A periodically activated optoelectronic neural network computing architecture

Abstract

The application discloses a kind of photonic neural network computing architecture based on periodic activation, including light source unit, photoelectric weighting computing unit and photoelectric nonlinear activation unit;The light source unit is used to split the light source and output n-way optical signals;The photoelectric weighting computing unit is used to receive the n-way optical signals output by the light source unit and is modulated, weighted and converted into electrical signals output to complete the matrix weighting calculation of photonic neural network;The photoelectric nonlinear activation unit is used to receive the electrical signals output by the photoelectric weighting computing unit and is nonlinearly activated to output the final electrical signals to complete the activation of photonic neural network;Wherein, n is positive integer, and n≥2.The present application can make photonic neural network have stronger fitting and information expression capacity.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and optical computing technology, and in particular to a photoelectric neural network computing architecture based on periodic activation. Background Technology

[0002] With the rapid development of artificial intelligence technology, the computational demands of various large-scale deep learning models are constantly increasing. Traditional electronic hardware such as FPGAs and GPUs are specifically used to accelerate neural network calculations, but due to the limitations of Moore's Law, the development of electrical hardware still struggles to meet the ever-growing demand for computing power. Optoelectronic neural networks, with their advantages of low power consumption, high parallelism, and bandwidth, have become strong contenders in the field of future high-performance artificial intelligence computing. Although significant progress has been made in optical computing research, current research mainly focuses on convolution matrix operations, with relatively little attention paid to nonlinear activation functions.

[0003] Current research on optical activation functions generally employs optoelectronic devices to simulate traditional nonlinear activation functions. However, for complex artificial intelligence tasks, this simulation method faces significant challenges in power efficiency and manufacturing cost when achieving precise alignment between optical devices and these nonlinear activation function curves. Therefore, exploring and utilizing the inherent characteristics of optoelectronic devices to construct a highly efficient and low-power optoelectronic nonlinear activation architecture is crucial.

[0004] In the field of information processing, photoelectric activation functions play a crucial role. Their repeatable modular design and scalability significantly contribute to the ever-increasing computing demands. These characteristics make photoelectric activation functions particularly important when handling advanced computing tasks, especially when dealing with complex application scenarios.

[0005] It should be noted that the information disclosed in the background section above is only for understanding the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, this invention provides a photoelectric neural network computing architecture based on periodic activation.

[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 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, modulate and weight them, and 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 calculation unit, perform nonlinear activation on the electrical signals, and output the final electrical signal, 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 systems are used to modulate and weight the n optical signals to output a weighted signal; the first photodetectors are used to sum the intensity of the weighted signal output by the photoelectric modulator systems and convert the optical signal into an electrical signal for output.

[0010] Furthermore, each optoelectronic modulator system includes multiple computational streams, each consisting of two cascaded first optoelectronic modulators and a second optoelectronic modulator. For each first optoelectronic modulator and second optoelectronic modulator in each computational stream, the optical signal output from the light source unit is input to the optical input terminal of the first optoelectronic modulator, and the electrical signal waveform generated by the electrical signal source is input from the radio frequency terminal of the first optoelectronic modulator. The input signal of the optoelectronic neural network is loaded at the radio frequency terminal of the first optoelectronic modulator, and the optical signal processed by the first optoelectronic modulator is input to the optical input terminal of the second optoelectronic modulator. The weight value of the optoelectronic neural network is loaded at the radio frequency terminal of the second optoelectronic modulator to output a weighted optical signal at the optical output terminal of the second optoelectronic modulator. In the same optoelectronic modulator system, the optical output terminals of the second optoelectronic modulators in all computational streams are connected to the optical input terminal of a first photodetector. 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 computation stream.

[0012] Furthermore, the photoelectric nonlinear activation unit includes n transimpedance amplifiers, n light sources, n Mach-Zehnder modulators, and n second photodetectors. The input terminals of the n transimpedance amplifiers are respectively connected to the n first photodetectors, and the output terminals of the n transimpedance amplifiers are respectively connected to the radio frequency (RF) terminals of the n Mach-Zehnder modulators. The transimpedance amplifiers amplify the electrical signals output from the first photodetectors and add the corresponding bias values ​​from the photoelectric neural network to adjust the amplitude to a predetermined range, thereby obtaining an information-carrying electrical signal, which is then input to the RF terminals of the n Mach-Zehnder modulators. The n light sources are respectively connected to the optical input terminals of the n Mach-Zehnder modulators, and are used to apply a fixed optical power to the Mach-Zehnder modulators. By driving the Mach-Zehnder modulators on the transfer function period, strong nonlinearity is achieved to complete the fitting of high-frequency signals. The optical output terminals 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, it also includes inputting a predetermined bias voltage at the bias terminal of the Mach-Zehnder modulator, so that the Mach-Zehnder modulator operates at the NULL point.

