A laser neuron based on polarization-selective feedback vcsel

Abstract

The application discloses a laser neuron based on polarization selective feedback VCSEL, comprising a single VCSEL and a feedback loop, laser is emitted by the VCSEL, and after passing through an optical circulator, the laser is separated into orthogonal polarization modes and parallel polarization modes by a polarization beam splitter, wherein the orthogonal polarization modes are taken as the output of the laser neuron; the parallel polarization modes are selected to enter the feedback loop, are rotated by 90 degrees through a Faraday rotator mirror and are converted into orthogonal polarization directions, then are modulated by a Mach-Zehnder modulator to simulate neuron stimulation, are subjected to feedback intensity control through a variable optical attenuator, and finally are fed back into the orthogonal polarization modes of the VCSEL to simulate the behaviors of Tonic Spiking and Phasic Spiking neurons. The application only needs a single laser to simulate a neuron, and thus the hardware structure is simplified; there is no need to match the optical frequency offset and phase between lasers, and thus the stability of the system is ensured.

Description

Technical Field

[0001] This invention belongs to the field of laser neuron technology, and particularly relates to a laser neuron based on polarization-selective feedback VCSEL. Background Technology

[0002] Binary logic computing based on the traditional von Neumann architecture lags far behind the human brain in terms of solving complex problems and power consumption. Against this backdrop, neuromorphic computing paradigms that mimic the function of biological neurons have gradually emerged as an alternative. In the field of neuromorphic computing, the traditional approach utilizes the nonlinear effects of circuits to mimic neuronal behaviors such as impulses; however, electronic chip technology is still constrained by Moore's Law. Compared to electronic technology, photonics technology typically offers advantages in parallelism, high speed, high bandwidth, and low power consumption. VCSELs exhibit complex nonlinear dynamic characteristics such as excitability and inhibition during the injection-locking process, providing a faster timescale than electrical methods and offering the advantage of low input power. Therefore, utilizing the nonlinear characteristics of VCSELs to simulate the behavior of biological neurons and construct laser neurons has become one of the effective means to solve the current problems of low speed and high power consumption in the field of neuromorphic computing.

[0003] To address the two major bottlenecks in current computer technology—the low energy efficiency caused by the von Neumann architecture and the approaching limits of Moore's Law—scientists worldwide are exploring various methods, principles, devices, and architectures to create a new generation of computers with better performance and lower energy consumption. Neuromorphic computing, drawing inspiration from the structure of the brain's nervous system and its fundamental mechanisms for information processing, is fundamentally transforming existing computing architectures and systems at multiple levels, including hardware implementation and software algorithms, thereby achieving low-energy, high-performance computing systems.

[0004] Biological neurons exhibit rich and complex firing behaviors when stimulated by electrical current, including Phasic Spiking and Tonic Spiking. Phasic Spiking refers to a neuron's sudden large discharge after stimulation, followed by a sharp decrease in discharge volume, and finally a return to its original level. This response is considered a phased pulse and is often used to detect the onset of stimulation. Tonic Spiking refers to a neuron's sustained output of a sequence of pulses after stimulation, typically used to detect persistent input. These firing behaviors help us understand the interactions between neurons, thus enabling better simulation of large-scale cortical neural networks. Laser neurons are a basic unit of neuromorphic computing based on photonics technology. They can simulate the behavior of neurons in the human brain, using light / electricity to transmit signals and build complex neural networks to achieve advanced computational functions.

[0005] VCSELs are a novel type of semiconductor laser characterized by an emission beam direction perpendicular to the laser surface, supporting dual polarization modes. VCSELs offer advantages such as good stability, low power consumption, low cost, small size, ease of integration into 2D / 3D arrays, and high coupling efficiency between mode patterns and optical fibers. Thanks to these advantages, VCSELs are considered one of the most popular devices for realizing laser neurons. At certain operating points, external light injection into a VCSEL induces nonlinear dynamic phenomena such as mode switching and injection lock-in. Common VCSEL neurons are based on external light injection architectures, utilizing the VCSEL's response to orthogonal or parallel polarized light injection to simulate neuron-like behavior. However, external light injection requires an additional laser to provide the light source; therefore, a single neuron typically requires two lasers to function, leading to a large hardware scale problem when building complex neural network systems.

