Optical diffractive neural network with nonlinear activation

Abstract

The application discloses an optical diffraction neural network with nonlinear activation, and belongs to the field of optical computing.The structure comprises an N*1 MZI array, a flat optical waveguide, an input light source, a control light source, a photoelectric detector and a control unit.Input light is input to the N*1 MZI array, and a data signal is loaded on a phase shifter electrode.The input light is modulated and then enters the flat optical waveguide.The phase change material and the nonlinear material are covered on the flat optical waveguide to form a layer of the neural network.When the light passes through the phase change material region, it is equivalent to performing a matrix multiplication operation, and when the light passes through the nonlinear material region, it is equivalent to running a nonlinear activation function.Setting multiple layers of the above structure, combining the control light source and the gradient descent algorithm, and purposefully erasing and writing the phase change material, a multi-layer diffraction neural network with small floor area, low power consumption, reconfiguration and nonlinear activation is realized, and all operations of the multi-layer neural network are realized in the optical domain.

Description

Technical Field

[0001] This invention belongs to the field of optical computing, and more specifically, relates to an optical diffraction neural network with nonlinear activation. Background Technology

[0002] Due to the immense convenience brought by artificial intelligence, it has gradually become one of the most watched fields. Neural networks are one of the most important components, with wide applications in classification, recognition, detection, and other fields. Matrix multiplication and nonlinear activation are its main operations, consuming most of the computing resources. However, with the rapid development of artificial intelligence, the scale of network models is increasing rapidly, and the demand for computing resources and energy efficiency in model training is also increasing. As Moore's Law is gradually approaching its limits, the slowdown in the growth rate of computing power and the "power wall" problem make traditional electronic chips unable to meet today's needs. Optical computing, with its low power consumption, high bandwidth, and multi-dimensional (wavelength, mode, and polarization) multiplexing physical characteristics, has become a feasible alternative to electrical computing.

[0003] To realize optical neural networks, the currently widely adopted approach for matrix computation is to build the network structure using a Mach-Zehnder interferometer (MZI) or a micro-ring resonator (MRR) array based on wavelength division multiplexing (WDM). However, optical neural networks built using these methods cannot perform nonlinear activation function operations, meaning that regardless of network size, they remain linear systems capable only of matrix multiplication. For AI (Artificial Intelligence) applications, nonlinearity is crucial, enabling neural networks to fit complex functions and achieve diverse functionalities. To compensate for this deficiency, some network structures convert optical signals into electrical signals, perform nonlinear activation function operations in the electrical domain, and then convert the electrical signals back into optical signals. However, this approach presents several problems. First, the light-to-electricity-to-light conversion increases power consumption and system complexity. Furthermore, the introduction of electrons limits the overall system bandwidth and slows down system performance. This approach, which utilizes a combination of optical and electrical domains to perform all neural network operations, does not fully leverage the advantages of optical computing; many performance characteristics remain limited by electronic properties. Summary of the Invention

[0004] To address the shortcomings of existing structures in terms of neural network scale and nonlinear activation capability, this invention proposes an optical diffraction neural network with nonlinear activation, aiming to achieve a low-footprint, low-power, and high-speed all-optical neural network. This network structure can execute nonlinear activation function operations with extremely high efficiency.

[0005] To achieve the above objectives, this invention proposes an optical diffraction neural network with nonlinear activation, comprising an N×1 MZI array, a planar optical waveguide, a phase change material, a nonlinear material, an input light source, a control light source, a three-dimensional displacement platform, a photodetector, and a control unit.

[0006] The N×1 MZI array is a signal modulation structure, with each MZI including two 3dB couplers and a phase shifter. The phase shifter can be based on the thermo-optic effect, consisting of electrodes and a silicon waveguide. Heating the silicon waveguide through the electrodes changes its effective refractive index. Alternatively, it can be based on the plasmon dispersion effect, heterogeneously integrating graphene, ITO, or other high-performance materials onto the silicon waveguide. Applying a voltage alters the carrier concentration within the structure, thus affecting the waveguide's effective refractive index. During signal loading, the FPGA applies the signal to the phase shifter electrodes, changing the waveguide's effective refractive index and consequently altering the phase change. As the light passes through the couplers, interference occurs between the two arms, converting phase information into intensity information, thereby achieving the modulation of the input light.

