Depth time delay reserve pool calculation method based on electric cascade photoelectric feedback laser

Abstract

The invention discloses a depth time delay reserve pool calculation method based on an electric cascade photoelectric feedback laser. An adopted depth time delay RC system comprises a T-shaped bias device, a semiconductor laser and a photoelectric detector. A signal of the input layer is injected into the first-layer storage pool through the T-shaped bias device; an optical signal output by a laser 1 is divided into three paths after passing through a photoelectric detector, one path returns to the laser 1 through a feedback time delay loop, one path is injected into a second-layer storage pool through a T-shaped bias device, and the other path reaches an output layer to obtain a virtual node state of a first layer; an optical signal output by the laser 2 is divided into two paths after passing through a photoelectric detector, one path returns to the laser 2 through a feedback time delay loop, and the other path reaches an output layer to obtain a virtual node state of a second layer; and finally, linearly weighting the virtual nodes of the two layers of reserve pools to obtain the output of the system. According to the invention, real layering, high dynamic adjustability and excellent integration feasibility of the system structure are realized, and the system is more suitable for processing complex tasks in the real world.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of time delay reservoir computing, and particularly relates to a deep time delay reservoir computing method based on an electrical cascade optoelectronic feedback laser. BACKGROUND

[0002] ANN algorithms have been widely used in complex information processing tasks such as image recognition, speech recognition, time series prediction and channel equalization. Among them, RNN with short-term memory capability performs particularly outstanding in time sequence related problems such as video, speech and text. However, RNN also faces some challenges in practical application, such as gradient disappearance or gradient explosion problem, and dependence on a large number of matrix multiplication operations. These factors result in low efficiency in long sequence training and high requirements for digital computing resources. To overcome the above problems, RC algorithm is proposed. RC can be regarded as a simplification of RNN, and its characteristics are that the connection weight between the input layer and the reservoir layer is a fixed random value, and only the weight of the output layer needs to be trained, so the training process is greatly simplified. At the same time, RC has the advantages of simple structure and efficient training, and is particularly suitable for physical implementation. In recent years, photonic RC based on optical implementation has attracted widespread attention due to its high speed, strong parallelism, high energy efficiency, and the potential to break through the limitations of von Neumann architecture.

[0003] At present, the physical implementation of RC mainly includes two ways: traditional RC and time delay RC. The reservoir of traditional RC is composed of multiple nonlinear physical nodes, and forms a complex network through spatial interconnection. This structure needs to control the state, connection relationship and output collection of a large number of nodes when implemented in hardware. With the increase of the number of nodes, the hardware system complexity increases exponentially, which limits the construction of large-scale RC network. To solve this problem, time delay RC architecture is proposed. Time delay RC uses a nonlinear node with delay feedback loop to build the reservoir, and replaces a large number of physical nodes in traditional RC by introducing virtual nodes in the time dimension, thereby greatly reducing the hardware complexity. In this architecture, the total delay time of the feedback loop is , by setting equal interval time , there are virtual nodes ( ). The input signal is mapped to each virtual node through a mask , which is composed of uniformly distributed random values within . The reservoir layer contains at least one nonlinear node and one delay feedback loop, and the output layer samples the system response at time interval to obtain the state of each virtual node. Finally, the output weight of the output layer is trained, and the output result of the RC system is obtained by linearly weighting and summing all virtual node states and weights.

[0004] In practical applications, RC systems require stronger representational and temporal modeling capabilities, which places higher demands on their scalability (width and depth). Deep structures can introduce multiple time scales and frequency components, enhancing the system's ability to express complex dynamic features. Existing research has shown that cascading multi-layer reservoirs or multi-stage delay structures can significantly improve RC performance. Semiconductor lasers, with their advantages of high-speed response, strong nonlinearity, compact structure, and ease of integration, are ideal platforms for constructing deep time-delay RC systems.