[0006] In the field of constructing laser neurons using laser nonlinear dynamics, the closest approach to this invention is the laser neuron technology based on optically injected VCSELs. This approach injects the output light from an externally tunable laser into the orthogonal polarization mode (perpendicular to the parallel polarization mode) of a VCSEL, inducing nonlinear dynamic phenomena such as polarization mode switching and injection locking, thereby generating neuron-like behavior. By adjusting the parameters of this optically injected VCSEL system, neuron-like behaviors such as tonic spinning and pulsed spinning can be obtained at the orthogonal polarization output. The device for this method is as follows: Figure 1 As shown. Figure 1 The scheme shown mainly consists of a tunable laser, a polarization controller, a Mach-Zehnder modulator, a stimulus signal generator, an optical circulator, and a vertical cavity surface-emitting laser (VCSEL).

[0007] The implementation steps of this scheme are as follows: First, the laser is emitted by a tunable laser. The stimulation signal generated by the stimulation signal generator is loaded onto the continuous light output by the tunable laser through a Mach-Zehnder modulator. After polarization tuning by a polarization controller, it is injected into the orthogonal polarization mode of the VCSEL. The light output by the VCSEL is separated into two mutually perpendicular polarization modes by a polarization beam splitter after passing through an optical circulator. The orthogonal polarization mode serves as the output of the laser's neurons.

[0008] Existing technologies rely on external optical injection-based kinetic excitation mechanisms, requiring a tunable laser as the external light source. Under suitable injection parameters (including injection frequency offset, injection intensity, and injection phase difference), nonlinear dynamics such as polarization mode switching and injection locking of the VCSEL can be induced, resulting in neuron-like responses output in orthogonal polarization modes. However, due to limitations of the optical injection mechanism, existing technologies using two lasers can only simulate a single neuron, leading to massive hardware requirements when constructing complex neural network systems. Furthermore, existing technologies depend on optical coupling for neuronal pulse signal transmission, placing stringent demands on the matching degree of optical frequency offset and phase between different lasers, posing challenges to system stability.

[0009] Terminology Explanation:

[0010] Neuromorphic computing: Using controllable spiking patterns to simulate neuronal behavior to achieve ultrafast neuromorphic computing for non-traditional computing applications.

[0011] Laser neurons: a new type of photonic neuron that uses light / electricity to transmit signals, mimicking the behavior of neurons in the human brain. It can build complex neural networks and achieve advanced computing functions.

[0012] Phasic Spiking: A neuronal behavior in which a neuron fires a large amount of electricity after being stimulated, then the amount of electricity drops sharply and eventually returns to the original level. This response is known as a phased pulse and is often used to detect the start time of stimulation.

[0013] Tonic Spiking: A neuronal behavior in which neurons continuously output a sequence of pulses after being stimulated, often used to detect the duration of stimulation.

[0014] Vertical-cavity surface-emitting laser (VCSEL): A new type of semiconductor laser characterized by the direction of the emitted beam being perpendicular to the surface of the laser.

[0015] Injection Lock: Injection lock is a special optical locking phenomenon that can occur when a laser is injected with an external light source, causing the output frequency and phase of the injected laser to synchronize with the external light source.

[0016] Feedback delay time: The time required for a signal to complete one cycle within the feedback loop.

[0017] Parallel / Orthogonal polarization: The polarization mode that dominates when the VCSEL is running freely is called parallel polarization, and the polarization perpendicular to it is called orthogonal polarization.

[0018] Polarization mode switching: The phenomenon that the polarization mode of the laser changes when it is subjected to external disturbances, as the light emitted by the VCSEL is polarized. Summary of the Invention

[0019] To address the above problems, this invention provides a laser neuron based on polarization-selective feedback VCSEL.

[0020] The present invention discloses a laser neuron based on polarization-selective feedback VCSEL, comprising a single VCSEL and a feedback loop, wherein the feedback loop includes an optical circulator, an optical fiber jumper, a polarization beam splitter, a Faraday rotator, a Mach-Zehnder modulator, a stimulus signal generator, and a variable optical attenuator.

[0021] The laser is emitted by the VCSEL and, after passing through an optical circulator, is separated into an orthogonal polarization mode and a parallel polarization mode by a polarization beam splitter. The orthogonal polarization mode serves as the output of the laser neuron. The parallel polarization mode is selected to enter the feedback loop, rotated 90° by a Faraday rotator to become orthogonal polarization, and then modulated by a Mach-Zehnder modulator to simulate neuron stimulation. The feedback intensity is then controlled by a variable optical attenuator and finally fed back into the orthogonal polarization mode of the VCSEL.

[0022] Under specific feedback parameters, the nonlinear dynamic state of feedback locking and polarization mode switching of VCSEL is used to simulate the behavior of Tonic Spiking and Phasic Spiking neurons.

[0023] Furthermore, it also includes a photodetector, which converts orthogonally polarized light signals into electrical signals for use in the output of laser neurons.

