Multi-feedback optical reserve pool system based on optical fiber ring
By introducing optical fiber rings into the optical delay storage pool calculation system, a multi-feedback optical storage pool system is solved, and a higher performance storage pool calculation is achieved, suitable for handling complex tasks.
Patent Information
- Application Number
- CN202411723078.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-28
- Publication Date
- 2025-03-07
AI Technical Summary
The network topology of the existing optical delay storage pool computing system is too simple, resulting in insufficient performance when handling complex tasks, and adding feedback loops will make the system bloated and difficult to achieve.
A multi-feedback optical reserve pool system based on optical fiber rings is adopted to achieve multiple optical delays through optical fiber rings, increase the connection density of virtual nodes, and enrich the connection state of nodes by changing the number and connection methods of optical fiber rings, and build a high-performance, highly interconnected all-optical RC network.
A more complex neural network topology is implemented, the performance of reserve pool computing is improved, the difficulty of device implementation is simplified, and the ability to handle more complex tasks.
Smart Images
Figure CN119578484A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial neural networks, and in particular to a new structure for physical realization of optical time-delay reservoir computing (RC). Background Art
[0002] Artificial neural network is a computing model inspired by biological neural networks, which is used to solve the von Neumann bottleneck problem in traditional computers. Reservoir computing is a special kind of artificial neural network. It is a new neuromorphic computing paradigm that mimics the way brain information is processed, and it originates from the evolution of recurrent neural networks. In terms of time series signal processing, such as chaotic time series prediction and nonlinear channel equalization, reservoir computing has shown excellent performance.
[0003] The advantage of this new computing paradigm lies in its simplified training process. Compared with traditional recurrent neural networks, reservoir computing only requires training output weights, while input weights and weights within the reservoir are randomly fixed. This reduces the cost of training, making it an attractive approach. Optical reservoir computing is a machine learning method implemented using photonic principles, integrating cutting-edge research fields in optics and computational science, providing a new way to process time series signals and information, and has unique advantages in processing time series data.
[0004] Optical reservoir computing can be realized through spatial structure and time delay structure. The optical reservoir computing of time delay structure only includes one nonlinear node, so it has the unique advantage of simplified structure. Optical reservoir computing of time delay structure includes photoelectric type and all-optical type. Photoelectric type reservoir computing has a photoelectric conversion process, and its noise and circuit bandwidth limit the data processing rate and reduce the energy utilization efficiency. All-optical reservoir computing avoids these drawbacks and has the advantages of wide bandwidth, high data processing rate and low loss.
[0005] Research has shown that by adding feedback loops within RC, RC parallel connection, RC series-parallel hybrid configurations, and mutual coupling, the predictive performance of the reserve pool calculation can be effectively improved and its memory capacity can be enhanced. However, these complex structures also increase the complexity of hardware implementation. Despite this, the single-node all-optical delay RC structure is still limited to a limited multi-feedback loop structure, the connection between virtual nodes is still too simple, and its performance still needs to be further improved. In addition, more feedback loops may cause the system to become bloated and difficult to implement in terms of hardware.
[0006] The training speed of the delay reservoir calculation is indeed impressive, but its accuracy still needs to be improved compared to a fully trainable neural network. The fundamental reason for this limitation is that the network topology of the delay reservoir is too simple and the connections between the internal neurons are sparse. If we increase the density of connections, the system will be able to exhibit more complex dynamic behaviors, thereby improving the ability to handle more complex tasks. Summary of the invention
[0007] The purpose of the present invention is to provide a multi-feedback optical reservoir system based on an optical fiber ring, which has a simple physical structure and can enhance the performance of reservoir calculation. The present invention uses a simple physical structure to realize a complex network topology, uses mature semiconductor lasers as nonlinear nodes of optical RC, and uses optical fiber rings as feedback loops to stimulate the complex dynamic characteristics of nonlinear nodes. By increasing the interconnection between RC virtual nodes, a high-performance, highly interconnected all-optical RC network is constructed. The technical solution is as follows:
[0008] A multi-feedback optical reservoir system based on an optical fiber ring comprises an input layer, a reservoir layer and an output layer connected in sequence, wherein the input layer is used to multiply a sample signal with a mask signal and modulate it on the light emitted by a narrow-band laser through a modulator, and the modulated light emitted by the narrow-band laser enters the reservoir layer. The reservoir layer is characterized in that: the reservoir layer is used to provide a laser feedback loop including a nonlinear response, and a fiber ring is used on the feedback loop to perform multiple optical delays.