[0005] In the technical field of deep time-delay RC systems, the closest to this invention is a deep time-delay RC solution based on optical feedback from a semiconductor laser. This solution is an algorithmic deep time-delay RC system formed in the data processing logic. Its core idea is to simulate multi-layer reservoir behavior through multiple rounds of input-output operations in a time-delay system composed of a single semiconductor laser and an optical feedback loop. The input signal is multiplied by a mask signal to modulate the intensity of the light emitted by the driving laser and injected into the reservoir laser with optical feedback. The time waveform output by the reservoir laser is sampled at fixed time intervals, and the resulting series of intensity values ​​are considered as virtual nodes. The number of virtual nodes is determined by the ratio of the feedback delay time to the sampling interval. The virtual node states of each layer are weighted and linearly summed to obtain the output of that layer, which is then used as the input signal for the next layer and injected into the same laser to repeat the process, thus realizing a logically hierarchical deep structure. A schematic diagram of this solution is shown below. Figure 2 As shown.

[0006] The implementation steps of this scheme are as follows: The input signal is multiplied by the first mask and then injected into the first-layer reservoir. The weighted linear sum of the virtual node states in the first-layer reservoir serves as the output signal of the first-layer reservoir. Subsequently, the output signal of the first layer is used as the input to the second-layer reservoir and multiplied by the second mask before being injected into the second-layer reservoir. The output of the second layer is also obtained by weighted linear summation of the virtual node states and is further used as the input signal of the third layer. This process is repeated sequentially to form a multi-layered progressive structure. Finally, the output of the entire deep reservoir system is obtained by linearly weighted summation of the virtual node states of all layers.

[0007] The existing technical solution is an algorithmic deep time-delay RC system based on optical feedback of a semiconductor laser. Although this solution has a simple structure and low hardware requirements, it has obvious disadvantages in several key aspects. First, this solution only simulates the behavior of multiple layers of reservoirs through multiple rounds of input-output operations, lacks decoupling and independent dynamic evolution between real physical layers, and cannot achieve deep modeling of multiple time scales. Second, all "layers" share the same laser, and its nonlinear dynamic capability is limited by the performance bottleneck of a single device, making it difficult to meet the needs of complex timing tasks. Therefore, although algorithmic deep RC has certain advantages in system simplification, it still has shortcomings in dynamic expression capability, system scalability, and physical authenticity.

[0008] Terminology:

[0009] Artificial Neural Network (ANN): Artificial Neural Network is a computational model inspired by the biological nervous system, composed of a large number of simple computational units (neurons), which simulate the way the brain processes information through connections and weights. It can learn the relationship between input and output through training, and is commonly used for function approximation, classification, and prediction tasks.

[0010] Recurrent Neural Network (RNN): Recurrent Neural Network is a type of neural network used to process sequential data. Its core feature is the introduction of "recurrent connections", allowing the network to utilize past state information when processing current input, thus possessing short-term memory capability. It is suitable for data analysis that relies on context, such as speech, text, and time series.

[0011] Reservoir Computing (RC): Reservoir Computing is an efficient neural network architecture derived from RNN. It divides the network into three parts: an input layer, a fixed randomly connected dynamic network (i.e. "reservoir") and a trainable output layer. Since the internal connections of the reservoir do not need to be trained, RC significantly reduces the training complexity.

[0012] Traditional RC: RC algorithm with a spatially nonlinear network as the reservoir simplifies the training process of the network and can be implemented in physical hardware.

[0013] Time-Delay RC: Time-Delay RC is a computational model that uses a single nonlinear node and a structure with delay feedback loops to construct a reservoir. By expanding the dynamic response of the system in time, it simulates the behavior of multiple virtual nodes, replacing the large network structure in traditional RC. This method greatly reduces the difficulty of implementing traditional RC in hardware.

[0014] Deep Time-Delay RC: Deep Time-Delay RC is an extension of a single Time-Delay RC, which builds a deep structure by stacking multiple reservoir layers, thereby enhancing the model's feature extraction and timing modeling capabilities.

[0015] Optoelectronic feedback: Optoelectronic feedback refers to the conversion of part of the optical signal output by the semiconductor laser into an electrical signal through a photodetector, and the feedback to the driving circuit of the laser after modulation or delay processing, thereby affecting its dynamic behavior.

