On-chip integrated photon reservoir structure and time series prediction method

Abstract

The present invention provides an on-chip integrated photon reservoir structure and a time series prediction method, comprising an input layer, an output layer and a diffraction reservoir layer arranged on a chip; the input layer comprises a task light input waveguide, a reference light input waveguide and a loop waveguide; the optical signals input by the task light input waveguide, the reference light input waveguide and the loop waveguide enter the diffraction reservoir layer at the same time; the diffraction reservoir layer comprises a preset number of diffraction groove structures; the output layer is used to output the optical signal processed by diffraction, and comprises a Y-type waveguide and an output waveguide; the input end of the Y-type waveguide is connected to the diffraction reservoir layer, one end of the output end of the Y-type waveguide is connected to the loop waveguide, and the other end of the output end of the Y-type waveguide is connected to the output waveguide. The present invention inputs an optical signal with an amplitude or phase according to a preset period through a reference light input waveguide, and designs a preset number of diffraction groove structures to form a diffraction reservoir layer, thereby enriching the internal state of the diffraction reservoir layer and making it adjustable.

Description

Technical Field

[0001] The present invention relates to the field of reservoir computing technology, and in particular to an on-chip integrated photon reservoir structure and a time series prediction method. Background Art

[0002] In recent years, the reduction in transistor size and the resulting exponential growth in the number of transistors in integrated circuits have driven the improvement of computer performance and led to the emergence of related applications such as artificial intelligence. However, the reduction in transistor size has begun to approach its physical limits and faces challenges such as quantum effects. At the same time, the inherent limitations of the separation of "storage" and "computation" in traditional hardware, known as the von Neumann architecture, have led to problems such as limitations in further improving its computing efficiency.

[0003] As a mathematical model that imitates the information processing method of the human brain, artificial neural network involves a large amount of parallel computing and data distributed processing. In recent years, it has made rapid progress in many fields such as target recognition, natural language processing, autonomous driving and safety monitoring. Its powerful information processing ability makes it a successful example in bionics. The current neural network can be divided into two categories. One is the feedforward neural network. In this type of model, neurons are divided into layers with a front-to-back order, and the signal only propagates forward. However, since this structure with only one-way propagation cannot handle time-related networks. The other type is the recursive neural network model. The neurons between recursive neural networks can be connected to each other and have a certain memory function, so this model can be used to solve time-related tasks. However, this recursive structure has the problem of gradient explosion or disappearance, and has high requirements for the algorithm and hardware implementation.

[0004] To address the above issues, the reservoir computing model was proposed. Among various artificial neural network architectures, the reservoir computing model has the advantage of only needing to train the network output layer, and has also been proven to be suitable for processing complex time series data with the lowest training cost.

[0005] The reservoir computing architecture includes an input layer, an output layer, and a reservoir layer. The current physical implementation schemes all use different devices and methods to implement the calculation, such as electronic devices such as memristors, or optical devices such as optical fibers, which poses challenges to the system's integration and process. At the same time, the above architecture uses the inherent differences of nonlinear devices to achieve different reservoir states. This method makes the richness of the system state fixed after the system is prepared, making it difficult to make further adjustments. Summary of the invention

[0006] To address the defects of the prior art, the present invention provides an on-chip integrated photon storage pool structure and a time series prediction method.

[0007] In order to achieve the above technical objectives, the specific technical solutions adopted by the present invention are as follows:

[0008] In one aspect, the present invention provides an on-chip integrated photon reservoir structure, comprising an input layer, an output layer and a diffraction reservoir layer arranged on the chip;

[0009] The input layer includes a task light input waveguide, a reference light input waveguide, and a loop waveguide; the task light input waveguide is used to input an optical signal containing a task time series data set; the reference light input waveguide is used to input an amplitude or phase modulated optical signal; the loop waveguide is used to input an optical signal output at the previous time step; the optical signals input by the task light input waveguide, the reference light input waveguide, and the loop waveguide enter the diffraction reserve pool layer at the same time;

[0010] The diffraction reserve pool layer includes a preset number of diffraction groove structures, which are used to perform diffraction processing on the optical signal input from the input layer;

[0011] The output layer is used to output the optical signal processed by diffraction, and includes a Y-type waveguide and an output waveguide; the input end of the Y-type waveguide is connected to the diffraction reserve pool layer, one end of the Y-type waveguide output end is connected to the loop waveguide, and the other end of the Y-type waveguide output end is connected to the output waveguide.

