Photoelectric neural network computing architecture based on periodic activation

Abstract

The invention discloses a photoelectric neural network computing architecture based on periodic activation. The photoelectric neural network computing architecture comprises a light source unit, a photoelectric weighted computing unit and a photoelectric nonlinear activation unit, the light source unit is used for splitting a light source and then outputting n paths of light signals; the photoelectric weighting calculation unit is used for receiving the n paths of optical signals output by the light source unit, modulating and weighting the n paths of optical signals, converting the n paths of optical signals into electric signals and outputting the electric signals so as to complete matrix weighting calculation of the photoelectric neural network; the photoelectric nonlinear activation unit is used for receiving the electric signal output by the photoelectric weighting calculation unit, performing nonlinear activation on the electric signal and then outputting a final electric signal so as to complete activation of the photoelectric neural network; wherein n is a positive integer, and n is greater than or equal to 2. According to the invention, the photoelectric neural network has stronger fitting and information expression capabilities.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and optical computing technology, and in particular to an optoelectronic neural network computing architecture based on periodic activation. Background Art

[0002] With the rapid development of artificial intelligence technology, the requirements for computing power of various deep learning models are also increasing. Traditional electronic hardware such as FPGA and GPU have been specifically used to accelerate neural network calculations, but due to the limitations of Moore's Law, the development of electrical hardware is still difficult to meet the growing demand for computing power. Optical neural networks have become a strong competitor in the field of high-performance artificial intelligence computing in the future with their low power consumption, high parallelism and bandwidth advantages. Although many advances have been made in the field of optical computing, existing research mainly focuses on convolution matrix operations, while relatively less attention has been paid to nonlinear activation functions.

[0003] Current research on optical activation functions generally uses optoelectronic devices to simulate traditional nonlinear activation functions. However, for complex artificial intelligence tasks, this simulation method faces huge challenges in power efficiency and manufacturing costs in the process of achieving precise alignment of optical devices with these nonlinear activation function curves. Therefore, it is crucial to explore and utilize the inherent properties of optoelectronic devices to build an optoelectronic nonlinear activation architecture that is both efficient and low-power.

[0004] In the field of information processing, photoelectric activation functions play a vital role. The repeatable modular units and scalability of their design play a significant role in meeting the growing computing needs. These characteristics make photoelectric activation functions particularly important in processing advanced computing tasks, especially in dealing with complex application scenarios.

[0005] It should be noted that the information disclosed in the above background technology section is only used for understanding the background of the present application, and therefore may include information that does not constitute prior art known to ordinary technicians in the field. Summary of the invention

[0006] In order to make up for the deficiencies of the prior art, the present invention provides an optoelectronic neural network computing architecture based on periodic activation, which can solve the problem.

[0007] The present invention adopts the following technical solution:

[0008] A photoelectric neural network computing architecture based on periodic activation includes a light source unit, a photoelectric weighted computing unit and a photoelectric nonlinear activation unit; the light source unit is used to output n optical signals after splitting the light source; the photoelectric weighted computing unit is used to receive the n optical signals output by the light source unit and modulate and weight them to convert them into electrical signals for output, thereby completing the matrix weighted calculation of the photoelectric neural network; the photoelectric nonlinear activation unit is used to receive the electrical signals output by the photoelectric weighted computing unit and perform nonlinear activation on the electrical signals to output the final electrical signals, thereby completing the activation of the photoelectric neural network; wherein n is a positive integer, and n≥2.

[0009] Furthermore, the photoelectric weighted calculation unit includes n photoelectric modulator systems and n first photodetectors; the n photoelectric modulator systems are respectively connected to the n first photodetectors, and the photoelectric modulator system is used to modulate and weight the n optical signals to output weighted signals; the first photodetector is used to sum the intensities of the weighted signals output by the photoelectric modulator system and convert the optical signals into electrical signals for output.

[0010] Furthermore, each optoelectronic modulator system includes multiple computing streams, each of which is composed of two first optoelectronic modulators and second optoelectronic modulators connected in series. For the first optoelectronic modulator and the second optoelectronic modulator on each computing stream, the optical signal output by the light source unit is input into the optical input end of the first optoelectronic modulator, the electrical signal waveform generated by the electrical signal source is input from the radio frequency end of the first optoelectronic modulator, and the input signal of the optoelectronic neural network is loaded at the radio frequency end of the first optoelectronic modulator, and the optical signal processed by the first optoelectronic modulator is input into the optical input end of the second optoelectronic modulator; the weight value of the optoelectronic neural network is loaded at the radio frequency end of the second optoelectronic modulator to output the weighted optical signal at the optical output end of the second optoelectronic modulator; in the same optoelectronic modulator system, the optical output ends of the second optoelectronic modulators on all computing streams are connected to the optical input end of a first photodetector, and the first photodetector is used to sum the intensity of the weighted optical signal and convert the optical signal into an electrical signal for output.

