A topology reconfigurable photonic reservoir computing method based on multiple optical feedback loops

Abstract

The application discloses a kind of topology reconfigurable photon reserve pool computing methods based on multiple optical feedback loop, including input mapping, optical injection and multiple feedback reservoir evolution, virtual node sampling, task adaptation configuration and linear readout training five main parts;By introducing multiple configurable optical feedback branches outside the reservoir semiconductor laser, the fixed virtual node connection relationship originally determined by single time delay is expanded into a reconfigurable connection network determined by multiple time delays;By adjusting the number of feedback branches, the minimum feedback delay and the time delay interval, the adaptation configuration of different task information structures is realized, so that the same reserve pool hardware can form a more suitable dynamic memory depth and feature mixing mode for different tasks.

Description

Technical Field

[0001] This invention belongs to the field of photonic information processing and neuromorphic computing, and particularly relates to a computational method for a topologically reconfigurable photonic reservoir based on multiple optical feedback loops. Background Technology

[0002] Pooled computation is a type of neural network computation framework with low training costs. Its basic structure includes an input layer, a reservoir layer, and an output layer. Unlike traditional recurrent neural networks, pooled computation typically trains only the output layer weights, utilizing a fixed nonlinear dynamic system to complete high-dimensional mapping and temporal memory. Therefore, it has high application value in tasks such as high-speed signal processing, chaotic prediction, speech recognition, and image classification.

[0003] In photonic storage, semiconductor lasers have become an important implementation platform due to their advantages such as high bandwidth, strong nonlinearity, and ease of integration. In particular, time-delay storage structures can utilize a single nonlinear optical node with feedback combined with time-division multiplexing to construct a large number of virtual nodes, thereby avoiding the use of a large number of physical neurons and facilitating high-speed, low-power information processing.

[0004] In existing photonic delay reservoirs based on a single feedback loop, the connectivity between virtual nodes is primarily determined by the single feedback delay and node spacing, resulting in a fixed ring or nearest-neighbor coupling structure in the equivalent topology. Once the single feedback delay is determined, the coupling pattern between reservoir virtual nodes is also fixed, making it difficult to flexibly change according to the information structure of specific tasks. For long-memory chaotic prediction tasks, more cross-period, non-local state coupling is required; for short-memory nonlinear tasks, more compact local coupling is needed; and for serialized image classification tasks, it is often necessary to recover the correlation between adjacent columns in the original two-dimensional space. Fixed topologies cannot simultaneously accommodate these differentiated requirements.

[0005] Some studies have attempted to improve performance by adding parallel reservoirs, constructing deep reservoirs, or introducing multiple feedback mechanisms. However, most schemes still treat multiple feedback merely as a means to enhance dynamic complexity, rather than explicitly defining it as a physical degree of freedom for controlling the topology of virtual nodes. Furthermore, a unified implementation method that can adaptively configure feedback topologies for different tasks is lacking. Therefore, there is an urgent need to propose a method that can reconstruct the virtual topology of photon reservoirs at the hardware level and optimize it according to different task requirements. Summary of the Invention

[0006] To address the above problems, this invention provides a method for calculating a topologically reconfigurable photon reservoir based on multiple optical feedback loops.

[0007] This invention discloses a computational method for a topologically reconfigurable photonic reservoir based on multiple optical feedback loops, comprising five main parts: input mapping, optical injection and multi-feedback reservoir evolution, virtual node sampling, task adaptation configuration, and linear readout training. The specific steps are as follows:

[0008] Step 1: Preprocess the input signal u(t).

[0009] Multiplying the original input by the mask signal Mask(t) yields the weighted input S(t) = Mask(t). u(t) is used to expand the low-dimensional input into a higher-dimensional dynamic state space.

[0010] Step 2: The weighted input S(t) is loaded onto the injection light via a modulator and injected into the semiconductor laser that serves as the reservoir core; a portion of the output from the reservoir laser is used for detection and sampling, and the other portion is injected back into the laser after entering M parallel optical feedback branches, thus forming a multi-optical feedback reservoir.