[0012] Furthermore, the optical signal input by the reference optical input waveguide is amplitude or phase modulated according to a preset period.

[0013] Furthermore, it also includes an optical amplifier for enhancing the light intensity of the loop optical signal, and the optical amplifier is arranged between the loop waveguide and the Y-type waveguide.

[0014] Furthermore, the diffraction groove structure is a sub-wavelength structure, which can achieve a 0-2π phase shift, and the width of each diffraction groove structure is equal.

[0015] Furthermore, the diffraction groove structure is composed of diffraction grooves of different lengths.

[0016] Furthermore, it also includes a driving laser and a modulator, wherein the driving laser is used to generate an input optical signal, and the modulator is used to modulate the amplitude or phase of the input optical signal; the driving laser is connected to the input end of the modulator, and the output end of the modulator is respectively connected to the task light input waveguide and the reference light input waveguide.

[0017] Furthermore, it also includes a photodetector, which is connected to the output end of the output waveguide and is used to collect and record the optical signal of each time step output by the output waveguide.

[0018] Furthermore, the target signal is trained by a fully connected neural network algorithm, a ridge regression algorithm, a linear regression algorithm and the optical signal output by the output waveguide to obtain the output weight.

[0019] Further, the reference light input waveguide includes a first reference light input waveguide and a second reference light input waveguide.

[0020] On the other hand, the present invention also provides a time series prediction method of an on-chip integrated photon reservoir structure, comprising the following steps:

[0021] S1, obtain the amplitude or phase of the task time series data set;

[0022] S2, amplitude or phase modulating the signal light input into the task light input waveguide according to the amplitude or phase of the task time series data set;

[0023] S3, amplitude or phase modulating the signal light of the input reference light input waveguide according to a preset period;

[0024] S4, inputting the optical signals of the task light input waveguide, the reference light input waveguide, and the loop waveguide into the diffraction reservoir layer at the same time, and collecting and recording the optical signal of each time step output by the output waveguide;

[0025] S5, training the target signal through the algorithm and the optical signal output by the output waveguide to obtain the output weight;

[0026] S6. Keep the output weight unchanged, continue to collect and record the optical signal of each time step output by the output waveguide, apply the output weight to the collected and recorded optical signal, and obtain the required prediction sequence information.

[0027] Compared with the prior art, the beneficial effects of the present invention are:

[0028] The present invention provides an on-chip integrated photon reservoir structure and a time series prediction method. By referring to the optical input waveguide input amplitude or phase of an optical signal with a preset period, the internal state richness of a determined reservoir structure is adjusted, thereby enriching the internal state of a diffraction reservoir layer. The diffraction reservoir layer of the present invention is composed of a preset number of diffraction groove structures. After the optical signal enters the diffraction reservoir layer, it is diffracted by multiple groups of diffraction groove structures in sequence, so that the internal state of the diffraction reservoir layer is sufficiently rich.

[0029] The present invention inputs the optical signal output at the previous time step through a loop waveguide, allowing the optical signal output at the previous time step to support the optical signal output at the current time step, that is, allowing historical data to provide support for predicted data, thereby making the on-chip integrated photon reserve pool structure have memory.

[0030] The present invention adopts an on-chip diffraction structure to integrate the photon reserve pool on the chip, which greatly improves the density of physical nodes, enhances the computing power of the reserve pool, and improves the robustness of the reserve pool.

[0031] The present invention uses optical signals for calculation in the entire time series prediction process, which greatly reduces the energy consumption of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying creative work.

[0033] Figure 1 A schematic diagram of the structure of an on-chip integrated photon storage pool provided by an embodiment;

[0034] Figure 2 A partially enlarged schematic diagram of a diffraction groove structure provided in an embodiment.

[0035] Figure annotation:

[0036] 1. Task light input waveguide; 2. Reference light input waveguide; 3. Loop waveguide; 4. Optical amplifier; 5. Output waveguide; 6. Y-type waveguide; 7. Diffraction slot structure; 8. Diffraction slot. DETAILED DESCRIPTION

[0037] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0038] Reference Figure 1 , one embodiment provides an on-chip integrated photon reservoir structure, including an input layer, an output layer and a diffraction reservoir layer arranged on the chip;

