A deep reservoir optical calculation method and system based on semiconductor lasers

Abstract

One technical solution of the present invention is to provide a deep reservoir optical computing method based on semiconductor lasers, and another technical solution of the present invention is to provide a deep reservoir optical computing system. In the present invention, each layer of the reservoir is mainly composed of a semiconductor laser and an optical delay line, and a large number of virtual neurons are generated thereby. The output of the laser in the upper layer of the reservoir is unidirectionally injected into the laser of the lower layer of the reservoir through the optical injection locking technology, thereby realizing the deep architecture of the reservoir computing. The depth of the reservoir optical computing system proposed by the present invention is not limited, so it has good scalability. The connection between the reservoir layers is an all-optical connection, so it has the advantages of simple device, low cost and low energy consumption.

Description

Technical Field

[0001] The present invention relates to a hardware implementation method and system for deep reservoir optical computing, and belongs to the intersecting field of machine learning and semiconductor optoelectronics. Background Art

[0002] The reservoir computing network is a special recurrent neural network. Unlike traditional recurrent neural networks, the connection weights of neurons in the input layer and the reservoir layer (i.e., the hidden layer) of the reservoir computing network do not require any training, but are set to random fixed values. Only the weights of the output layer of the reservoir computing network need to be trained, and in most cases only a simple linear regression is required to obtain the desired target value. Therefore, reservoir computing has the advantages of low training cost and fast training speed. Typical reservoir computing networks include echo-state networks (Echo-state network [Jaeger and Haas, Science 304, 78 (2004)]) and liquid state machines (Liquid state machine, Maass, and Markram, Neural Computation 14, 2531 (2002). Reservoir computing networks excel at processing time-series-related tasks, including but not limited to speech recognition, nonlinear channel equalization, and complex time series prediction such as chaotic sequences. Software-based reservoir computing networks typically run on general-purpose computing platforms such as CPUs or GPUs.

[0003] Thanks to the fixed weights of the input layer and the reservoir layer, the network architecture of reservoir computing is very conducive to its hardware implementation. The core devices used in verified hardware reservoir computing networks include but are not limited to memristors, spintronic devices, optical modulators, and semiconductor lasers. Reservoir optical computing networks based on semiconductor lasers have the advantages of high speed and low latency. Generally, large-scale reservoir computing networks require the use of a large number of the above-mentioned devices to be implemented. However, this implementation method poses huge challenges to the integration and cost of devices in reality. In 2011, researchers such as Appeltant proposed a system composed of nonlinear devices with delay lines that can generate a large number of virtual neurons under dynamic working conditions [Appeltant, Nature Communications 2, 468 (2011)]. Based on this principle, researchers successfully verified a reservoir optical computing network based on semiconductor lasers and optical delay lines [Brunner, Nature Communications 4, 1364 (2013)]. However, current semiconductor laser-based reservoir optical computing networks typically have only a few hundred virtual neurons [Chembo, Chaos 30, 013111 (2020)], which is insufficient for real-world complex tasks. To increase the "width" of reservoir computing networks, that is, to increase the number of neurons in the reservoir, researchers have proposed parallel solutions, including wavelength division multiplexing [Tang, IEEE Journal of Quantum Electronics 58, 8100109 (2022)]. On the other hand, handling complex tasks also requires increasing the "depth" of the reservoir computing network, that is, increasing the number of reservoir layers. Theoretical studies have shown that multi-layered deep reservoir computing networks offer significant advantages over single-layer reservoir computing networks in terms of neuronal dynamics richness and storage capacity, and the former demonstrates significantly superior performance on numerous benchmark tasks [Gallicchio, Micheli, and Pedrelli, Neurocomputing 268, 87 (2017)]; Goldmann, Chaos 30, 093124 (2020)]. However, there is currently a lack of hardware implementation solutions for deep reservoir optical computing based on semiconductor lasers.

[0004] The traditional single-layer reservoir computing network architecture based on semiconductor lasers is as follows: Figure 1As shown, the preprocessed input data is injected into the semiconductor laser through the optical injection locking technology. The laser output by the semiconductor laser is fed back into the laser through the optical delay line part (i.e., optical feedback [Tang, IEEE Journal of Quantum Electronics 58, 8100109 (2022)]. This system generates a large number of virtual neurons under stable and dynamic working conditions. By observing the state of the neurons and applying appropriate weights, the target value of the output layer is finally obtained. In most cases, the weights of the output layer can be optimized by linear regression or logistic regression. Summary of the Invention

[0005] The purpose of the present invention is to solve the depth problem of the reservoir light calculation.