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

[0016] Furthermore, it also includes inputting a bias voltage at the bias terminal of each of the first and second photoelectric modulators to make the first and second photoelectric modulators operate in the linear region.

[0017] This invention offers the following advantages: It solves the problem of implementing nonlinear activation functions in complex artificial intelligence tasks via optoelectronics. By leveraging the characteristics of periodic activation functions in optoelectronic devices, it designs a periodic activation-based optoelectronic neural network architecture, improving the fitting ability for high-frequency information and effectively reducing the number of optoelectronic neural network layers required for complex tasks such as image reconstruction, thereby reducing inference time and power consumption. Starting from the inherent characteristics of optoelectronic devices, this invention avoids the problem of precisely matching optical devices with traditional electronic activation function curves, significantly improving power efficiency and reducing manufacturing costs. Unlike traditional optical computing network designs, this invention utilizes the inherent characteristics of optical devices to process complex signals (especially high-frequency signals) more effectively, giving the optoelectronic neural network stronger fitting and information expression capabilities. This represents a revolutionary shift in the design philosophy of optoelectronic neural networks, marking a new direction in design paradigms and providing an innovative computing solution for improving the energy efficiency of optical computing systems. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of a photoelectric neural network computing architecture based on periodic activation in a specific embodiment of the present invention.

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

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

[0021] Figure 4 This is a comparison chart showing the fitting results of a randomly generated grayscale image in a specific embodiment of the present invention. Detailed Implementation

[0022] The embodiments of the present invention will be described in detail below. It should be emphasized that the following description is merely exemplary and not intended to limit the scope and application of the present invention. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other.

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

[0024] This invention provides a computational architecture for photoelectric neural networks based on periodic activation, such as... Figure 1As shown, the system includes a light source unit 1, a photoelectric weighted calculation unit 2, and a photoelectric nonlinear activation unit 3. The light source unit 1 is used to split the light source and output n optical signals. The photoelectric weighted calculation unit 2 is used to receive the n optical signals output by the light source unit 1, modulate and weight them, and convert them into electrical signals for output, thereby completing the matrix weighted calculation of the photoelectric neural network. The photoelectric nonlinear activation unit 3 is used to receive the electrical signals output by the photoelectric weighted calculation unit, perform nonlinear activation on the electrical signals, and output the final electrical signal, thereby completing the activation of the photoelectric neural network. Wherein, n is a positive integer, and n≥2.

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

[0026] This invention performs matrix weighting calculations of the photoelectric neural network through the photoelectric weighting calculation unit 2. In some embodiments, such as... 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 systems (MDS) are used to modulate and weight the n optical signals to output a weighted signal; the first photodetectors (PD1) are used to sum the intensity of the weighted signal output by the photoelectric modulator systems (MDS) and convert the optical signal into an electrical signal for output.

[0027] In some implementations, such as Figure 3As shown, each optoelectronic modulator system (MDS) includes multiple computational streams. The light source unit 1 loads the optical signal into each computational stream. Specifically, each computational stream consists of two cascaded first optoelectronic modulators MD1 and second optoelectronic modulators MD2. For each first optoelectronic modulator MD1 and second optoelectronic modulator MD2 in each computational stream, the optical signal output by the light source unit 1 is input to the optical input terminal of the first optoelectronic modulator MD1. The electrical signal waveform generated by the electrical signal source is input from the radio frequency (RF) terminal of the first optoelectronic modulator MD1, and the input signal of the optoelectronic neural network is loaded at the RF terminal of the first optoelectronic modulator MD1. The optical signal processed by the first optoelectronic modulator MD1 is input to the optical input terminal of the second optoelectronic modulator MD2. The weight values ​​of the optoelectronic neural network are loaded at the RF terminal of the second optoelectronic modulator MD2 to output the weighted optical signal at the optical output terminal of the second optoelectronic modulator MD2. In the same optoelectronic modulator system (MDS), the optical output terminals of the second optoelectronic modulators MD2 in all computational streams are connected to the optical input terminal of a first photodetector PD1. 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, a bias voltage is input at the bias terminal of each first opto-modulator MD1 and each second opto-modulator MD2 to make the first opto-modulator MD1 and the second opto-modulator MD2 operate in the linear region, such that the input of the opto-neural network and the weight values ​​of the opto-neural network are loaded into the optical signal through the radio frequency terminals of the first opto-modulator MD1 and the second opto-modulator MD2 in the computation stream, respectively.