[0009] Furthermore, the modulated light output from the input layer is stored in a pool layer through a first polarization controller (PC1); in the pool layer, the light coming out of the first coupler enters port 1 of the circulator, and enters a laser providing a nonlinear response from port 2 of the circulator, the output light of the laser providing a nonlinear response returns to port 2 of the circulator, the light coming out of port 3 of the circulator passes through a second polarization controller (PC2) and then enters the second coupler, the light coming out of port 2 of the second coupler passes through an attenuator (VA) and then enters the third coupler, a portion of the light that enters the third coupler passes through at least one optical fiber ring and then returns to the third coupler, another portion of the light comes out of the third coupler and then returns to the first coupler, forming a feedback loop, and the light coming out of port 3 of the second coupler enters the output layer.
[0010] Furthermore, the optical fiber ring is a single optical fiber ring or a nested optical fiber ring, or a cascade and combination of one or two optical fiber rings, and the number of cascaded and combined optical fiber rings is more than two.
[0011] Furthermore, the single optical fiber ring includes a third coupler and a section of feedback optical fiber connected between two ports of the third coupler.
[0012] Further, the nested fiber ring is a nest of two single fibers, including a third coupler, a fourth coupler and three sections of feedback fiber, wherein the third coupler is connected between the first coupler and the attenuator (VA), two of the three sections of feedback fiber are respectively connected between the third coupler and the fourth coupler, and the third section of feedback fiber is connected between two ports of the fourth coupler.
[0013] Furthermore, let γ be the coupling ratio of the third coupler. By changing the coupling ratio γ, the feedback strength in the optical fiber ring is adjusted. By adjusting the feedback strength, the connection weights in the reserve pool network structure are adjusted and reshaped, thereby changing the structural composition of the neural network.
[0014] Furthermore, in the input layer, the modulator is a Mach-Zehnder modulator.
[0015] Furthermore, the output layer includes: an optical amplifier, a photodetector and a digitizer connected in sequence, and the digital signal is collected through the output layer for training the output weights.
[0016] Furthermore, the narrowband laser is a tunable laser.
[0017] Further, the laser providing the nonlinear response is a semiconductor laser, a fiber laser or a VCSEL laser.
[0018] The present invention introduces a fiber optic ring on the basis of a single feedback loop to achieve the realization of a complex neural network topology, which enables multiple feedback loops with better performance to be realized in hardware, greatly simplifies the difficulty of device implementation, and enhances the predictive performance of the reserve pool. Similarly, based on the technical solution proposed by the present invention, the connection state of the node can be further enriched to change its complexity by changing the number and connection method of the fiber optic ring. The reserve pool system generated by this method can be equivalent to the cascade effect of multiple reserve pools, which is equivalent to the neural network architecture of deep learning, and contains multiple hidden layers for processing complex training tasks. BRIEF DESCRIPTION OF THE DRAWINGS
[0019] Figure 1 A structural diagram of a multi-feedback optical reservoir system including a single optical fiber ring according to an embodiment of the present invention.
[0020] Figure 2 A structural diagram of a multi-feedback optical reservoir system including nested fiber rings according to an embodiment of the present invention.
[0021] Figure 3 A structural diagram of a multi-feedback optical reservoir system including cascaded optical fiber rings according to an embodiment of the present invention.
[0022] Figure 4A simple structural diagram of the basic connection method of the optical fiber ring in the storage tank layer of an embodiment of the present invention.
[0023] Figure 5 Schematic diagram of various combined connection structures of optical fiber rings in a storage pool layer according to an embodiment of the present invention. Specific implementation methods
[0025] The present invention provides a multi-feedback optical reservoir system based on a fiber optic ring. The system consists of three main parts: an input layer, a reservoir layer and an output layer. The system aims to enhance the performance of reservoir calculation by increasing the connection density between neurons in the reservoir layer.
[0026] At the input layer, the system includes a narrowband laser as a driving laser, a modulator, and an arbitrary waveform generator. The narrowband laser is connected to the modulator via an optical fiber, and the modulator is directly connected to the reservoir layer. The function of the modulator is to multiply the sample data with the chaotic mask, and the modulated signal obtained after processing is used for the optical signal emitted by the narrowband laser.