[0011] Step 3: Set the feedback delay τ for the m-th feedback branch. m The preferred configuration for each feedback delay is a linear sequence: τ m = τ min + (m-1)∆τ, m=1,2,...,M; by adjusting the minimum feedback delay τ min The time delay interval ∆τ between adjacent feedback branches directly changes the equivalent coupling offset between virtual nodes, thereby realizing the hardware reconstruction of the reservoir virtual topology.

[0012] Step 4: Time-division multiplexing sampling of the reservoir continuous response is performed according to the node interval θ, in each sampling period T=N R ∙N is obtained within θ R The state of each virtual node; arbitrary feedback delay τ m All are mapped to τ m ≈q m ∙T + D m ∙θ, where q m D represents the number of complete sampling periods spanned. m This represents the virtual node offset within a sampling period; thus, each feedback branch introduces a directed coupling path from a different historical position to the current virtual node, and multiple feedbacks together form a reconfigurable virtual connection network.

[0013] Step 5: Denote the sampled reservoir state matrix as X, the target output as Y, and use ridge regression to solve for the output layer weights W. out After training, the linear readout layer is used to predict or classify the signal to be tested.

[0014] Furthermore, the reservoir laser satisfies the extended Lang-Kobayashi dynamic equations, which include multiple feedback terms, and its complex electric field amplitude E(t) and carrier density N(t) are expressed as:

[0015] ;

[0016] ;

[0017] Where α is the linewidth enhancement factor, G N Let N0 be the differential gain, N0 be the transparent carrier density, ε be the gain saturation coefficient, and τ be the differential gain. p and τ s These represent the photon lifetime and the carrier lifetime, respectively, k in For injection intensity, k feedback and τ m These represent the feedback strength and feedback delay of the m-th feedback branch, respectively. For the optical signal modulated by S(t), ω0, ω s These are the operating angular frequencies of the reservoir laser and the main laser, respectively.

[0018] Through the superposition of multiple feedback items, the current state is coupled with multiple historical states simultaneously.

[0019] The method of this invention can be adapted to different information structures for different tasks, and can be used for chaotic time series prediction, short memory nonlinear system modeling, and Deccan task image classification.

[0020] The beneficial technical effects of this invention compared to the prior art are as follows:

[0021] (1) By using multiple optical feedback loops, the fixed ring topology can be extended into a reconfigurable virtual topology, and the connection relationship of virtual nodes can be directly adjusted at the hardware level.

[0022] (2) The ability to select different feedback delay configurations based on the information structure of different tasks, thereby enabling the same reserve pool hardware to adapt to different tasks, is one of the important innovations of this invention.

[0023] (3) The reservoir state complexity, nonlinear mapping capability and time memory capability can be improved without significantly increasing the number of physical lasers.

[0024] (4) Compared with parallel reservoir and deep reservoir schemes, the system structure is more compact, the feedback loop is easier to implement, and it has better potential for miniaturization and integration. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall structure and virtual topology of the topologically reconfigurable photon reservoir based on multi-optical feedback loops according to the present invention.

[0026] Figure 2 The figure shows the impact of total feedback strength on time series forecasting performance under different feedback loop configurations.

[0027] Figure 3 A graph showing the performance differences for different tasks under different latency intervals (used to illustrate the optimal topology-dependent task type). Detailed Implementation

[0028] The present invention will be further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0029] This invention discloses a topologically reconfigurable photonic reservoir computation method based on multiple optical feedback loops. On one hand, by introducing multiple configurable optical feedback branches outside the reservoir semiconductor laser, the fixed virtual node connection relationship originally determined by a single time delay is expanded into a reconfigurable connection network jointly determined by multiple time delays. On the other hand, by adjusting the number of feedback branches, the minimum feedback time delay, and the time delay interval, adaptive configuration for different task information structures is achieved, enabling the same reservoir hardware to form a more suitable dynamic memory depth and feature mixing mode for different tasks. The overall structural principle is as follows: Figure 1 As shown, it includes five main parts: input mapping, optical injection and multi-feedback reservoir evolution, virtual node sampling, task adaptation configuration, and linear readout training. The specific steps are as follows:

[0030] Step 1: Preprocess the input signal u(t).