[0029] In some implementations, each computation stream modulates the optical signal to load different weight values, and the optical signals obtained from all computation streams in each opto-modulator system (MDS) are finally connected to a first photodetector (PD1) to sum the intensity of the weighted optical signals and convert them into electrical signals.

[0030] In some implementations, such as Figure 2 As shown, the photoelectric nonlinear activation unit 3 includes n transimpedance amplifiers (TIAs), n light sources, n Mach-Zehnder modulators (MZIs), and n second photodetectors (PD2s). The input terminals of the n transimpedance amplifiers (TIAs) are respectively connected to the n first photodetectors (PD1s), and the output terminals of the n transimpedance amplifiers (TIAs) are respectively connected to the radio frequency terminals of the n Mach-Zehnder modulators (MZIs). The transimpedance amplifiers (TIAs) are used to amplify the electrical signals output by the first photodetectors (PD1s) and add the corresponding n bias values ​​(V0, V0) in the photoelectric neural network. i bias(i=1, 2, 3, ..., n) to adjust the amplitude to a predetermined range (preferably, to the operating range of the RF terminal of the Mach-Zehnder modulator MZI), obtain an electrical signal carrying information, and input it to the RF terminals of the n Mach-Zehnder modulators MZI; the n light sources are respectively connected to the optical input terminals of the n Mach-Zehnder modulators MZI, and are used to apply 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), through the transfer function period of the Mach-Zehnder modulator MZI (i.e., the optical...) of the Mach-Zehnder modulator MZI... The ratio of the optical power at the output end to the optical power at the optical input end has a periodic response curve as a function of voltage. 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) as a function of voltage being periodic. This drives the Mach-Zehnder modulator (MZI) to achieve strong nonlinearity in order to fit 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 photodetectors (PD2).

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

[0032] In some embodiments, the light source unit 1 connected to the photoelectric weighted calculation unit 2 is a laser. The light source emitted by the laser is split into n optical signals, which enter the optical input terminals of n photoelectric modulator systems (MDS) for signal modulation, weighting and other processing.

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

[0034] like Figure 4 As shown, using Figure 2 and 3 The photoelectric neural network computing architecture shown in this invention compares the results under the same conditions for a randomly generated 16*16 grayscale image (i.e., Figure 4The fitting results of the image (corresponding to the "original image" in the text) are compared with those of the periodically activated photoelectric neural network computing architecture of this invention and the existing ReLU activation function. The photoelectric neural networks used are all 6-layer fully connected neural networks, with 64 neurons in each layer, and the half-wave voltage of the nonlinear activation Mach-Zehnder modulator (MZI) is set to 1V. It can be seen that the periodically activated photoelectric neural network computing architecture of this invention achieves better image fitting results (e.g., ...). Figure 4 The figure corresponding to "this invention" in the text is significantly better than the effect of using the ReLU activation function to fit the image (e.g., Figure 4 The diagram corresponding to "ReLU" in the figure.

[0035] The specific embodiments of the present invention have the following advantages:

[0036] 1. In existing technologies, optical activation functions are simulated using optoelectronic devices to mimic traditional nonlinear activation functions. This simulation method faces significant challenges in terms of power consumption and cost when aligning the optoelectronic system with these activation function curves. The optoelectronic neural network computing architecture based on periodic activation, as proposed in this invention, leverages the inherent physical characteristics of Mach-Zehnder modulators. It does not rely on matching and aligning with traditional activation functions, significantly reducing the energy consumption of nonlinear activation systems and possessing stronger reconstructive capabilities, thus revolutionizing the design approach of optoelectronic neural networks.

[0037] 2. In existing technologies, a small portion of the linear modulation interval in the photoelectric modulator is typically used as the optical activation function. Compared to this traditional method, this invention achieves nonlinear functionality by driving the photoelectric modulator across multiple cycles of the transfer function. This extremely strong nonlinear capability, due to its periodicity and the resulting non-monotonic behavior, enables the network to capture more high-frequency information details and has a stronger fitting ability.