[0011] Furthermore, different weight values ​​are applied to the optical signal in each computing flow.

[0012] Further, the photoelectric nonlinear activation unit includes n transimpedance amplifiers, n light sources, n Mach-Zehnder modulators and n second photodetectors; the input ends of the n transimpedance amplifiers are respectively connected to the n first photodetectors, and the output ends of the n transimpedance amplifiers are respectively connected to the radio frequency ends of the n Mach-Zehnder modulators; the transimpedance amplifier is used to amplify the electrical signal output by the first photodetector and add the corresponding bias value in the photoelectric neural network to adjust the amplitude to a predetermined range, obtain an electrical signal carrying information, and input it to the radio frequency end of the n Mach-Zehnder modulators; the n light sources are respectively connected to the optical input ends of the n Mach-Zehnder modulators, used to apply fixed optical power to the Mach-Zehnder modulators, and realize strong nonlinearity by driving the Mach-Zehnder modulators on the transfer function period of the Mach-Zehnder modulators to complete the fitting of high-frequency signals; the optical output ends of the n Mach-Zehnder modulators are respectively connected to the n second photodetectors, and the final electrical signal processed by the Mach-Zehnder modulators is output through the second photodetectors.

[0013] Furthermore, the light source unit is a laser.

[0014] Furthermore, the method further includes inputting a predetermined bias voltage into a bias end of the Mach-Zehnder modulator so that the Mach-Zehnder modulator operates at a NULL point.

[0015] Furthermore, the transimpedance amplifier is used to adjust the amplitude to within the working range of the radio frequency end of the Mach-Zehnder modulator.

[0016] Furthermore, the method also includes inputting a bias voltage to the bias end of each of the first electro-optical modulators and each of the second electro-optical modulators so that the first electro-optical modulators and the second electro-optical modulators operate in a linear region.

[0017] The present invention has the following beneficial effects: The present invention solves the problem of optoelectronic realization of nonlinear activation functions in existing complex artificial intelligence tasks. Based on the characteristics of the periodic activation function of optoelectronic devices, an optoelectronic neural network computing architecture based on periodic activation is designed, which improves the fitting ability of high-frequency information, effectively reduces the number of optoelectronic neural network layers required to perform complex tasks such as image reconstruction, and thus reduces the reasoning time and power consumption. Starting from the intrinsic characteristics of optoelectronic devices, the present invention avoids the problem of precise matching of optical devices with traditional electronic activation function curves in traditional solutions, significantly improves power efficiency and reduces manufacturing costs. Unlike traditional optical computing network designs, the present invention utilizes the intrinsic characteristics of optical devices to process complex signals (especially high-frequency signals) more effectively, so that the optoelectronic neural network has stronger fitting and information expression capabilities, realizes a revolutionary shift in the design concept of optoelectronic neural networks, marks a new direction for the design paradigm, and provides an innovative computing solution for optical computing systems to improve computing energy efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a schematic diagram of an optoelectronic neural network computing architecture based on periodic activation in a specific implementation of the present invention.

[0019] Figure 2 It is a schematic diagram of the optoelectronic neural network computing architecture based on periodic activation in a preferred embodiment of the present invention.

[0020] Figure 3 yes Figure 2 Schematic diagram of the structure of an MDS.

[0021] Figure 4 It is a comparison diagram of the fitting of a randomly generated grayscale image in a specific implementation manner of the present invention. DETAILED DESCRIPTION

[0022] The following is a detailed description of the embodiments of the present invention. It should be emphasized that the following description is only exemplary and is not intended to limit the scope and application of the present invention. In the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other.

[0023] The terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the embodiments of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0024] The specific embodiment of the present invention provides an optoelectronic neural network computing architecture based on periodic activation, such as Figure 1As shown, it includes a light source unit 1, an optoelectronic weighted calculation unit 2 and an optoelectronic nonlinear activation unit 3; the light source unit 1 is used to output n optical signals after splitting the light source; the optoelectronic weighted calculation unit 2 is used to receive the n optical signals output by the light source unit 1 and modulate and weight them to convert them into electrical signals for output, thereby completing the matrix weighted calculation of the optoelectronic neural network; the optoelectronic nonlinear activation unit 3 is used to receive the electrical signals output by the optoelectronic weighted calculation unit and output the final electrical signals after nonlinear activation of the electrical signals, thereby completing the activation of the optoelectronic neural network; wherein n is a positive integer, and n≥2.