[0027] The core of the reservoir layer is a feedback loop composed of a laser that provides nonlinear response. This laser can also be called a response laser. The loop uses a fiber optic ring to achieve multiple optical delays. By adjusting the length of the fiber optic ring, the delay loop in the reservoir can be precisely adjusted, thereby affecting the dynamic characteristics of the system.
[0028] The output layer includes an optical amplifier for amplifying the output optical signal; a photodetector for converting light intensity into electrical signals; and a digitizer for converting electrical signals into analog-to-digital signals and collecting digital signals for subsequent processing and analysis.
[0029] The multi-feedback optical reserve pool calculation method based on the fiber optic ring provided by the present invention first modulates the sample data to the light emitted by the narrowband laser through a modulator after chaos mask processing, injects it into the laser providing nonlinear response, and returns to the laser providing nonlinear response through a feedback loop. The number of virtual nodes in the reserve pool layer is determined by the length of the feedback loop and the time delay between virtual nodes, and then the fiber optic ring is introduced into the feedback loop to achieve the effect of multi-segment feedback. The status of these virtual nodes in the feedback loop is recorded, and their output weights are obtained through the ridge regression algorithm to complete the training task of the reserve pool calculation. The reserve pool layer can increase the number of fiber optic rings to complete a more complex network, such as cascaded fiber optic rings, nested fiber optic rings, and can also be used in combination. The variable connection method can freely change the network state to complete the required requirements of different tasks.
[0030] Through this structural design, the present invention not only improves the performance of the reserve pool calculation, but also provides strong support for processing more complex tasks through the multi-feedback loop network topology realized by the optical fiber ring.
[0031] The entire structure diagram is attached Figure 1 As shown, the operation process is as follows:
[0032] In the input layer, the sample signal is first multiplied by the mask signal through data preprocessing, and then modulated on the light emitted by the narrow-band laser through the Mach-Zehnder modulator. The light enters the reservoir layer through the polarization controller (PC1).
[0033] In the reservoir layer, the light from the first coupler 1 enters port 1 of the circulator 1, enters the laser providing nonlinear response from port 2 of the circulator 1, and the output light of the laser providing nonlinear response returns to port 2 of the circulator 1, exits from port 3 and enters the polarization controller (PC2), and then enters the second coupler 2, dividing the light into two paths. Most of the light exits from port 2 of the second coupler 2 and passes through the attenuator (VA), and then enters the third coupler 3 again. A part of the light enters the fiber ring through port 4 of the third coupler 3, exits from port 2 and returns to the coupler again, and another part of the light exits from port 1 of the third coupler 3 and returns to the first coupler 1 again, completing the feedback loop. Another part of the light from port 3 of the second coupler 2 enters the output layer, and its status is recorded for training and testing on the computer side.
[0034] In the output layer, an erbium-doped fiber amplifier is used to amplify the output optical signal, a photodetector converts the light intensity into an electrical signal, a digitizer performs analog-to-digital conversion on the electrical signal and collects digital signals, and the collected data is used to train the output weights.
[0035] The multi-feedback optical reservoir based on fiber ring and its calculation method proposed in the present invention are intended to improve the complexity of the neural network topology, thereby enhancing the performance of the reservoir calculation. The method first involves processing the sample data through a chaotic mask, and then modulating the processed data onto an optical signal emitted by a narrowband laser. These optical signals are then injected into a laser that provides a nonlinear response and returned to the laser that provides a nonlinear response via a feedback loop. In this process, the length of the feedback loop and the time delay between virtual nodes jointly determine the number of virtual nodes in the reservoir layer, and the introduction of the fiber ring achieves the effect of multi-segment feedback.
[0036] By recording the states of the virtual nodes in the feedback loop and applying the ridge regression algorithm, we can determine the output weights to complete the training task of the reservoir calculation. A significant advantage of this method is that it can achieve better performance at the hardware level. Multiple feedback loops, simplifying the difficulty of device implementation, and significantly improving the prediction performance of the reservoir.
[0037] In addition, the present invention further enriches the connection status of the nodes and increases the complexity of the system by changing the number and connection mode of the optical fiber rings. This design enables the reserve pool system to simulate the cascade effect of multiple reserve pools, which is equivalent to building a deep learning neural network architecture containing multiple hidden layers, enabling it to cope with more complex training tasks. In this way, the present invention not only improves computing efficiency, but also provides a powerful tool for solving complex practical problems.