[0038] The above description provides a further detailed explanation of the present invention in conjunction with specific / preferred embodiments, and it should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various substitutions or modifications can be made to these described embodiments without departing from the concept of the present invention, and all such substitutions or modifications should be considered within the scope of protection of the present invention. In the description of this specification, the reference to terms such as "an embodiment," "some embodiments," "preferred embodiment," "example," "specific example," or "some examples," etc., indicates that the specific features, structures, materials, or characteristics described in connection with that embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any suitable manner in one or more embodiments or examples. Without contradiction, those skilled in the art can combine and integrate 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. A photoelectric neural network computing system 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 into n optical signals and output them. The photoelectric weighted calculation unit is used to receive the n optical signals output by the light source unit, modulate and weight them, and then convert them into electrical signals for output, thereby completing the matrix weighted calculation of the photoelectric neural network; the photoelectric weighted calculation unit includes n photoelectric modulator systems (MDS) and n first photodetectors (PD1), and the n photoelectric modulator systems (MDS) are respectively connected to the n first photodetectors (PD1); The photoelectric nonlinear activation unit receives the electrical signal output by the photoelectric weighted calculation unit, performs nonlinear activation on the electrical signal, and outputs the final electrical signal to complete the activation of the photoelectric neural network. The photoelectric nonlinear activation unit includes n transimpedance amplifiers (TIAs), n light sources, n Mach-Zehnder modulators (MZIs), and n second photodetectors (PD2). The input terminals of the n transimpedance amplifiers (TIAs) are respectively connected to the n first photodetectors (PD1), and the output terminals of the n transimpedance amplifiers (TIAs) are respectively connected to the radio frequency terminals of the n Mach-Zehnder modulators (MZIs). The transimpedance amplifiers (TIAs) amplify the electrical signal output by the first photodetectors (PD1) and add the corresponding bias signal from the photoelectric neural network. The amplitude is adjusted to a predetermined range to obtain an electrical signal carrying information, which is then input to the radio frequency (RF) terminals of the n Mach-Zehnder modulators (MZIs). The n light sources are respectively connected to the optical input terminals of the n Mach-Zehnder modulators (MZIs) to apply a fixed optical power to the MZIs. Strong nonlinearity is achieved by driving the MZIs along their transfer function period to fit the high-frequency signal. The optical output terminals of the n Mach-Zehnder modulators (MZIs) are respectively connected to the n second photodetectors (PD2), through which the final electrical signal processed by the MZIs is output. Where n is a positive integer, and n≥2.

2. The photoelectric neural network computing system based on periodic activation as described in claim 1, characterized in that, The photoelectric modulator system (MDS) is used to modulate and weight the n optical signals to output a weighted signal; the first photodetector (PD1) is used to sum the intensity of the weighted signal output by the photoelectric modulator system (MDS) and convert the optical signal into an electrical signal for output.

3. The photoelectric neural network computing system based on periodic activation as described in claim 2, characterized in that, Each optoelectronic modulator system (MDS) includes multiple computational streams. Each computational stream consists of two cascaded first optoelectronic modulators (MD1) and second optoelectronic modulators (MD2). For each first optoelectronic modulator (MD1) and second optoelectronic modulator (MD2) in each computational stream, the optical signal output from the light source unit is input to the optical input terminal of the first optoelectronic modulator (MD1). The electrical signal waveform generated by the electrical signal source is input from the radio frequency terminal of the first optoelectronic modulator (MD1), and the input signal of the optoelectronic neural network is loaded at the radio frequency terminal of the first optoelectronic modulator (MD1). The optical signal processed by the first optoelectronic modulator (MD1) is input to the optical input terminal of the second optoelectronic modulator (MD2). Weight values ​​of a photoelectric neural network are loaded at the radio frequency terminal of the second photoelectric modulator (MD2) to output a weighted optical signal at the optical output terminal of the second photoelectric modulator (MD2); In the same photoelectric modulator system (MDS), the optical output of the second photoelectric modulator (MD2) on all computational streams is connected to the optical input of a first photodetector (PD1). The first photodetector (PD1) is used to sum the intensity of the weighted optical signal and convert the optical signal into an electrical signal output.

4. The photoelectric neural network computing system based on periodic activation as described in claim 3, characterized in that, Different weight values ​​are applied to the optical signal in each computation stream.

5. The photoelectric neural network computing system based on periodic activation as described in claim 1, characterized in that, The light source unit is a laser.

6. The photoelectric neural network computing system based on periodic activation as described in claim 3, characterized in that, It also includes inputting a predetermined bias voltage at the bias terminal of the Mach-Zehnder modulator (MZI) so that the Mach-Zehnder modulator (MZI) operates at the NULL point.

7. The photoelectric neural network computing system based on periodic activation as described in claim 1, characterized in that, The transimpedance amplifier (TIA) is used to adjust the amplitude to the operating range of the radio frequency end of the Mach-Zehnder modulator (MZI).

8. The photoelectric neural network computing system based on periodic activation as described in claim 3, characterized in that, It also includes inputting a bias voltage at the bias terminal of each of the first photoelectric modulators (MD1) and each of the second photoelectric modulators (MD2) to make the first photoelectric modulators (MD1) and the second photoelectric modulators (MD2) operate in the linear region.

Images