[0007] The input end of the planar optical waveguide is an optical wave input end, which receives the signal light modulated by the N×1MZI array, and the output end is connected to the input end of the photodetector for detecting the optical power of the light wave output from the planar optical waveguide.

[0008] The phase change material (such as Sb2Se3) is a long rectangular strip covering the flat optical waveguide. The state of the phase change material can be controlled by an externally controlled laser to realize matrix operations in the optical domain. Other materials can also be used to replace it, but they must be able to locally change the effective refractive index of the optical waveguide without introducing additional losses, such as metal thermoelectric electrodes.

[0009] The nonlinear material is germanium (Ge), which is partially covered on the optical waveguide. Different intensities of input light result in different complex effective refractive indices, enabling nonlinear operation in the optical domain. Furthermore, this material is compatible with current CMOS processes.

[0010] The input light source is located at the input end of the N×1MZI array and is used to provide a light source with adjustable wavelength and power for the entire structure.

[0011] The controlled light source is placed above the phase change material to irradiate it, causing a change in the crystallization state of the phase change material, resulting in different refractive indices, thereby controlling the light propagation state.

[0012] The three-dimensional displacement platform is installed above the non-volatile phase change material and can move precisely in the X, Y, and Z directions to adjust the control light source to a suitable position.

[0013] The photodetector is placed at the output end of the planar optical waveguide, converts the output light into an electrical signal, and then transmits the detection signal to the FPGA.

[0014] The control unit includes an FPGA chip, a digital-to-analog converter (DAC), and an analog-to-digital converter (ADC). The DAC converts data into analog signals, which are applied to the N×1 MZI electrodes to modulate the input light. The photodetector transmits the detection results to the ADC, which converts the analog electrical signal into a digital signal. Then, the FPGA chip executes a gradient descent algorithm and, based on the results, controls the control light source and the three-dimensional displacement platform to erase and write the phase change material at a specific location.

[0015] In summary, compared with existing technologies, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0016] 1. By depositing phase change materials and nonlinear materials on planar optical waveguides, all-optical matrix operations and nonlinear activation function operations can be performed on a smaller footprint without the introduction of electrical components, thereby increasing the operating bandwidth and reducing power consumption.

[0017] 2. Due to the non-volatility of phase change materials, no static voltage needs to be applied after the optical diffraction neural network has been trained. Therefore, during the inference process, energy is only consumed during signal loading, detection, and processing, significantly reducing operating costs.

[0018] 3. Based on the gradient descent algorithm and the erasable and rewritable properties of phase change materials, the entire structure can be rapidly reconstructed and trained online.

[0019] 4. This structure has a high tolerance for process errors, and the materials used in the structure are CMOS process compatible, making the structure easy to mass-produce. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of an optical diffraction neural network with nonlinear activation, provided as an example of the present invention.

[0021] Figure 2 This invention provides an example of light transmission through a planar optical waveguide covered with a nonlinear material in an optical diffraction neural network with nonlinear activation.

[0022] Figure 3 This is a schematic diagram illustrating the nonlinearity of germanium material in an optical diffraction neural network with nonlinear activation, as provided as an example of the present invention.

[0023] Figure 4 This is a schematic diagram illustrating the operation of an optical diffraction neural network with nonlinear activation, provided as an example of the present invention.

[0024] Figure 5 This is a flowchart illustrating an optical diffraction neural network with nonlinear activation, provided as an example of the present invention.

[0025] Figure 6 The loss function iteration curve of an optical diffraction neural network with nonlinear activation is provided as an example of the present invention.