[0025] In the above technical solution, the optoelectronic weighted calculation unit 2 completes the matrix weighted calculation of the optoelectronic neural network, and the optoelectronic nonlinear activation unit 3 completes the activation of the optoelectronic neural network. The present invention is more effective in processing complex signals (especially high-frequency signals), and the present invention can make the optoelectronic neural network have stronger fitting and information expression capabilities.

[0026] The present invention uses the photoelectric weighted calculation unit 2 to complete the matrix weighted calculation of the photoelectric neural network. In some embodiments, Figure 2 As shown, the photoelectric weighted calculation unit 2 includes n photoelectric modulator systems MDS and n first photodetectors PD1; the n photoelectric modulator systems MDS are respectively connected to the n first photodetectors PD1, and the photoelectric modulator system MDS is used to modulate and weight the n optical signals to output weighted signals; the first photodetector PD1 is used to sum the intensities of the weighted signals output by the photoelectric modulator system MDS, and convert the optical signals into electrical signals for output.

[0027] In some embodiments, Figure 3As shown, each optoelectronic modulator system MDS includes multiple computing flows, and the light source unit 1 loads the optical signal into each computing flow. Specifically, each computing flow is composed of two first optoelectronic modulators MD1 and second optoelectronic modulators MD2 connected in series. For the first optoelectronic modulator MD1 and the second optoelectronic modulator MD2 on each computing flow, the optical signal output by the light source unit 1 is input into the optical input end of the first optoelectronic modulator MD1, and the electrical signal waveform generated by the electrical signal source is input from the radio frequency end of the first optoelectronic modulator MD1, and the input signal of the optoelectronic neural network is loaded at the radio frequency end of the first optoelectronic modulator MD1, and the optical signal processed by the first optoelectronic modulator MD1 is input into the optical input end of the second optoelectronic modulator MD2; the weight value of the optoelectronic neural network is loaded at the radio frequency end of the second optoelectronic modulator MD2 to output the weighted optical signal at the optical output end of the second optoelectronic modulator MD2; in the same optoelectronic modulator system MDS, the optical output ends of the second optoelectronic modulators MD2 on all computing flows are connected to the optical input end of a first photodetector PD1, and the first photodetector PD1 is used to sum the intensity of the weighted optical signal and convert the optical signal into an electrical signal for output.

[0028] In some embodiments, it also includes inputting a bias voltage at the bias end of each first optoelectronic modulator MD1 and each second optoelectronic modulator MD2 so that the first optoelectronic modulator MD1 and the second optoelectronic modulator MD2 operate in a linear region, so that the input of the optoelectronic neural network and the weight value of the optoelectronic neural network are respectively loaded into the optical signal through the RF end of the first optoelectronic modulator MD1 and the second optoelectronic modulator MD2 in the calculation flow.

[0029] In some embodiments, each computing flow will modulate the optical signal to load different weight values, and the optical signals obtained by all computing flows in each optoelectronic modulator system MDS are finally connected to a first photodetector PD1, and the intensities of the weighted optical signals are summed and converted into electrical signals.

[0030] In some embodiments, Figure 2 As shown, the photoelectric nonlinear activation unit 3 includes n transimpedance amplifiers TIA, n light sources, n Mach-Zehnder modulators MZI and n second photodetectors PD2; the input ends of the n transimpedance amplifiers TIA are respectively connected to the n first photodetectors PD1, and the output ends of the n transimpedance amplifiers TIA are respectively connected to the radio frequency ends of the n Mach-Zehnder modulators MZI; the transimpedance amplifier TIA is used to amplify the electrical signal output by the first photodetector PD1 and add the corresponding n bias values ​​(V i bias, i=1, 2, 3..., n) to adjust the amplitude to a predetermined range (preferably, to the working range of the radio frequency end of the Mach-Zehnder modulator MZI), obtain an electrical signal carrying information, and input it to the radio frequency end of the n Mach-Zehnder modulators MZI; the n light sources are respectively connected to the optical input ends of the n Mach-Zehnder modulators MZI, for applying a uniform and fixed optical power to the Mach-Zehnder modulators MZI (the optical power applied by the light source to each Mach-Zehnder modulator MZI is fixed and the same), and the optical power is adjusted to the predetermined range (preferably, to the working range of the radio frequency end of the Mach-Zehnder modulator MZI) by adjusting the amplitude to the ...) by adjusting the amplitude to the predetermined range The response curve of the ratio of the optical power at the output end to the optical power at the optical input end with voltage change is periodic. Since the optical power at the optical input end of the Mach-Zehnder modulator MZI is constant, it can also be expressed as the response curve of the optical power at the optical output end of the Mach-Zehnder modulator MZI with voltage change is periodic) drives the Mach-Zehnder modulator MZI to achieve strong nonlinearity to complete the fitting of the high-frequency signal; the optical output ends of the n Mach-Zehnder modulators MZI are respectively connected to the n second photodetectors PD2, and the final electrical signal processed by the Mach-Zehnder modulator MZI is output through the second photodetector PD2.