[0039] The input layer includes a task light input waveguide 1, a reference light input waveguide 2, and a loop waveguide 3; the task light input waveguide 1 is used to input an optical signal containing a task time series data set; the reference light input waveguide 2 is used to input an amplitude or phase modulated optical signal; the loop waveguide 3 is used to input an optical signal output at the previous time step; the optical signals input by the task light input waveguide 1, the reference light input waveguide 2, and the loop waveguide 3 enter the diffraction reserve pool layer at the same time;

[0040] The diffraction reserve pool layer includes a preset number of diffraction groove structures 7, which are used to perform diffraction processing on the optical signal input from the input layer;

[0041] The output layer is used to output the optical signal processed by diffraction, and includes a Y-type waveguide 6 and an output waveguide 5; the input end of the Y-type waveguide 6 is connected to the diffraction reserve pool layer, one end of the output end of the Y-type waveguide 6 is connected to the loop waveguide 3, and the other end of the output end of the Y-type waveguide 6 is connected to the output waveguide 5.

[0042] An optical amplifier 4 is also arranged between the loop waveguide 3 and the Y-type waveguide 6, and the optical amplifier 4 is used to enhance the light intensity of the loop optical signal, thereby enhancing the light signal output at the previous time step, so that the light intensity of the light signal output at the previous time step is enhanced to be able to support the light signal output at the current time step.

[0043] The input ends of the task light input waveguide 1 and the reference light input waveguide 2 are connected to a modulator, which is used to modulate the amplitude or phase of the input light signal and map the required data to the amplitude or phase of the light signal; the other end of the modulator is connected to a driving laser, which is used to generate the input light signal.

[0044] The optical signal of the input reference light input waveguide 2 is amplitude or phase modulated by the modulator at a preset period, thereby adjusting the richness of the internal state of the determined reserve pool structure and enriching the internal state of the diffraction reserve pool layer.

[0045] The output end of the output waveguide 5 is connected to a photodetector, and the photodetector is used to collect and record the optical signal of each time step output by the output waveguide.

[0046] In one embodiment, the target signal is trained in a computer through a fully connected neural network algorithm, a ridge regression algorithm, a linear regression algorithm combined with the light signal of each time step collected and recorded by the photodetector to obtain the output weight.

[0047] The reference light input waveguide 2 includes a first reference light input waveguide and a second reference light input waveguide. By using two reference light input waveguides to input a phase- or amplitude-modulated optical signal, the internal state of the diffraction reservoir layer is enriched, and the richness is adjustable.

[0048] The diffraction groove structure 7 is a sub-wavelength structure, which can achieve a 0-2π phase shift, and the widths of the diffraction groove structures 7 are equal. Figure 2 The diffraction groove structure 7 is composed of diffraction grooves 8 of different lengths, and the specific lengths are randomly generated. After the light signal input from the input layer enters the diffraction reserve pool layer, it is diffracted by multiple groups of diffraction groove structures in sequence, so that the internal state of the diffraction reserve pool layer is rich enough.

[0049] In one embodiment, a time series prediction method using the above-mentioned on-chip integrated photon reservoir structure is provided, comprising the following steps:

[0050] S1, obtain the amplitude or phase of the task time series data set;

[0051] S2, amplitude or phase modulating the signal light input into the task light input waveguide according to the amplitude or phase of the task time series data set;

[0052] S3, amplitude or phase modulating the signal light of the input reference light input waveguide according to a preset period;

[0053] S4, inputting the optical signals of the task light input waveguide, the reference light input waveguide, and the loop waveguide into the diffraction reservoir layer at the same time, and collecting and recording the optical signal of each time step output by the output waveguide;

[0054] S5, training the target signal through the algorithm and the optical signal output by the output waveguide to obtain the output weight;

[0055] S6. Keep the output weight unchanged, continue to collect and record the optical signal of each time step output by the output waveguide, apply the output weight to the collected and recorded optical signal, and obtain the required prediction sequence information.

[0056] In one embodiment, the on-chip integrated photon reservoir structure adopts four sets of diffraction groove structures, and the length of the diffraction groove is randomly generated between 0.2μm-3μm based on the actual preparation process, and the width is 0.2μm. The above-mentioned on-chip integrated photon reservoir structure is used to perform a prediction task on the Mackey-Glass time series data set, and the prediction is performed according to the following steps:

[0057] S1, phase mapping after normalization of the Mackey-Glass time series dataset;

[0058] S2, phase modulating the signal light of the input task light input waveguide using the phase obtained in S1;

[0059] S3, phase modulating the signal light of the input reference light input waveguide by adding one degree of phase per step length;

[0060] S4, inputting the optical signals of the task light input waveguide, the reference light input waveguide, and the loop waveguide into the diffraction reservoir layer at the same time, and collecting and recording the optical signal of each time step output by the output waveguide;

[0061] S5, training the target signal by using a ridge regression algorithm and the optical signal output by the output waveguide to obtain an output weight;

[0062] S6. Keep the output weight unchanged, continue to collect and record the optical signal of each time step output by the output waveguide, apply the output weight to the collected and recorded optical signal, and obtain the predicted sequence information of the Mackey-Glass time series data set.