[0026] Figure 7 This provides partial numerical simulation results for an optical diffraction neural network with nonlinear activation, as an example of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0028] like Figure 1 As shown in the example, the present invention proposes an optical diffraction neural network with nonlinear activation. The system includes: an N×1 MZI array 1, a planar optical waveguide 2, a phase change material 3, a nonlinear material 4, an input light source 5, a control light source 6, a three-dimensional displacement platform 7, a photodetector 8, and a control unit 9.

[0029] The optical signal, used as the input signal of this invention, is generated by the input light source 5 and input to the N×1 MZI array 1. The input light source 5 has one output port, which can transmit the light into the N×1 MZI array 1 through a 1×N beam splitter. The planar optical waveguide 2 has N input ports and is connected to the N×1 MZI array 1. By preprocessing the dataset, the 28*28 images in the dataset are mapped into an N×1 vector, where each element corresponds to the voltage applied to the phase shifter electrode of the N×1 MZI array 1, and the mapping range is determined by the range of voltages that can be applied to the electrode. The light after passing through the N×1 MZI array 1 is input to the planar optical waveguide 2. Multiple rows of elongated rectangular phase change materials 3 and an elongated rectangular nonlinear material 4 constitute a layer in the neural network. The portion of the planar optical waveguide 2 covered with the phase change material 3 constitutes the matrix operation section. By changing the phase transition state of the phase change material 3, different propagation paths of light in the planar optical waveguide 2 can be altered, enabling different matrix operation functions. The portion of the planar optical waveguide covered with the nonlinear material 4 constitutes the nonlinear activation operation section. Light of different intensities will have different effective complex refractive indices after passing through the nonlinear material, meaning the output response changes nonlinearly with the input light intensity. The entire planar optical waveguide 2 can include multiple layers of the above structure, thereby achieving the purpose of realizing a multilayer neural network.

[0030] When light passes through the portion of the planar optical waveguide 3 covered by the nonlinear material 4, the light transmission is as follows: Figure 2 As shown, light is coupled from the planar optical waveguide 3 into the nonlinear material 4, and then coupled back. The intensity and attenuation of light transmitted through germanium vary. For example... Figure 3 As shown, the output light intensity gradually decreases as the light intensity increases. After transmission, it is equivalent to performing a nonlinear function operation. In this example, the length of the germanium material is only 1.3 μm, which means that a planar optical waveguide combined with a nonlinear material can perform nonlinear function operations with extremely high efficiency.

[0031] After being modulated by the planar optical waveguide 2, the light is detected by the photodetector 8 to obtain the output result. The number of output channels is determined by the classification task requirements. This example uses a handwritten digit training set, requiring 10 output channels. The overall process overview is as follows: Figure 4 As shown.

[0032] In this example, the specific process of calculating the gradient is as follows:

[0033] After the preprocessed data is loaded onto the electrodes of the N×1MZI array 1, the control light source 6 is moved by the three-dimensional displacement platform 7 to perturb the corresponding strip-shaped phase change material 3 (slightly increasing the area Δx occupied by the crystalline state of the strip-shaped phase change material). The output results of the photodetector 8 before and after the perturbation are respectively... The control unit 9 first preprocesses the data according to the following formula:

[0034]

[0035] The control unit 9, based on the processed... Calculate the corresponding loss function value L2

[0036]

[0037] Based on the gradient value, the area occupied by the crystalline state of the phase change material 3 is changed according to the following formula.

[0038]

[0039] Where α is the learning rate, and if the value is negative, it corresponds to an increase in the area occupied by the amorphous state.

[0040] After loading multiple data sets, the calculated gradients are summed and averaged. Then, the phase transition state of the phase transition material 3 is changed using the three-dimensional displacement platform 7 and the control light source 6, completing one iteration. Repeating the above operations completes the model training. The overall operation process is as follows: Figure 5 As shown.