[0006] In order to achieve the above object, the present invention provides a method for calculating the depth of a reservoir based on a semiconductor laser, which is characterized by comprising the following steps:

[0007] Step 1: The preprocessed input data is injected into the semiconductor laser of the first layer reservoir. Part of the light from the semiconductor laser is output to the optical delay line, which provides optical feedback for the semiconductor laser and M1 virtual neurons for the first layer reservoir.

[0008] Step 2: Set n = 1;

[0009] Step 3: Another portion of the light from the semiconductor laser in the nth layer reservoir is unidirectionally injected into the semiconductor laser in the n+1th layer reservoir through optical injection locking technology;

[0010] Step 4: Part of the light from the semiconductor laser in the n+1 layer storage pool is output to the optical delay line, which provides optical feedback for the semiconductor laser and M for the n+1 layer storage pool. (n+1) virtual neurons;

[0011] Step 5: n=n+1. If n≥N, proceed to step 6; otherwise, return to step 3.

[0012] Step 6: Collect the states of all virtual neurons in the N-layer reserve pool, N ≥ 2, and combine them with the weights of the output layer to obtain the desired target value.

[0013] Preferably, in step 3, when part of the light from the semiconductor laser in the nth layer reservoir is unidirectionally injected into the semiconductor laser in the n+1th layer reservoir through the optical injection locking technology, the optical injection locking parameters operate in a stable locking range surrounded by the Hopf bifurcation and the saddle-node bifurcation.

[0014] The technical solution of the present invention is to provide a deep reservoir optical computing system based on the aforementioned deep reservoir optical computing method, characterized in that it is composed of N layers of reservoirs, each layer of reservoirs including a semiconductor laser and an optical delay line. Part of the light from the semiconductor laser in each layer of reservoirs is output to the optical delay line, which provides optical feedback for the semiconductor laser and provides multiple virtual neurons for the current layer of reservoirs.

[0015] In the N-layer reservoir, except for the semiconductor laser of the first-layer reservoir whose input is preprocessed input data, for the remaining N-1-layer reservoirs, part of the light from the semiconductor laser of the previous layer of reservoir is unidirectionally injected into the semiconductor laser of the current layer of reservoir through optical injection locking technology, thereby constructing a deep reservoir optical computing network architecture with N layers of reservoir in a cascade manner, N≥2.

[0016] Preferably, the number of virtual neurons belonging to each layer of the reserve pool is the same or different.

[0017] The solution disclosed in this invention not only increases the depth of the reservoir, but also increases the total number of virtual neurons in the entire network or improves the processing rate of input data. The connection between reservoirs at different levels in this solution is an all-optical connection, requiring no photoelectric or electro-optical conversion processes. Therefore, it has the advantages of simple device, low energy consumption, low latency, and high speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 The traditional semiconductor laser-based single-layer reservoir computing network architecture is demonstrated;

[0019] Figure 2 The deep reservoir optical computing network architecture proposed by the present invention is demonstrated;

[0020] Figure 3 The figure shows an embodiment of the deep reservoir optical calculation proposed by the present invention;

[0021] Figure 4 The diagram shows the implementation effect of the deep reserve pool optical calculation proposed in the present invention. DETAILED DESCRIPTION

[0022] Below in conjunction with specific embodiment, further set forth the present invention.Should be understood that these embodiments are only used to illustrate the present invention and are not used in limiting the scope of the present invention.In addition, should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall equally within the scope limited by the appended claims of the application.

[0023] like Figure 2As shown in the figure, the deep reservoir optical computing network architecture proposed by the present invention consists of multiple layers of reservoirs. Each layer of reservoirs is mainly composed of semiconductor lasers and optical delay lines. The semiconductor lasers provide optical feedback through the optical delay lines and provide virtual neurons for each layer of reservoirs. Figure 2 In the deep reservoir optical computing network architecture shown, the output signal of the first layer of reservoir is injected into the laser of the second layer of reservoir through optical injection locking technology. The optical injection locking parameters operate in a stable locking range surrounded by Hopf bifurcation and saddle-node bifurcation [Tang, IEEE Journal of Quantum Electronics 58, 8100109 (2022)]. Similarly, the output signal of the second layer of reservoir is injected into the laser of the third layer of reservoir through optical injection locking technology. In this way, a deep reservoir optical computing network architecture with multiple layers of reservoirs can be constructed in a cascade manner through the technology of optical injection locking.