[0016] Feedback delay time: The time required for a signal to circulate in a feedback loop. SUMMARY

[0017] To solve the above problems, the application provides a deep time delay reservoir pool calculation method based on an electrically cascaded optoelectronic feedback laser.

[0018] The deep time delay RC system of the deep time delay reservoir pool calculation method based on an electrically cascaded optoelectronic feedback laser includes a T-shaped biasing device, a semiconductor laser, and a photodetector.

[0019] The input signal is multiplied by a mask in the input layer to obtain a masked electrical signal, and the masked electrical signal and the bias current are injected into the pump current of the semiconductor laser 1 through the T-shaped biasing device, so as to inject the input layer information into the first layer reservoir pool; in the first layer reservoir pool, the optical signal emitted by the semiconductor laser 1 is converted into an electrical signal through a photodetector and is divided into three paths: one of the electrical signals is injected into the pump current of the semiconductor laser 1 through a feedback delay loop, one of the electrical signals is injected into the pump current of the semiconductor laser 2 through the T-shaped biasing device and the bias current, and the other electrical signal is sent to the output layer to obtain the virtual node state of the first layer; in the second layer reservoir pool, the optical signal emitted by the semiconductor laser 2 is converted into an electrical signal through a photodetector and is divided into two paths: one of the electrical signals is injected into the pump current of the semiconductor laser 2 through a feedback delay loop, and the other electrical signal is sent to the output layer to obtain the virtual node state of the second layer; in the output layer, the electrical signals output by each layer are equally spaced sampled to obtain the virtual node state of each layer, and finally the virtual nodes of each layer are added and linearly weighted to obtain the output result of the deep time delay RC.

[0020] Further, the deep time delay RC system is constructed by the optoelectronic feedback and electrical cascade of the semiconductor laser, and the theoretical model is described by the following rate equation:

[0021]

[0022] Wherein: represents a slowly varying complex electric field, represents the corresponding carrier density; is a line width enhancement factor, is a differential gain coefficient, is a transparent carrier number, is a gain saturation factor, is a photon lifetime, is the carrier lifetime, is the threshold current density, is the output light intensity of the laser when it is free running, is the normalized bias current, is the feedback strength, is the feedback delay time; is the Langevin term, which models the inherent spontaneous emission noise of the laser, satisfying the following statistical properties: where denotes the ensemble average, denotes the noise intensity; here equals 1 and 2, corresponding to laser 1 and laser 2, respectively; for the 1st layer, is the information injection term, representing the electrically injected mask signal; where, is the normalized input signal, is the input scaling factor; the mask has a period equal to the clock time , i.e. where, denotes the number of virtual nodes in the first layer, denotes the time interval between adjacent virtual nodes in the first layer; for the 2nd layer, is the coupling term, representing the electrical signal transferred from the first layer to the second layer, where is the coupling strength, is the coupling delay time; in the output stage, the light intensity of the first layer laser is sampled at time intervals to form the virtual node state ; the light intensity of the second layer laser is sampled at time intervals to form the virtual node state ; the output weights are trained and optimized by the ridge regression algorithm; the final calculation result is obtained by weighted summation: .

[0023] The beneficial technical effects of the present application compared with the prior art are:

[0024] This invention constructs a deep time-delay reservoir structure based on cascaded opto-feedback semiconductor lasers, where each reservoir layer consists of a semiconductor laser with an independent opto-feedback loop. State transfer between layers is achieved via electrical signals, forming a true physical hierarchical network. This scheme not only possesses high modularity and scalability in structure but also avoids dependence on the stability of the light injection phase through opto-feedback, significantly improving the system's stability and robustness. The dynamic parameters of each laser layer can be independently adjusted, giving the entire system stronger dynamic adaptability and feature extraction capabilities. In summary, this invention, while maintaining the advantages of deep RC computing, further achieves true hierarchical structure, high dynamic adjustability, and excellent integration feasibility, making it an ideal implementation path for high-performance neural optical computing platforms. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the deep time delay RC of the cascaded opto-feedback semiconductor laser of the present invention.