[0031] Multiplying the original input by the mask signal Mask(t) yields the weighted input S(t) = Mask(t). u(t) is used to expand the low-dimensional input into a higher-dimensional dynamic state space.

[0032] Step 2: The weighted input S(t) is loaded onto the injection light via a modulator and injected into the semiconductor laser that serves as the reservoir core; a portion of the output from the reservoir laser is used for detection and sampling, and the other portion is injected back into the laser after entering M parallel optical feedback branches, thus forming a multi-optical feedback reservoir.

[0033] Step 3: Set the feedback delay τ for the m-th feedback branch. m The preferred configuration for each feedback delay is a linear sequence: τ m = τ min + (m-1)∆τ, m=1,2,...,M; by adjusting the minimum feedback delay τ min The time delay interval ∆τ between adjacent feedback branches directly changes the equivalent coupling offset between virtual nodes, thereby realizing the hardware reconstruction of the reservoir virtual topology.

[0034] Step 4: Time-division multiplexing sampling of the reservoir continuous response is performed according to the node interval θ, in each sampling period T=N R ∙N is obtained within θ R The state of each virtual node; arbitrary feedback delay τ m All are mapped to τ m ≈q m ∙T + D m ∙θ, where q m D represents the number of complete sampling periods spanned. m This represents the virtual node offset within a sampling period; thus, each feedback branch introduces a directed coupling path from a different historical position to the current virtual node, and multiple feedbacks together form a reconfigurable virtual connection network.

[0035] Step 5: Denote the sampled reservoir state matrix as X, the target output as Y, and use ridge regression to solve for the output layer weights W. out After training, the linear readout layer is used to predict or classify the signal to be tested.

[0036] Furthermore, the reservoir laser satisfies the extended Lang-Kobayashi dynamic equations, which include multiple feedback terms, and its complex electric field amplitude E(t) and carrier density N(t) are expressed as:

[0037] ;

[0038] ;

[0039] Where α is the linewidth enhancement factor, G N Let N0 be the differential gain, N0 be the transparent carrier density, ε be the gain saturation coefficient, and τ be the differential gain. p and τ s These represent the photon lifetime and the carrier lifetime, respectively, k in For injection intensity, k feedback and τ m These represent the feedback strength and feedback delay of the m-th feedback branch, respectively.

[0040] Through the superposition of multiple feedback items, the current state is coupled with multiple historical states simultaneously.

[0041] Task adaptation strategy:

[0042] Different topology configuration strategies can be adopted for different tasks. For long-memory chaotic prediction tasks, it is preferable to use ∆τ, which is not an integer multiple of the sampling period T, to avoid feedback topological resonance and enhance non-redundant coupling, state diversity, and long-term memory capabilities. For short-memory nonlinear tasks, it is preferable to use a smaller ∆τ to form a more compact local coupling and highlight the influence of recent states. For serialized image classification tasks, it is preferable to set the feedback delay near an integer multiple of the scan period T to restore the spatial correlation between adjacent image columns.

[0043] Example:

[0044] 1. A simulation example for a long-memory chaotic time series prediction task. Taking the Santa Fe chaotic time series prediction task as an example, the number of virtual nodes N... R The optimal value is 100, the node spacing θ is preferably 10 ps, ​​and the sampling period T is approximately 1 ns. Simulation results are as follows: Figure 2 As shown, the prediction error changes with the number of feedback loops M and the total feedback strength, indicating that the number of feedback loops and the feedback strength jointly determine the reservoir topology characteristics and performance. For this type of long memory task, it is preferable to use τ min By setting the time to approximately 1.01 ns and setting ∆τ to approximately 1.65 ns, a non-resonant multi-delay topology can be obtained, thereby enhancing state complexity and long-term memory capability.

[0045] 2. Simulation Implementation Examples for Short-Memory Nonlinear Tasks. Taking the NARMA2 task as an example, under the condition of keeping the total feedback strength similar, ∆τ can be reduced to a range closer to local coupling, such as around 0.9 ns, so that adjacent virtual nodes form more compact and effective connections, improving the ability to represent short-term nonlinear dependencies. Combined with... Figure 3 It is evident that the optimal latency intervals for NARMA2 and Santa Fe tasks are different, indicating that there is no uniform optimal topology applicable to all tasks.