[0024] Figure 3 A specific embodiment of the present invention is shown in FIG. Figure 3 In the scheme shown, laser 1, circulator and optical fiber delay line 1 constitute the first layer of reserve pool, laser 2, circulator and optical fiber delay line 2 constitute the second layer of reserve pool, and laser 3, circulator and optical fiber delay line 3 constitute the third layer of reserve pool.

[0025] Preprocessed data is injected into Laser 1 through the circulator's P1 port. Laser 1's output is connected to a fiber-optic delay line 1 (made of optical fibers) through the circulator's P3 port. The delayed laser light is then fed back into Laser 1 through the circulator again. Laser 1 provides optical feedback via fiber-optic delay line 1 and contributes virtual neurons to the first reservoir layer. Meanwhile, a portion of Laser 1's output is injected into Laser 2 through the circulator in the second reservoir layer using optical injection locking. Laser 2 provides optical feedback via fiber-optic delay line 2 and contributes virtual neurons to the second reservoir layer. Similarly, a portion of Laser 2's output is injected into Laser 3 through the circulator in the third reservoir layer using optical injection locking. Simultaneously, Laser 3 provides optical feedback via fiber-optic delay line 3 and contributes virtual neurons to the third reservoir layer. Finally, the desired target value is obtained by collecting the states of all virtual neurons in the three reservoir layers and combining them with the weights of the output layer.

[0026] Based on the above technical principles, a four-layer reservoir optical computing network was constructed. The total number of virtual neurons in the four-layer reservoir optical computing network is 400, that is, the number of neurons in each layer of the reservoir is 100. The total number of virtual neurons in the single-layer reservoir optical computing network is also 400. Under this condition, the input data processing speed of the four-layer reservoir computing network is 4 times faster than that of the single-layer reservoir computing network. Figure 4 As shown in the figure, the linear storage capacity of the former is much greater than that of the latter under different optical feedback ratios. In addition, the performance of the deep reservoir optical computing network on other standard test tasks such as chaotic sequence prediction and nonlinear channel equalization is significantly better than that of the traditional single-layer reservoir optical computing network.

Claims

1. A method for calculating the depth of a reservoir based on a semiconductor laser, characterized in that: The following steps are involved: Step 1: The preprocessed input data is injected into the semiconductor laser of the first layer reservoir. Part of the light from the semiconductor laser is output to the optical delay line, which provides optical feedback for the semiconductor laser and M1 virtual neurons for the first layer reservoir. Step 2: Set n = 1; Step 3: Another portion of the light from the semiconductor laser in the nth layer reservoir is unidirectionally injected into the semiconductor laser in the n+1th layer reservoir through optical injection locking technology; Step 4: Part of the light from the semiconductor laser in the n+1 layer storage pool is output to the optical delay line, which provides optical feedback for the semiconductor laser and M for the n+1 layer storage pool. (n+1) virtual neurons; Step 5: n=n+1. If n≥N, proceed to step 6; otherwise, return to step 3. Step 6: Collect the states of all virtual neurons in the N-layer reserve pool, N ≥ 2, and combine them with the weights of the output layer to obtain the desired target value.

2. A semiconductor laser-based depth reservoir optical calculation method according to claim 1, characterized in that: In step 3, when part of the light from the semiconductor laser in the nth layer reservoir is unidirectionally injected into the semiconductor laser in the n+1th layer reservoir through the optical injection locking technology, the optical injection locking parameters operate in a stable locking range surrounded by the Hopf bifurcation and the saddle-node bifurcation.

3. A depth reserve pool optical calculation system implemented based on the depth reserve pool optical calculation method according to claim 1, characterized in that: It consists of N layers of reservoirs, each of which includes a semiconductor laser and an optical delay line. Part of the light from the semiconductor laser in each layer of the reservoir is output to the optical delay line, which provides optical feedback for the semiconductor laser and provides multiple virtual neurons for the current layer of the reservoir. In the N-layer reservoir, except for the semiconductor laser of the first-layer reservoir whose input is preprocessed input data, for the remaining N-1-layer reservoirs, part of the light from the semiconductor laser of the previous layer of reservoir is unidirectionally injected into the semiconductor laser of the current layer of reservoir through optical injection locking technology, thereby constructing a deep reservoir optical computing network architecture with N layers of reservoir in a cascade manner, N≥2.

4. A deep reservoir optical computing system as claimed in claim 3, characterized in that: The number of virtual neurons belonging to the reserve pools in each layer is the same or different.

Images