[0041] In this example, the specific operation to change the state of the phase change material 3 is as follows:

[0042] The three-dimensional moving platform 7 is controlled to move the control light source 6 directly above the phase change material 3 that needs to be changed. Based on the gradient value calculated by the control unit 9, the pulse intensity and duration of the light pulse from the control light source 6 are controlled so that the phase change material 3 reaches different phase change states. Once the state of the non-volatile phase change material is determined, it can remain unchanged for a long time, avoiding additional power consumption.

[0043] Numerical simulations were performed based on the above example, and the curve of the loss function is shown below. Figure 6 As shown, the loss function value gradually decreases with increasing iterations and stabilizes after approximately 100 iterations, indicating that the structure can be trained online and reconstructed quickly. Some of the trained results are shown below. Figure 7 As shown, this structure can implement the MNIST ten-class task. However, it should be noted that only a two-layer structure was constructed in this numerical simulation. More complex inference tasks can be achieved by increasing the number of layers.

[0044] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An optical diffraction neural network with nonlinear activation, characterized in that, include: N×1MZI array, planar optical waveguide, input light source, control light source, photodetector, control unit; The input light emitted by the input light source is input to the N×1 MZI array. The N×1 MZI array includes N MZIs, each MZI including two MMIs and two straight waveguides, one of which is covered with a thermoelectric electrode to form a phase shifter. The N×1 MZI array is used to load the data signal onto the phase shifter to modulate the input light and generate signal light; N is a positive integer greater than 1. The input end of the planar optical waveguide receives signal light, and the output end is connected to the input end of the photodetector. The photodetector is used to detect the optical power of the light wave output from the planar optical waveguide. The planar optical waveguide includes multiple repeating units, each of which includes a phase transition region and a nonlinear region. The phase transition region is covered with a phase transition material, and the nonlinear region is covered with a nonlinear material. The control light source is used to control the state of the phase transition material, thereby controlling the transmission state of light in the phase transition region and realizing matrix operations in the optical domain. The nonlinear region is used to realize nonlinear operations in the optical domain. The photodetector is placed at the output end of the planar optical waveguide and transmits the detection signal to the control unit. The control unit is used to execute the gradient descent algorithm to control the electrodes of the phase shifter and update the state of the phase change material.

2. The optical diffraction neural network with nonlinear activation according to claim 1, characterized in that, The data signal is mapped to an N×1 vector, where each element corresponds to the voltage applied to the electrode of the phase shifter in the N×1 MZI array, and the mapping range is determined by the range of voltages that can be applied to the electrode.

3. The optical diffraction neural network with nonlinear activation according to claim 1, characterized in that, It also includes a three-dimensional moving platform, which moves in the X, Y, and Z directions to move the control light source above the phase change material.

4. The optical diffraction neural network with nonlinear activation according to claim 1, characterized in that, The controlled light source is used to change the phase transition state of the phase change material. By outputting light pulses of different intensities and durations, the phase change material can achieve different degrees of crystallization, including crystalline, amorphous, and multiple amorphous states.

5. An optical diffraction neural network with nonlinear activation according to claim 1, characterized in that, The phase change material is in the form of long rectangular strips covering the flat optical waveguide, with multiple long rectangular strips of phase change material placed in parallel in each column.

6. The optical diffraction neural network with nonlinear activation according to claim 1, characterized in that, The control unit includes an FPGA chip, a digital-to-analog converter (DAC), and an analog-to-digital converter (ADC). The DAC converts the data signal into an analog signal and applies it to the electrodes of the phase shifter. The ADC converts the received analog signal into a data signal for processing. The FPGA chip encodes the data signal into voltage information, executes a gradient descent algorithm to control the electrodes of the phase shifter, and updates the state of the phase change material.

7. An optical diffraction neural network with nonlinear activation according to claim 1, characterized in that, The phase change material is Sb2Se3.

8. An optical diffraction neural network with nonlinear activation according to claim 1, characterized in that, The nonlinear material is germanium (Ge).

Images