[0038] In this embodiment, the first coupler 1 is a 1x2 coupler, the second coupler 2 is 80:20, and 80% of the light is retained in the loop, and the third coupler 3 is a 2x2 coupler.
[0039] The laser providing nonlinear response used in this embodiment is a DFB laser, and the selection of the laser providing nonlinear response in the reserve pool can be a semiconductor laser, a fiber laser, a VCSEL laser, and all other nonlinear devices that can provide nonlinearity to the reserve pool. As a response laser (that is, a nonlinear node) providing nonlinear response, its main purpose is to make different input signals have different transient responses after passing through the nonlinear node, and its main function is similar to the activation function in the traditional artificial neural network. The modulator uses a Mach-Zehnder modulator, and the narrowband laser uses a tunable laser.
[0040] According to the structure proposed in this scheme, the rate equation for the optical signal output from the second coupler 3 port is established as follows:
[0041]
[0042] +k inj E inj exp(2iπΔν)+F(t)
[0043]
[0044] Where E represents the complex amplitude of the slowly varying electric field of the laser providing the nonlinear response, N represents the average carrier density, α represents the linewidth enhancement factor, g is the differential gain, N0 represents the transparent carrier density, ε is the gain saturation coefficient, k represents the optical feedback intensity of the feedback loop, v represents the frequency of the laser providing the nonlinear response when it is running freely, Δv represents the frequency detuning between the narrowband laser and the laser providing the nonlinear response, τ p and τ s represent the photon lifetime and carrier lifetime respectively, J is the injection current of the laser providing nonlinear response, It represents the phase difference between the feedback field of the delayed feedback loop and the field cavity in the laser at the feedback injection point. γ is the coupling ratio of the third coupler. F(t) is the noise term.
[0045] The instrument parameters used are as follows:
[0046] α=3.0,g=8.4x10 -13 m 3 s -1 ,N0=1.4x10 24 m -3 ,ε=2.0x10 -23 ,τ p =1.927ps,
[0047] τ s =2.04ns,k inj =12.43ns -1 ,v=1.96x10 14 Hz,Δv=-4.0GHz,J=1.037x10 33 m - 3 s -1 .
[0048] By changing the coupling ratio γ, the feedback strength in the fiber ring can be adjusted. By adjusting the injection intensity and feedback intensity, the connection weights in the reserve pool network structure can be adjusted and reshaped, thereby changing the structural composition of the neural network and increasing the system's adjustability and adaptability to tasks. By adding a fiber ring, when the coupling ratio is 50:50, the light is output from the response laser, passes through the second coupler port 2, and then directly outputs from port 1 through the third coupler port 4. The time it takes to pass through the first coupler port 3 is 40.05ns. When the fiber ring delay is 1ns, its NMSE reaches 4.75%
[0049] The application of the present invention realizes the complex neural network topology by introducing an optical fiber ring on the basis of a single feedback loop, which enables multiple feedback loops with better performance to be realized in hardware, greatly simplifies the difficulty of device implementation, and enhances the predictive performance of the reserve pool.
[0050] Practical application method (forecasting the Santa Fe time series)
[0051] First, the Santa Fe time series prepared for training is multiplied by the prepared mask. This step is equivalent to the input weight in the traditional artificial neural network. The signal after the mask multiplication is generated by an arbitrary waveform generator, modulated into an optical signal by a modulator, and passed through the reservoir layer. It is converted into an electrical signal by a photodetector in the output layer, and the signal output from the reservoir layer is received and recorded by a digitizer. The output signal (that is, the part of the data called activation value or feature representation in the artificial neural network) is trained by the ridge regression algorithm to obtain the weight of the output layer. After obtaining the trained output weight, you can start to actually invest in the prediction task. After the parameter data of the new Santa Fe time series enters the reservoir through the same operation as the training data, the new data is received in the digitizer. These data are directly weighted and summed with the trained output weights to obtain the required prediction results.
[0052] The present invention can also further enrich the connection state of the node and change its complexity by changing the number and connection mode of the optical fiber ring. The specific method is as follows:
[0053] 1. Use fiber optic loops in a feedback loop to delay the light multiple times.