[0063] Matters not covered by the present invention are known technologies.

[0064] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0065] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention. It should be pointed out that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

[0066] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An on-chip integrated photon reservoir structure, characterized in that: It includes an input layer, an output layer and a diffraction reservoir layer arranged on the chip; The input layer includes a task light input waveguide, a reference light input waveguide, and a loop waveguide; the task light input waveguide is used to input an optical signal containing a task time series data set; the reference light input waveguide is used to input an amplitude or phase modulated optical signal; the loop waveguide is used to input an optical signal output at the previous time step; the optical signals input by the task light input waveguide, the reference light input waveguide, and the loop waveguide enter the diffraction reserve pool layer at the same time; The diffraction reserve pool layer includes a preset number of diffraction groove structures, which are used to perform diffraction processing on the optical signal input from the input layer; The output layer is used to output the optical signal processed by diffraction, and includes a Y-type waveguide and an output waveguide; the input end of the Y-type waveguide is connected to the diffraction reserve pool layer, one end of the Y-type waveguide output end is connected to the loop waveguide, and the other end of the Y-type waveguide output end is connected to the output waveguide.

2. The on-chip integrated photon reservoir structure according to claim 1, characterized in that: The optical signal input by the reference optical input waveguide is amplitude or phase modulated according to a preset period.

3. The on-chip integrated photon reservoir structure according to claim 1, characterized in that: It also includes an optical amplifier for enhancing the light intensity of the loop optical signal, and the optical amplifier is arranged between the loop waveguide and the Y-type waveguide.

4. The on-chip integrated photon reservoir structure as claimed in claim 1, characterized in that: The diffraction groove structure is a sub-wavelength structure, which can achieve a 0-2π phase shift, and the widths of each diffraction groove structure are equal.

5. The on-chip integrated photon reservoir structure as claimed in claim 4, characterized in that: The diffraction groove structure is composed of diffraction grooves of different lengths.

6. The on-chip integrated photon reservoir structure according to claim 1, characterized in that: It also includes a driving laser and a modulator, wherein the driving laser is used to generate an input optical signal, and the modulator is used to modulate the amplitude or phase of the input optical signal; the driving laser is connected to the input end of the modulator, and the output end of the modulator is respectively connected to the task optical input waveguide and the reference optical input waveguide.

7. The on-chip integrated photon reservoir structure according to claim 1, characterized in that: It also includes a photodetector, which is connected to the output end of the output waveguide and is used to collect and record the optical signal of each time step output by the output waveguide.

8. The on-chip integrated photon reservoir structure as claimed in claim 7, characterized in that: The target signal is trained through a fully connected neural network algorithm, a ridge regression algorithm, a linear regression algorithm and combined with the optical signal output by the output waveguide to obtain the output weight.

9. The on-chip integrated photon reservoir structure as claimed in claim 1, characterized in that: The reference light input waveguide includes a first reference light input waveguide and a second reference light input waveguide.

10. A time series prediction method based on the on-chip integrated photon reservoir structure as claimed in claim 1, characterized in that: The following steps are involved: S1, obtain the amplitude or phase of the task time series data set; S2, amplitude or phase modulating the signal light input into the task light input waveguide according to the amplitude or phase of the task time series data set; S3, amplitude or phase modulating the signal light of the input reference light input waveguide according to a preset period; S4, inputting the optical signals of the task light input waveguide, the reference light input waveguide, and the loop waveguide into the diffraction reserve pool layer at the same time, wherein the diffraction reserve pool layer includes a preset number of diffraction groove structures, and collects and records the optical signal of each time step output by the output waveguide; S5, training the target signal through the algorithm and the optical signal output by the output waveguide to obtain the output weight; S6. Keep the output weight unchanged, continue to collect and record the optical signal of each time step output by the output waveguide, apply the output weight to the collected and recorded optical signal, and obtain the required prediction sequence information.

Images