[0026] Figure 2 This is a schematic diagram of the deep delay RC of the traditional algorithm. Detailed Implementation

[0027] The present invention will be further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0028] This invention provides a method for calculating a deep time-delay reservoir based on an electrically cascaded opto-feedback laser, employing a deep time-delay RC system such as... Figure 1 As shown, the system includes a T-type bias, a semiconductor laser, and a photodetector. The input signal is injected into the first-layer reservoir through the T-type bias. Laser 1, acting as a nonlinear node in the first-layer reservoir, outputs a light signal that is split into three paths after passing through the photodetector: one path returns to laser 1 via a feedback delay loop, another path is injected into the second-layer reservoir via the T-type bias, and the third path goes to the output layer to obtain the virtual node state of the first layer. Laser 2, acting as a nonlinear node in the second-layer reservoir, outputs a light signal that is split into two paths after passing through the photodetector: one path returns to laser 2 via a feedback delay loop, and the other path goes to the output layer to obtain the virtual node state of the second layer. Finally, the virtual nodes of the two reservoirs are linearly weighted to obtain the system output.

[0029] The deep time delay RC of the application is electric information injection and photoelectric feedback, and layers are electrically cascaded. The implementation process is as follows: the input signal is multiplied by the mask in the input layer to obtain a masked electric signal, and the masked electric signal and the bias current are injected into the pump current of the semiconductor laser 1 through the T-type biasing device, so that the input layer information is injected into the first layer reservoir. In the first layer reservoir, the light signal emitted by the semiconductor laser 1 is converted into an electric signal by passing through a photodetector and is divided into three paths: one of the electric signals is injected into the pump current of the semiconductor laser 1 through the feedback time delay loop, one is injected into the pump current of the semiconductor laser 2 through the T-type biasing device and the bias current, and the other is to the output layer to obtain the virtual node state of the first layer. In the second layer reservoir, the light signal emitted by the semiconductor laser 2 is converted into an electric signal by passing through a photodetector and is divided into two paths: one of the electric signals is injected into the pump current of the semiconductor laser 2 through the feedback time delay loop, and the other is to the output layer to obtain the virtual node state of the second layer. In the output layer, the electric signals output by each layer are equally spaced sampled to obtain the virtual node state of each layer, and finally the virtual nodes of each layer are added and linearly weighted to obtain the output result of the deep time delay RC.

[0030] Compared with the deep time delay RC based on the cascaded injection-locked semiconductor laser, the deep time delay RC system of the application is constructed by the way of photoelectric feedback and electrical cascade of semiconductor lasers. Unlike optical feedback, photoelectric feedback does not depend on the phase of the optical field, so the system is not sensitive to phase disturbance and has higher stability and robustness. Since it does not depend on external injection laser source and complex optical alignment structure, the application has obvious advantages in system integration and is suitable for realizing high-density and integrated optical computing hardware. Further, the theoretical model of the application is described by the following rate equation:

[0031]

[0032] Wherein: represents a slowly varying complex electric field, represents the corresponding carrier density; is a linewidth enhancement factor, is a differential gain coefficient, is a transparent carrier number, is a gain saturation factor, is a photon lifetime, is a carrier lifetime, is a threshold current density, is the output light intensity when the laser is running freely, is a normalized bias current, is a feedback strength, is a feedback delay time; is the Langevin term, which models the inherent spontaneous emission noise of the laser, satisfying the following statistical properties: where denotes the ensemble average, denotes the noise intensity; here equals 1 and 2, corresponding to laser 1 and laser 2, respectively; for the 1st layer, is the information injection term, representing the electronically injected mask signal; where is the normalized input signal, is the input scaling factor; the mask has a period equal to the clock time , i.e. where denotes the number of virtual nodes in the 1st layer, denotes the time interval between adjacent virtual nodes in the 1st layer; for the 2nd layer, is the coupling term, representing the electrical signal transferred from the 1st layer to the 2nd layer, where is the coupling strength, is the coupling delay time; in the output stage, the optical intensity of the 1st layer laser is sampled at time intervals to form the virtual node state ; the optical intensity of the 2nd layer laser is sampled at time intervals to form the virtual node state ; the output weights are trained and optimized by the ridge regression algorithm; the final calculation result is obtained by weighted summation: .