[0031] In the above technical solution, the periodicity of the transfer function of the Mach-Zehnder modulator MZI is used to complete the fitting of the high-frequency signal, which has the advantage of strong fitting ability.

[0032] In some embodiments, the light source unit 1 connected to the optoelectronic weighted calculation unit 2 is a laser. The light source emitted by the laser outputs n optical signals after beam splitting, which enter the optical input ends of n optoelectronic modulator systems MDS for signal modulation, weighting and other processing.

[0033] In some implementations, the method further includes inputting a predetermined bias voltage to a bias terminal of the Mach-Zehnder modulator (MZI) so that the Mach-Zehnder modulator (MZI) operates at a NULL point.

[0034] like Figure 4 As shown, use Figure 2 and 3 The optical neural network computing architecture shown in the figure compares the results of the present invention on a randomly generated 16*16 grayscale image (i.e. Figure 4The figure corresponding to the "original image" in the figure compares the fitting effect of the optoelectronic neural network computing architecture based on periodic activation in the present invention and the existing activation function ReLU. The optoelectronic neural networks used are all 6-layer fully connected neural networks, the number of neurons in each layer is 64, and the half-wave voltage of the nonlinear activation Mach-Zehnder modulator MZI is set to 1V. It can be seen that the effect of the optoelectronic neural network computing architecture based on periodic activation in the present invention on image fitting (such as Figure 4 The corresponding figure of "the present invention" in FIG) is obviously better than the effect of using the activation function ReLU to fit the image (such as Figure 4 The corresponding figure of “ReLU” in .

[0035] The specific implementation of the present invention has the following advantages:

[0036] 1. In the prior art, optical activation functions use optoelectronic devices to simulate traditional nonlinear activation functions. This simulation method faces huge challenges in terms of power consumption and cost in the process of aligning optoelectronic systems with these activation function curves. The optoelectronic neural network computing architecture based on periodic activation adopted by the present invention is based on the physical properties of the Mach-Zehnder modulator itself, does not rely on the matching alignment of traditional activation functions, greatly reduces the energy consumption of nonlinear activation systems, and has stronger reconstruction capabilities, bringing revolutionary changes to the design ideas of optoelectronic neural networks.

[0037] 2. In the prior art, a small portion of the linear modulation interval in the optoelectronic modulator is usually used as the optical activation function. Compared with this traditional method, the present invention realizes the nonlinear function by driving the optoelectronic modulator over multiple cycles of the transfer function. This extremely strong nonlinear capability, due to its periodicity and the non-monotonic behavior caused by it, enables the network to capture more high-frequency information details and has a stronger fitting ability.

[0038] The above content is a further detailed description of the present invention in combination with specific / preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, without departing from the concept of the present invention, it can also make several substitutions or modifications to these described embodiments, and these substitutions or modifications should be regarded as belonging to the protection scope of the present invention. In the description of this specification, the description of the reference terms "an embodiment", "some embodiments", "preferred embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in combination with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily target the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In the absence of mutual contradiction, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples. Although the embodiments of the present invention and their advantages have been described in detail, it should be understood that various changes, substitutions and modifications can be made herein without departing from the scope of protection of the patent application.

Claims

1. An optoelectronic neural network computing architecture based on periodic activation, characterized in that: It includes a light source unit, a photoelectric weighted calculation unit and a photoelectric nonlinear activation unit; The light source unit is used to split the light source and output n optical signals; The photoelectric weighted calculation unit is used to receive the n optical signals output by the light source unit and modulate and weight them to convert them into electrical signals for output, thereby completing the matrix weighted calculation of the photoelectric neural network; The photoelectric nonlinear activation unit is used to receive the electrical signal output by the photoelectric weighted calculation unit and output the final electrical signal after performing nonlinear activation on the electrical signal, thereby completing the activation of the photoelectric neural network; Wherein, n is a positive integer and n≥2.