[0046] 3. Application example for serialized image classification tasks. Taking MNIST handwritten digit recognition as an example, the two-dimensional grayscale image can be first projected through a random mask matrix and then serialized column-wise before being input into the reservoir. Preferably, the number of virtual nodes N R With a value of 600 and a node spacing θ of 10 ps, ​​the corresponding scan period T is approximately 6 ns. In this case, a portion of the feedback delay can be set to a value close to an integer multiple of T, creating controlled coupling between different scan columns. This helps the reservoir recover the original two-dimensional spatial correlation of the image, thereby improving the classification performance under serial input conditions.

[0047] As can be seen from the above embodiments, the key to this invention does not lie in a fixed parameter value, but in utilizing multiple feedback loops and minimizing the feedback delay τ. minThe physical topology parameter set, consisting of the time delay interval ∆τ, is adjustable. Based on the information structure of different tasks, the virtual node connection method is actively matched to realize topology reconfigurability and photon reservoir computing adapted to different tasks.

Claims

1. A computational method for a topologically reconfigurable photon reservoir based on multiple optical feedback loops, characterized in that, It includes five main parts: input mapping, optical injection and multi-feedback reservoir evolution, virtual node sampling, task adaptation configuration, and linear readout training. The specific steps are as follows: Step 1: Preprocess the input signal u(t); Multiplying the original input by the mask signal Mask(t) yields the weighted input S(t) = Mask(t). u(t) is used to expand the low-dimensional input into a higher-dimensional dynamic state space. Step 2: The weighted input S(t) is modulated onto the injection light and injected into the semiconductor laser that serves as the reservoir core; A portion of the reservoir laser output is used for detection and sampling, while another portion is injected back into the laser after entering M parallel optical feedback branches, thus forming a multi-optical feedback reservoir. Step 3: Set the feedback delay τ for the m-th feedback branch. m The preferred configuration for each feedback delay is a linear sequence: τ m =τ min + (m-1)∆τ, m=1,2,...,M; by adjusting the minimum feedback delay τ min The time delay interval ∆τ between adjacent feedback branches directly changes the equivalent coupling offset between virtual nodes, thereby realizing the hardware reconstruction of the reservoir virtual topology; Step 4: Time-division multiplexing sampling of the reservoir continuous response is performed according to the node interval θ, in each sampling period T=N R ∙N is obtained within θ R The state of each virtual node; arbitrary feedback delay τ m All are mapped to τ m ≈q m ∙T + D m ∙θ, where q m D represents the number of complete sampling periods spanned. m This represents the virtual node offset within a sampling period; thus, each feedback branch introduces a directed coupling path from a different historical position to the current virtual node, and multiple feedbacks together form a reconfigurable virtual connection network. Step 5: Denote the sampled reservoir state matrix as X, the target output as Y, and use ridge regression to solve for the output layer weights W. out After training, the linear readout layer is used to predict or classify the signal to be tested.

2. The method for calculating a topologically reconfigurable photon reservoir based on a multi-optical feedback loop according to claim 1, characterized in that, The reservoir laser satisfies the extended Lang-Kobayashi dynamic equations, which include multiple feedback terms, and its complex electric field amplitude E(t) and carrier density N(t) are expressed as: ； ； Where α is the linewidth enhancement factor, G N Let N0 be the differential gain, N0 be the transparent carrier density, ε be the gain saturation coefficient, and τ be the differential gain. p and τ s These represent the photon lifetime and the carrier lifetime, respectively, k in For injection intensity, k feedback and τ m These represent the feedback strength and feedback delay of the m-th feedback branch, respectively. For the optical signal modulated by S(t), ω0, ω s These are the operating angular frequencies of the reservoir laser and the main laser, respectively. Through the superposition of multiple feedback items, the current state is coupled with multiple historical states simultaneously.

3. A method for calculating a topologically reconfigurable photon reservoir based on multiple optical feedback loops according to any one of claims 1-2, characterized in that, This method can be adapted to different information structures for various tasks, including chaotic time series prediction, short-memory nonlinear system modeling, and image classification.

Images