[0024] Furthermore, the above model simulating the neuronal behavior of Tonic Spiking and Phasic Spiking is described as follows:

[0025]

[0026]

[0027]

[0028]

[0029] Among them: E x / y γ is the complex amplitude of the slowly changing photoelectric field, with subscripts x and y representing orthogonal and parallel polarization modes, respectively; N is the normalized value of the total reversal rate of carriers between the conduction and valence bands, and n is the normalized difference between the reversal rates of sine and anti-sine carriers; κ is the photon decay rate, and γ... N It is the carrier decay rate, γs It is the spin homogenization rate, γ a It is linear cavity dichroism, γ p It is a linear birefringence, α is the linewidth enhancement factor, μ is the normalized injection current, and β is the linear birefringence. x / y =β1+iβ2 is the spontaneous emission noise of the laser, and the real part β1 and the imaginary part β2 satisfy <β1(t)β2(t′)>=0. The time self-reflection satisfies The relationship is given by Δν, where Δν is the full width at half maximum optical linewidth of the freely moving laser; the time correlation of spontaneous emission noise from two orthogonally polarized lasers satisfies... The feedback light is described by the second-to-last term in formula (1), where ξ s τ is the normalized stimulus intensity of the external stimulus, and τ is the feedback delay time.

[0030] The beneficial technical effects of this invention are as follows:

[0031] 1. This invention realizes nonlinear dynamic excitation of lasers based on polarization-selective feedback, avoiding the need for additional lasers in optical injection mechanisms. Thus, only a single laser is needed to simulate a neuron, simplifying the hardware structure.

[0032] 2. This invention uses a neuronal pulse transmission mechanism based on electrical signals, eliminating the need for matching optical frequency offset and phase between lasers, thus ensuring the stability of the system. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of a traditional laser neuron based on light-injected VCSELs.

[0034] Figure 2 This is a schematic diagram of the laser neuron based on polarization-selective feedback VCSEL of the present invention.

[0035] Figure 3 This is a schematic diagram of a laser neuron based on an electrically injected VCSEL. Detailed Implementation

[0036] The present invention will be further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0037] This invention discloses a laser neuron based on a polarization-selective feedback VCSEL, comprising a single VCSEL and a feedback loop. The feedback loop includes an optical circulator, fiber optic jumpers, a polarization beam splitter, a Faraday rotator, a Mach-Zehnder modulator, a stimulus signal generator, and a variable optical attenuator, enabling the selection, rotation, and feedback modulation of the feedback light's polarization. It also includes a photodetector to convert orthogonally polarized light signals into electrical signals for use in the laser neuron's output.

[0038] Unlike traditional optically injected VCSEL laser neurons, this invention provides a laser neuron based on polarization-selective feedback VCSELs, such as... Figure 2 As shown, external stimulation of the neuron is achieved by modulating the intensity of the optical feedback. The tonic and phasic-spiking responses of biological neurons are simulated using the feedback locking and polarization mode switching of a VCSEL. The implementation process is as follows: First, the laser is emitted from the VCSEL, passes through an optical circulator, and is separated into orthogonal and parallel polarization modes by a polarization splitter. The orthogonal polarization mode serves as the output of the laser neuron. The parallel polarization mode is selected to enter the feedback loop, rotated 90° by a Faraday rotator to become orthogonal polarization, and then modulated by a Mach-Zehnder modulator to simulate neuronal stimulation. The feedback intensity is then controlled by a variable optical attenuator, and finally, the feedback is injected back into the orthogonal polarization mode of the VCSEL. Under specific feedback parameters, a nonlinear dynamic state of self-injection locking and polarization mode switching can be excited, successfully simulating tonic and phasic-spiking neuronal behaviors.

[0039] Compared to traditional optically injected VCSEL laser neurons, this scheme requires only one laser to simulate one neuron, halving the network size. Furthermore, it eliminates the need for precise control over the frequency offset and phase of the light, resulting in a simpler and easier-to-implement neuron structure. The theoretical model of this scheme can be described by the following equation:

[0040]

[0041]

[0042]

[0043]

[0044] Among them: E x / y γ is the complex amplitude of the slowly changing photoelectric field, with subscripts x and y representing orthogonal and parallel polarization modes, respectively; N is the normalized value of the total reversal rate of carriers between the conduction and valence bands, and n is the normalized difference between the reversal rates of sine and anti-sine carriers; κ is the photon decay rate, and γ... N It is the carrier decay rate, γ s It is the spin homogenization rate, γ a It is linear cavity dichroism, γ p It is the linear birefringence, α is the linewidth enhancement factor, μ is the normalized injection current (μ = 1 is the threshold), β x / y =β1+iβ2 is the spontaneous emission noise of the laser, and the real part β1 and the imaginary part β2 satisfy <β1(t)β2(t′)>=0. The time self-reflection satisfies The relationship is given by Δv, where Δv is the full width at half maximum optical linewidth of the freely running laser; the time correlation of spontaneous emission noise from two orthogonally polarized lasers satisfies... The feedback light is described by the second-to-last term in formula (1), where ξ s τ is the normalized stimulus intensity of the external stimulus, and τ is the feedback delay time.