[0054] 2. In the feedback loop, the network state of the reservoir layer is changed by cascading fiber rings to complete a more complex connection matrix.
[0055] 3. In the feedback loop, the network state of the reservoir layer is changed by nesting fiber rings to complete a more complex connection matrix.
[0056] 4. In the feedback loop, the network state of the reservoir layer is changed by simultaneously cascading and nesting fiber rings to complete a more complex connection matrix. BRIEF DESCRIPTION OF THE DRAWINGS
[0058] Figure 1 A structural diagram of a multi-feedback optical reservoir system including a single optical fiber ring according to an embodiment of the present invention.
[0059] Figure 2 A structural diagram of a multi-feedback optical reservoir system including nested fiber rings according to an embodiment of the present invention.
[0060] Figure 3 A structural diagram of a multi-feedback optical reservoir system including cascaded optical fiber rings according to an embodiment of the present invention.
[0061] Figure 4 A simple structural diagram of the basic connection method of the optical fiber ring in the storage tank layer of an embodiment of the present invention.
[0062] Figure 5Schematic diagram of various combined connection structures of optical fiber rings in a storage pool layer according to an embodiment of the present invention. Specific implementation methods
[0064] The present invention provides a multi-feedback optical reservoir system based on a fiber optic ring. The system consists of three main parts: an input layer, a reservoir layer and an output layer. The system aims to enhance the performance of reservoir calculation by increasing the connection density between neurons in the reservoir layer.
[0065] At the input layer, the system includes a narrowband laser as a driving laser, a modulator, and an arbitrary waveform generator. The narrowband laser is connected to the modulator via an optical fiber, and the modulator is directly connected to the reservoir layer. The function of the modulator is to multiply the sample data with the chaotic mask, and the modulated signal obtained after processing is used for the optical signal emitted by the narrowband laser.
[0066] The core of the reservoir layer is a feedback loop composed of a laser that provides nonlinear response. This laser can also be called a response laser. The loop uses a fiber optic ring to achieve multiple optical delays. By adjusting the length of the fiber optic ring, the delay loop in the reservoir can be precisely adjusted, thereby affecting the dynamic characteristics of the system.
[0067] The output layer includes an optical amplifier for amplifying the output optical signal; a photodetector for converting light intensity into electrical signals; and a digitizer for converting electrical signals into analog-to-digital signals and collecting digital signals for subsequent processing and analysis.
[0068] Through this structural design, the present invention not only improves the performance of the reserve pool calculation, but also provides strong support for processing more complex tasks through the multi-feedback loop network topology realized by the optical fiber ring.
[0069] The entire structure diagram is attached Figure 1 As shown, the operation process is as follows:
[0070] In the input layer, the sample signal is first multiplied by the mask signal through data preprocessing, and then modulated on the light emitted by the narrow-band laser through the Mach-Zehnder modulator. The light enters the reservoir layer through the polarization controller (PC1).
[0071] In the reservoir layer, the light from the first coupler 1 enters port 1 of the circulator 1, enters the laser providing nonlinear response from port 2 of the circulator 1, and the output light of the laser providing nonlinear response returns to port 2 of the circulator 1, exits from port 3 and enters the polarization controller (PC2), and then enters the second coupler 2, dividing the light into two paths. Most of the light exits from port 2 of the second coupler 2 and passes through the attenuator (VA), and then enters the third coupler 3 again. A part of the light enters the fiber ring through port 4 of the third coupler 3, exits from port 2 and returns to the coupler again, and another part of the light exits from port 1 of the third coupler 3 and returns to the first coupler 1 again, completing the feedback loop. Another part of the light from port 3 of the second coupler 2 enters the output layer, and its status is recorded for training and testing on the computer side.
[0072] In the output layer, an erbium-doped fiber amplifier is used to amplify the output optical signal, a photodetector converts the light intensity into an electrical signal, a digitizer performs analog-to-digital conversion on the electrical signal and collects digital signals, and the collected data is used to train the output weights.
[0073] The multi-feedback optical reservoir based on fiber ring and its calculation method proposed in the present invention are intended to improve the complexity of the neural network topology, thereby enhancing the performance of the reservoir calculation. The method first involves processing the sample data through a chaotic mask, and then modulating the processed data onto an optical signal emitted by a narrowband laser. These optical signals are then injected into a laser that provides a nonlinear response and returned to the laser that provides a nonlinear response via a feedback loop. In this process, the length of the feedback loop and the time delay between virtual nodes jointly determine the number of virtual nodes in the reservoir layer, and the introduction of the fiber ring achieves the effect of multi-segment feedback.
