to this. The number of turns of the spiral shape may be one or more. It is sufficient that it is not a straight line.

Here, the speckles generated from a multimode waveguide will be described. FIG. 5 is a diagram showing a schematic configuration of a multimode waveguide used in the simulation. FIG. 5(b) is a cross-sectional view taken along a-a in FIG. 5(a). A simulation is performed to show that different speckles are generated by using multimode waveguides of different sizes. In the simulation, multimode waveguides of the following three sizes were used. All of them have a curved region with R=25 μm.
(1) Width: 25 μm, thickness: 0.25 μm, length: 230 μm
(2) Width: 25 μm, thickness: 0.22 μm, length: 230 μm
(3) Width: 25 μm, thickness: 0.22 μm, length: 280 μm

Figures 6 (a) to (c) are diagrams showing the light intensity output from the multimode waveguides (1) to (3). In Figure 6, the vertical axis indicates the light intensity [a.u.], and the horizontal axis y [μm] indicates the position in the lateral (width) direction of the waveguide when the central axis of the waveguide is the origin. As shown in Figures 6 (a) to (c), the light intensity output from the multimode waveguide differs by changing the size of the multimode waveguide. In other words, it can be seen that different speckles are generated at the output end of the multimode waveguide.

Figure 7 is a diagram showing a schematic configuration of the connection portion of the single-mode waveguide and the multi-mode waveguide according to this embodiment. As shown in Figure 7, the second single-mode waveguide 123 and the first multi-mode waveguide 125, the third single-mode waveguide 223 and the second multi-mode waveguide 225 are connected at a position where they are shifted by the same length from the center of each waveguide, that is, the distance between the center axis of the second single-mode waveguide 123 and the center axis of the first multi-mode waveguide 125 connected to the second single-mode waveguide 123, and the distance between the center axis of the third single-mode waveguide 223 and the center axis of the second multi-mode waveguide 225 connected to the third single-mode waveguide 223 are the same.

The modulated light branched by the optical branching element 117 propagates from the second single-mode waveguide 123 to the first multi-mode waveguide 125, and from the third single-mode waveguide 223 to the second multi-mode waveguide 225.

The modulated light propagated to the first multimode waveguide 125 and the second multimode waveguide 225 is converted into a number of spatial mode lights in each multimode waveguide 125, 225, which has a curved region that serves to excite higher-order spatial modes, and a time difference in propagation between the modes is given, generating a time-varying speckle (shade image) at the end of the waveguide.

65 modes are excited in the first multimode waveguide 125 with a width of 23.9 μm, and 51 modes are excited in the second multimode waveguide 225 with a width of 18.8 μm. Since the speckles generated in the multimode waveguides of different sizes (different widths in this embodiment) are different, the independence of neurons between the waveguides is maintained. The generated speckles are photoelectrically converted into electrical signals by each light receiving element, which is an array of multiple photodiodes.

Each light receiving element 119, 219 is formed of elements according to the number of modes. In FIG. 1, for simplicity, each light receiving element is illustrated as eight light receiving elements, but for example, since the first multimode waveguide has 65 modes, the first light receiving element 119 connected to the first multimode waveguide 125 is made up of 65 light receiving elements. Also, since the second multimode waveguide 225 has 51 modes, the light receiving element 219 connected to the second multimode waveguide 225 is made up of 51 light receiving elements.

The electrical signals photoelectrically converted by each light receiving element 119, 219 are converted to digital signals by A/D converters 131, 231 and output to digital circuit 133. In digital circuit 133, a weighting electrical circuit multiplies each signal by its respective weight, and an electrical circuit that adds the weighted electrical signals sums up all the values multiplied by each respective weight. An inference value is then output via a D/A converter. Note that each weight is preset so that it approximately matches the target data y(n) during the learning stage.