[0045] Traditional optically injected VCSEL neurons require an additional laser to provide an external injection induction light source. Therefore, simulating a single neuron requires at least two lasers, which presents a large hardware scale problem when building complex neural network systems. In addition, its neuronal pulse transmission mechanism depends on the optical coupling effect between lasers, so the requirements for the optical frequency offset and phase matching between different lasers are stringent, which limits the system stability.

[0046] Traditional optical injection into VCSEL neurons induces neuron-like behaviors by modulating externally injected light. In addition, there are also... Figure 3 The diagram illustrates an alternative approach to generating neuron-like behavior by modulating the injected current. Specifically, a tunable laser emits a constant, continuously orthogonally polarized beam that induces injection lock-in in the VCSEL. Light from another tunable laser is modulated by a Mach-Zehnder modulator, converted to the electrical domain by a photodetector, and finally injected into the VCSEL via a bias transistor. This approach utilizes a neuronal pulse propagation mechanism based on electrical signals, thus requiring stringent matching of optical frequency offset and phase when pulse propagation is absent. However, it requires two lasers as external light sources, meaning at least three lasers are needed to simulate a single neuron. This further exacerbates the problem of large hardware scale and high cost when constructing complex neural network systems.

[0047] This invention utilizes a classic feedback loop structure, employing optical feedback to rotate the parallel polarization mode emitted by the VCSEL itself into an orthogonal polarization mode before feeding it back to the laser. This excites the laser's polarization dynamics, simulating neuron-like behavior. On one hand, because the induction light source is achieved through self-emission feedback, the need for an additional laser is avoided; a single laser can simulate a single neuron, simplifying the neuronal architecture and halving the laser hardware overhead of the neural network. On the other hand, since an electrical signal-based neuronal pulse transmission mechanism is used, there is no requirement for matching the optical frequency offset and phase between lasers, ensuring system stability. Therefore, this invention is simpler, more operable, and less expensive.

Claims

1. A laser neuron based on polarization-selective feedback VCSEL, characterized in that, It includes a single VCSEL and a feedback loop, which includes an optical circulator, fiber optic patch cord, polarization beam splitter, Faraday rotator, Mach-Zehnder modulator, stimulus signal generator and variable optical attenuator; The laser is emitted by the VCSEL and, after passing through the optical circulator, is separated into an orthogonal polarization mode and a parallel polarization mode by the polarization beam splitter. The orthogonal polarization mode serves as the output of the laser neuron. The parallel polarization mode is selected to enter the feedback loop, rotated 90° by the Faraday rotator to become orthogonal polarization, and then modulated by the Mach-Zehnder modulator to simulate neuron stimulation. The feedback intensity is then controlled by the variable optical attenuator and finally fed back into the orthogonal polarization mode of the VCSEL. Under specific feedback parameters, the nonlinear dynamic state of feedback locking and polarization mode switching of VCSEL is used to simulate the behavior of Tonic Spiking and Phasic Spiking neurons. The model simulating the neuronal behavior of Tonic Spiking and Phasic Spiking is described as follows: ； ； ； ； Among them: E x / y It is the complex amplitude of the slowly changing photoelectric field, with subscripts x and y representing orthogonal and parallel polarization modes, respectively; N is the normalized value of the total reversal rate of carriers between the conduction and valence bands, and n is the normalized difference between the reversal rates of sine and anti-sine carriers. It is the photon decay rate. It is the carrier decay rate. It is the spin homogenization rate. It is linear cavity dichroism. It is a linear birefringence. It is the linewidth enhancement factor. It is the normalized injection current. This is the spontaneous emission noise of the laser, real part. and the virtual part satisfy Time self-consistency is satisfied Relationship, among which It is the full width of the half-maximum optical linewidth of a freely moving laser; the time correlation of spontaneous emission noise from two orthogonally polarized lasers satisfies The feedback light is described by the second-to-last term in formula (1), where τ is the normalized stimulus intensity of the external stimulus, and τ is the feedback delay time.

2. The laser neuron based on polarization-selective feedback VCSEL according to claim 1, characterized in that, It also includes photodetectors, which convert orthogonally polarized light signals into electrical signals for use in the output of laser neurons.

Images