[0074] By recording the states of the virtual nodes in the feedback loop and applying the ridge regression algorithm, we can determine the output weights to complete the training task of the reservoir calculation. A significant advantage of this method is that it can achieve better performance at the hardware level. Multiple feedback loops, simplifying the difficulty of device implementation, and significantly improving the prediction performance of the reservoir.
[0075] In addition, the present invention further enriches the connection status of the nodes and increases the complexity of the system by changing the number and connection mode of the optical fiber rings. This design enables the reserve pool system to simulate the cascade effect of multiple reserve pools, which is equivalent to building a deep learning neural network architecture containing multiple hidden layers, enabling it to cope with more complex training tasks. In this way, the present invention not only improves computing efficiency, but also provides a powerful tool for solving complex practical problems.
[0076] In the embodiment of the present invention, the first coupler 1 is a 1x2 coupler, the second coupler 2 is 80:20, and 80% of the light is retained in the loop, and the third coupler 3 is a 2x2 coupler.
[0077] The laser providing nonlinear response used in the embodiment of the present invention is a DFB laser, and the laser providing nonlinear response in the reserve pool can be a semiconductor laser, a fiber laser, a VCSEL laser, or any other nonlinear device that can provide nonlinearity to the reserve pool. As a response laser (that is, a nonlinear node) providing nonlinear response, its main purpose is to make different input signals have different transient responses after passing through the nonlinear node, and its main function is similar to the activation function in a traditional artificial neural network. The modulator uses a Mach-Zehnder modulator, and the narrowband laser uses a tunable laser.
[0078] According to the structure proposed in this scheme, the rate equation for the optical signal output from the second coupler 3 port is established as follows:
[0079]
[0080] +k inj E inj exp(2iπΔν)+F(t)
[0081]
[0082] Where E represents the complex amplitude of the slowly varying electric field of the laser providing the nonlinear response, N represents the average carrier density, α represents the linewidth enhancement factor, g is the differential gain, N0 represents the transparent carrier density, ε is the gain saturation coefficient, k represents the optical feedback intensity of the feedback loop, v represents the frequency of the laser providing the nonlinear response when it is running freely, Δv represents the frequency detuning between the narrowband laser and the laser providing the nonlinear response, τ p and τ s represent the photon lifetime and carrier lifetime respectively, J is the injection current of the laser providing nonlinear response, It represents the phase difference between the feedback field of the delayed feedback loop and the field cavity in the laser at the feedback injection point. γ is the coupling ratio of the third coupler. F(t) is the noise term.
[0083] The instrument parameters used are as follows:
[0084] α=3.0,g=8.4x10 -13 m 3 s -1 ,N0=1.4x10 24 m -3 ,ε=2.0x10 -23 ,τp =1.927ps,
[0085] τ s =2.04ns,k inj =12.43ns -1 ,v=1.96x10 14 Hz,Δv=-4.0GHz,J=1.037x10 33 m - 3 s -1 .
[0086] By changing the coupling ratio γ, the feedback strength in the fiber ring can be adjusted. By adjusting the injection intensity and feedback intensity, the connection weights in the reserve pool network structure can be adjusted and reshaped, thereby changing the structural composition of the neural network and increasing the system's adjustability and adaptability to tasks. By adding a fiber ring, when the coupling ratio is 50:50, the light is output from the response laser, passes through the second coupler port 2, and then directly outputs from port 1 through the third coupler port 4. The time it takes to pass through the first coupler port 3 is 40.05ns. When the fiber ring delay is 1ns, its NMSE reaches 4.75%
[0087] The application of the present invention realizes the complex neural network topology by introducing an optical fiber ring on the basis of a single feedback loop, which enables multiple feedback loops with better performance to be realized in hardware, greatly simplifies the difficulty of device implementation, and enhances the predictive performance of the reserve pool.