By using the optical neural network device 1 in this embodiment, a number of neurons of 116 (65 + 51) can be obtained. In this embodiment, a two-branch element is used for the optical branching element 117, but it is possible to further increase the number of neurons by using an element with four branches, eight branches, or more branches and connecting multimode waveguides with different waveguide widths to the branch outputs. In this way, the number of neurons can be significantly increased without using multiple wavelengths or division on the time axis.

In this embodiment, the first multimode waveguide 125 and the second multimode waveguide 225 are the same, 0.22 μm in thickness and 39 mm in length, but have different widths of 23.9 μm and 18.8 μm, respectively. In other words, the two multimode waveguides differ only in width, and a configuration is described in which different speckle patterns are generated. Another configuration for generating different speckle patterns is described below as a modified example.

(Variation 1)
In the first modification, the first multimode waveguide 125 and the second multimode waveguide 225 have the same length and width, but the multimode waveguides have different thicknesses. That is, the thickness is 0.22 μm, and the width is one of 23.9 μm or 18.8 μm, for example, and the thickness of one multimode waveguide is 0.22 μm and the thickness of the other multimode waveguide is 0.25 μm. In this way, even if the width of the multimode waveguide is the same, by making the thickness different, it is possible to generate different speckle patterns, and the number of neurons can be significantly increased.

(Variation 2)
8 is a diagram showing a schematic configuration of the length of the multimode waveguide in the modified example 2. In the modified example 2, as shown in FIG. 8, the first multimode waveguide 125 and the second multimode waveguide 225 have the same width and thickness, and multimode waveguides of different lengths are used. That is, the width is, for example, 23.9 μm or 18.8 μm, and the thickness is, for example, 0.22 μm or 0.25 μm, and the length of one multimode waveguide is 39 mm and the length of the other multimode waveguide is 35 mm. In this way, even if the width and thickness of the multimode waveguide are the same, different speckle patterns can be generated by changing the length, and the number of neurons can be significantly increased.

(Variation 3)
Fig. 9 is a diagram showing a connection portion of a single-mode waveguide and a multi-mode waveguide in Modification 3. In Modification 3, each multi-mode waveguide uses a waveguide having the same size in width, length, and thickness, and as shown in Fig. 9, the second single-mode waveguide 123 and the first multi-mode waveguide 125, and the third single-mode waveguide 223 and the second multi-mode waveguide 225 are connected at positions shifted by different lengths from the centers of the respective waveguides, that is, the distance between the central axis of the second single-mode waveguide 123 and the central axis of the first multi-mode waveguide 125 connected to the second single-mode waveguide 123, and the distance between the central axis of the third single-mode waveguide 223 and the central axis of the second multi-mode waveguide 225 connected to the third single-mode waveguide 223 are different. Even if the first multimode waveguide 125 and the second multimode waveguide 225 have the same width, length, and thickness, by changing the distance between the central axes of each waveguide at the connection portion between the single-mode waveguide and the multimode waveguide, it is possible to generate different speckle patterns and significantly increase the number of neurons.

(Variation 4)
In the above-described embodiment and modification, the input polarization to the second single mode waveguide 123 and the third single mode waveguide 223 is set to TE mode (transverse electric modes), and a branching element of a multimode interference waveguide is applied to the optical branching element 117. In modification 4, each multimode waveguide uses a waveguide having the same size in width, length, and thickness, and the input polarization to the second single mode waveguide 123 and the third single mode waveguide 223 is set to excite both TE mode (transverse electric modes) in which the vibration direction of the electric field of the light wave is in the width direction of the waveguide and TM mode (transverse magnetic modes) in which the vibration direction of the electric field of the light wave is in the thickness direction by changing the input angle to the first single mode waveguide 121, and a TE/TM mode separator is applied to the branching element. In other words, by configuring the second single-mode waveguide 123, the first multi-mode waveguide 125, and the first light-receiving element 119 for the TE mode and the third single-mode waveguide 223, the second multi-mode waveguide 225, and the second light-receiving element 219 for the TM mode, it is possible to double the number of neurons even if the multi-mode waveguide width, thickness, and length are the same.