[0033] Although the traditional algorithmic deep RC scheme realizes a similar multi-layer structure computing function through time multiplexing and data concatenation, all operations of its "layers" are completed in the same laser and feedback loop, lacking physical coupling, which leads to the system unable to form a real deep structure with multiple time scales and independent dynamic responses. In addition, the inter-layer connection relies on software weighting implementation, lacking a physical inter-layer coupling mechanism, which limits its ability in complex time series feature modeling and dynamic hierarchical expression.

[0034] This invention achieves deep time-delay RC (Recursive Conversion) based on cascaded opto-feedback semiconductor lasers. It constructs a physical hierarchy through opto-feedback of multiple semiconductor lasers, introducing an opto-feedback mechanism in each stage to form multiple layers of reservoirs with independent nonlinear dynamic response capabilities, thereby enabling efficient modeling and prediction of complex time series. Unlike existing deep RC algorithms that rely on a single laser and simulate multi-layer behavior through time multiplexing and sequential propagation of input and output, lacking physical layer separation and independent dynamic evolution, this invention achieves true inter-layer dynamic coupling and heterogeneous modeling capabilities through cascaded physical devices and feedback control, making it more suitable for handling complex real-world tasks.

Claims

1. A method for calculating a deep time delay reservoir based on an electrically cascaded optoelectronic feedback laser, characterized in that, The deep time-delay RC system used includes a T-type bias, a semiconductor laser, and a photodetector; The input signal is multiplied by the mask in the input layer to obtain the masked electrical signal. The masked electrical signal and the bias current are injected together into the pump current of the semiconductor laser 1 through the T-type biaser, thereby injecting the input layer information into the first layer reservoir. In the first-layer reservoir, the optical signal emitted by semiconductor laser 1 is converted into an electrical signal by a photodetector and divided into three paths: one electrical signal is injected into the pump current of semiconductor laser 1 through a feedback delay loop, another is injected into the pump current of semiconductor laser 2 through a T-type biaser and the bias current, and the third path goes to the output layer to obtain the virtual node state of the first layer; in the second-layer reservoir, the optical signal emitted by semiconductor laser 2 is converted into an electrical signal by a photodetector and divided into two paths: one electrical signal is injected into the pump current of semiconductor laser 2 through a feedback delay loop, and the third path goes to the output layer to obtain the virtual node state of the second layer; In the output layer, the electrical signals output from each layer are sampled at equal intervals to obtain the virtual node state of each layer. Finally, the virtual nodes of each layer are added together and linearly weighted to obtain the output result of the depth delay RC.

2. The method for calculating the deep time delay reservoir based on an electrically cascaded optoelectronic feedback laser according to claim 1, characterized in that, A deep time-delay RC system is constructed by photoelectric feedback and electrical cascading of semiconductor lasers. The theoretical model is described by the following rate equation: ； in: This represents a slowly changing complex electric field. This represents the corresponding carrier density; Linewidth enhancement factor The differential gain coefficient, Transparent carrier number, This is the gain saturation factor. For photon lifetime, For carrier lifetime, For the threshold current density, The intensity of the light output by the laser when it is running freely. For normalized bias current, For feedback strength, For feedback delay time; This is the Langevin term, used to simulate the inherent spontaneous emission noise of the laser, which satisfies the following statistical properties: ,in Represents the set average. Indicates noise intensity; here The values ​​1 and 2 correspond to laser 1 and laser 2, respectively; for the first layer, This is an information injection item, representing a mask signal injected electrically; among which, It is the normalized input signal. It is the input scaling factor; mask The period is equal to the clock time. ,Right now ,in, This indicates the number of virtual nodes in the first layer. This represents the time interval between adjacent virtual nodes in the first layer; for the second layer, It is a coupling term, representing the electrical signal transmitted from the first layer to the second layer, where It is the coupling strength. It is the coupling delay time; in the output stage, the light intensity of the first-layer laser. By time interval Sampling to form virtual node states The light intensity of the second-layer laser By time interval Sampling to form virtual node states Output weights The algorithm is trained and optimized using ridge regression; the final result is obtained by weighted summation. .

Images