[0088] Practical application method (forecasting the Santa Fe time series)
[0089] First, the SantaFe time series prepared for training is multiplied by the prepared mask. This step is equivalent to the input weight in the traditional artificial neural network. The signal after the mask multiplication is generated by an arbitrary waveform generator, modulated into an optical signal by a modulator, and passed through the reservoir layer. It is converted into an electrical signal by the photodetector in the output layer, and the signal output from the reservoir layer is received and recorded by the digitizer. The output signal (that is, the part of the data called activation value or feature representation in the artificial neural network) is trained by the ridge regression algorithm to obtain the weight of the output layer. After obtaining the trained output weights, you can start to actually invest in the prediction task. After the parameter data of the new SantaFe time series enters the reservoir through the same operation as the training data, the new data is received in the digitizer. These data are directly weighted and summed with the trained output weights to get the required prediction results.
[0090] The present invention can also further enrich the connection state of the node and change its complexity by changing the number and connection mode of the optical fiber ring. The specific method is as follows:
[0091] 1. Use fiber optic loops in a feedback loop to delay the light multiple times.
[0092] 2. In the feedback loop, the network state of the reservoir layer is changed by cascading fiber rings to complete a more complex connection matrix.
[0093] 3. In the feedback loop, the network state of the reservoir layer is changed by nesting fiber rings to complete a more complex connection matrix.
[0094] 4. In the feedback loop, the network state of the reservoir layer is changed by simultaneously cascading and nesting fiber rings to complete a more complex connection matrix.
[0095] The specific connection method is as follows Figure 4 and attached Figure 5 shown.
Claims
1. A multi-feedback optical reservoir system based on an optical fiber ring, comprising an input layer, a reservoir layer and an output layer connected in sequence, wherein the input layer is used to multiply a sample signal with a mask signal, modulate it on the light emitted by a narrow-band laser through a modulator, and the modulated light emitted by the narrow-band laser enters the reservoir layer, characterized in that: The reservoir layer is used to provide a laser feedback loop including a nonlinear response, and a fiber ring is used in the feedback loop to perform multiple optical delays.
2. The multi-feedback optical reservoir system according to claim 1, characterized in that: The modulated light output from the input layer enters the reservoir layer through the first polarization controller (PC1); in the reservoir layer, the light coming out of the first coupler enters port 1 of the circulator, and enters the laser providing nonlinear response from port 2 of the circulator, the output light of the laser providing nonlinear response returns to port 2 of the circulator, the light coming out of port 3 of the circulator passes through the second polarization controller (PC2) and then enters the second coupler, the light coming out of port 2 of the second coupler passes through the attenuator (VA) and enters the optical fiber ring, and the light comes out of the optical fiber ring and returns to the first coupler again to form a feedback loop, and the light coming out of port 3 of the second coupler enters the output layer.
3. The multi-feedback optical reservoir system according to claim 1, characterized in that: The optical fiber ring is a single optical fiber ring or a nested optical fiber ring, or a cascade and combination of one or two optical fiber rings, and the number of cascaded and combined optical fiber rings is more than two.
4. The multi-feedback optical reservoir system according to claim 3, characterized in that: The single optical fiber ring includes a third coupler and a section of feedback optical fiber connected between two ports of the third coupler.
5. The multi-feedback optical reservoir system according to claim 3, characterized in that: The nested fiber ring is a nest of two single optical fibers, including a third coupler, a fourth coupler and three sections of feedback optical fiber. The third coupler is connected between the first coupler and the attenuator (VA). Two of the three sections of feedback optical fiber are respectively connected between the third coupler and the fourth coupler, and the third section of feedback optical fiber is connected between two ports of the fourth coupler.
6. The multi-feedback optical reservoir system according to claim 1, characterized in that: Let γ be the coupling ratio of the third coupler. By changing the coupling ratio γ, the feedback strength in the optical fiber ring is adjusted. By adjusting the feedback strength, the connection weights in the reserve pool network structure are adjusted and reshaped, thereby changing the structural composition of the neural network.
7. The multi-feedback optical reservoir system according to claim 1, characterized in that: In the input layer, the modulator is one of a Mach-Zehnder modulator and a phase modulator.
8. The multi-feedback optical reservoir system according to claim 1, characterized in that: The output layer consists of: The optical amplifier, photodetector and digitizer are connected in sequence, and the digital signals are collected through the output layer for training the output weights.
9. The multi-feedback optical reservoir system according to claim 1, characterized in that: Narrowband lasers are tunable lasers; lasers that provide nonlinear response are semiconductor lasers, fiber lasers or VCSEL lasers.