(Variation 5)
Fig. 10 is a diagram showing a schematic configuration of an optical neural network device 2 of the modified example 5. As shown in Fig. 10, the modified example 5 has a multi-stage configuration of multi-mode waveguides in which a second multi-mode waveguide 225 is further connected to a third multi-mode waveguide 227. By adopting such a structure, further mixing between modes is performed, the correlation of the output signal is reduced, and it is possible to effectively increase the number of neurons.

It goes without saying that the number of neurons can be significantly increased by combining any of the multimode waveguide width, length, offset from the waveguide center, and application of TE/TM mode separators to the branching elements described in this embodiment and variants 1 to 5.

In the previous example, the multimode waveguide was connected to a single-mode waveguide, but if the output light from the phase modulator can be made multimode at any stage and further coupled to the multimode waveguides 125 and 225, the number of neurons can be significantly increased with the above-mentioned configuration.

Furthermore, the multimode waveguide input photodetector converts the signal into an electrical signal, then converts it into a digital signal using an A/D converter, multiplies each signal by a weight in a digital circuit, and then sums them all together, converts them back into an analog signal using a D/A converter to obtain the output (inferred value). However, the output can be obtained by applying a circuit that directly integrates the electrical output from the photodetector.

As described above, according to this embodiment, by increasing the number of neurons in an optical neural network device, it is possible to improve the inference performance of artificial intelligence.

1, 2 Optical neural network device 111 Single wavelength light source 113 Phase modulator 117 Optical branching element 119 First light receiving element 121 First single mode waveguide 123 Second single mode waveguide 125 First multimode waveguide 131, 231 A/D converter 133 Digital circuit 219 Second light receiving element 223 Third single mode waveguide 225 Second multimode waveguide 227 Third multimode waveguide 310 Silicon substrate (chip)

Claims (9)

A speckle generating circuit for use in an optical neural network device, comprising:
a phase modulator that receives an optical signal from a light source and performs phase modulation;
a single single-mode waveguide for receiving and propagating the phase-modulated light;
an optical branching element connected to the single-mode waveguide for branching the propagated modulated light into a plurality of beams;
a plurality of single mode waveguides for propagating the respective branched modulated lights;
a plurality of multi-mode waveguides each connected to each of the branched single-mode waveguides, for generating speckles from the propagated modulated light. The speckle generating circuit according to claim 1, characterized in that each of the multimode waveguides has a different waveguide width. The speckle generating circuit according to claim 1 or 2, characterized in that each of the multimode waveguides has a different waveguide thickness. The speckle generating circuit according to any one of claims 1 to 3, characterized in that each of the multimode waveguides has a different waveguide length. The speckle generation circuit according to any one of claims 1 to 4, characterized in that the multi-mode waveguides are connected at positions where the distance between the central axis of each branched single-mode waveguide and the central axis of each multi-mode waveguide connected to each branched single-mode waveguide is the same. The speckle generating circuit according to any one of claims 1 to 4, characterized in that the multi-mode waveguides are connected at positions where the distances between the central axis of each branched single-mode waveguide and the central axis of each multi-mode waveguide connected to each branched single-mode waveguide are all different. The speckle generation circuit according to any one of claims 1 to 6, characterized in that the optical branching element splits the input modulated light into TE mode (transverse electric modes) and TM mode (transverse magnetic modes) and outputs different signals to each of the branched single mode waveguides. An optical neural network device for generating a large number of neurons,
A speckle generating circuit according to any one of claims 1 to 7,
and a plurality of light receiving elements that photoelectrically convert the speckles generated by the speckle generating circuit into electrical signals. 9. The optical neural network device according to claim 8, wherein the number of said light receiving elements corresponds to the number of modes output from said multi-mode waveguides connected thereto.