2. The periodically activated optoelectronic neural network computing architecture according to claim 1, characterized in that: The photoelectric weighted calculation unit comprises n photoelectric modulator systems (MDS) and n first photoelectric detectors (PD1); the n photoelectric modulator systems (MDS) are respectively connected to the n first photoelectric detectors (PD1), and the photoelectric modulator system (MDS) is used to modulate and weight the n optical signals to output weighted signals; the first photodetector (PD1) is used to sum the intensities of the weighted signals output by the photoelectric modulator system (MDS) to convert the optical signals into electrical signals for output.

3. The periodically activated optoelectronic neural network computing architecture according to claim 2, characterized in that: Each optoelectronic modulator system (MDS) includes multiple computing streams, each of which is composed of two first optoelectronic modulators (MD1) and second optoelectronic modulators (MD2) connected in series. For the first optoelectronic modulator (MD1) and the second optoelectronic modulator (MD2) on each computing stream, the optical signal output by the light source unit is input to the optical input end of the first optoelectronic modulator (MD1), the electrical signal waveform generated by the electrical signal source is input from the radio frequency end of the first optoelectronic modulator (MD1), and the input signal of the optoelectronic neural network is loaded at the radio frequency end of the first optoelectronic modulator (MD1), and the optical signal processed by the first optoelectronic modulator (MD1) is input to the optical input end of the second optoelectronic modulator (MD2); Loading a weight value of the optoelectronic neural network at the radio frequency end of the second optoelectronic modulator (MD2) to output a weighted optical signal at the optical output end of the second optoelectronic modulator (MD2); In the same optoelectronic modulator system (MDS), the optical output end of the second optoelectronic modulator (MD2) on all computing flows is connected to the optical input end of a first photodetector (PD1), and the first photodetector (PD1) is used to sum the intensity of the weighted optical signal and convert the optical signal into an electrical signal for output.

4. The periodically activated optoelectronic neural network computing architecture according to claim 3, characterized in that: Different weight values ​​are applied to the optical signal in each computing flow.

5. The periodically activated optoelectronic neural network computing architecture according to claim 3, characterized in that: The photoelectric nonlinear activation unit includes n transimpedance amplifiers (TIA), n light sources, n Mach-Zehnder modulators (MZI) and n second photodetectors (PD2); The input ends of the n transimpedance amplifiers (TIA) are respectively connected to the n first photodetectors (PD1), and the output ends of the n transimpedance amplifiers (TIA) are respectively connected to the radio frequency ends of the n Mach-Zehnder modulators (MZI); the transimpedance amplifier (TIA) is used to amplify the electrical signal output by the first photodetector (PD1) and add the corresponding bias value in the optoelectronic neural network to adjust the amplitude to a predetermined range, thereby obtaining an electrical signal carrying information, and inputting the electrical signal into the radio frequency end of the n Mach-Zehnder modulators (MZI); The n light sources are respectively connected to the optical input ends of the n Mach-Zehnder modulators (MZIs) to apply fixed optical power to the Mach-Zehnder modulators (MZIs), and drive the Mach-Zehnder modulators (MZIs) on the transfer function period of the Mach-Zehnder modulators (MZIs) to achieve strong nonlinearity, so as to complete the fitting of high-frequency signals; The optical output ends of the n Mach-Zehnder modulators (MZI) are respectively connected to the n second photodetectors (PD2), and the final electrical signal processed by the Mach-Zehnder modulator (MZI) is output through the second photodetectors (PD2).

6. The periodically activated optoelectronic neural network computing architecture according to claim 1, characterized in that: The light source unit is a laser.

7. The periodically activated optoelectronic neural network computing architecture according to claim 3, characterized in that: The method also includes inputting a predetermined bias voltage into the bias terminal of the Mach-Zehnder modulator (MZI) so that the Mach-Zehnder modulator (MZI) operates at a NULL point.

8. The periodically activated optoelectronic neural network computing architecture according to claim 5, characterized in that: The transimpedance amplifier (TIA) is used to adjust the amplitude to be within the working range of the radio frequency end of the Mach-Zehnder modulator (MZI).

9. The periodically activated optoelectronic neural network computing architecture according to claim 3, characterized in that: It also includes inputting a bias voltage into the bias end of each of the first photoelectric modulators (MD1) and each of the second photoelectric modulators (MD2) so that the first photoelectric modulator (MD1) and the second photoelectric modulator (MD2) operate in a linear region.

Images