Nonlinear activation unit, optoelectronic neural network and photonic chip

Abstract

The nonlinear activation unit, the optoelectronic neural network and the photon chip, the nonlinear activation unit comprises: an optoelectronic detection module comprising a first photodiode and a second photodiode connected in series, the first photodiode is used to receive a first optical input signal and carry out photoelectric conversion, and the second photodiode is used to receive a second optical input signal and carry out photoelectric conversion;Signal receiving module, the signal receiving module is coupled with the anode of the first photodiode, and is coupled with the cathode of the second photodiode to receive the current signal between the anode of the first photodiode and the cathode of the second photodiode;Wherein, the anode of the first photodiode is connected with the cathode of the second photodiode;A positive voltage is applied to the cathode of the first photodiode, and the anode of the second photodiode is grounded. The above scheme can reduce the calculation resource consumption of nonlinear operation in neural network algorithm and improve operation speed.

Description

Technical Field

[0001] This invention relates to the field of neural network algorithm technology, and in particular to a nonlinear activation unit, an opto-neural network, and a photonic chip. Background Technology

[0002] In deep learning, the most popular field of artificial intelligence algorithms today, the computation process mainly involves two parts: matrix multiplication and nonlinear activation functions. Nonlinear operations are the source of the powerful expressive power of artificial neural networks. They can accelerate the convergence speed of the network and improve the accuracy of recognition, making them an indispensable component of neural networks.

[0003] Artificial intelligence algorithms are characterized by processing unstructured data (such as video, images, and audio), requiring extensive linear algebraic operations, and involving a large number of parameters. Computing hardware, primarily based on central processing units (CPUs), cannot meet the computational demands of artificial intelligence, necessitating the use of electronic artificial intelligence (AI) chips. Currently, electronic AI chips are mainly based on Complementary Metal Oxide Semiconductor (CMOS) technology. However, as integrated circuit device dimensions approach physical limits, Moore's Law is slowing down. Simultaneously, microelectronic processors suffer from declining energy efficiency, limited clock frequencies (difficult to exceed 6GHz), electronic crosstalk, high power consumption, and heat generation, severely restricting further performance improvements in existing electronic AI chips.

[0004] To overcome the problems faced by electronic AI chips, photonic AI chips have emerged. Using light to perform deep learning algorithms essentially involves using light to implement matrix multiplication (linear operations) and nonlinear activation functions. In existing technologies, Mach-Zehnder interferometer (MZI) arrays and microring resonator (MRR) arrays are commonly used to implement matrix multiplication operations in the optical domain. However, implementing nonlinear activation function operations in the optical domain still faces significant challenges, mainly including: (1) the generation of optical nonlinear effects generally requires high optical power; (2) the photonic AI chip architecture requires high flexibility for nonlinear activation functions, but the control of existing optical nonlinear effects is difficult and cannot meet this flexibility requirement; (3) the integration of nonlinear activation units with existing linear operation units on the chip also presents problems such as process compatibility and device consistency. In view of this, in existing related technologies, nonlinear operations are still usually performed in the electrical domain, which requires additional equipment and computing resources (e.g., the use of digital chips or analog-to-digital converters (A / D) devices), and the limited conversion speed of related additional equipment in the electrical domain may also become a performance bottleneck for the entire opto-neural network. Summary of the Invention

[0005] One of the objectives of this invention is to provide a nonlinear activation unit that can reduce the computational resource consumption of nonlinear operations in neural network algorithms and improve the computational speed.

[0006] To achieve the above objectives, embodiments of the present invention provide a nonlinear activation unit, comprising: a photodetector module, the photodetector module including a first photodiode and a second photodiode connected in series, the first photodiode being used to receive a first optical input signal and perform photoelectric conversion, and the second photodiode being used to receive a second optical input signal and perform photoelectric conversion; and a signal receiving module, the signal receiving module being coupled to the positive terminal of the first photodiode and to the negative terminal of the second photodiode, to receive a current signal between the positive terminal of the first photodiode and the negative terminal of the second photodiode; wherein, the positive terminal of the first photodiode is connected to the negative terminal of the second photodiode; a positive voltage is applied to the negative terminal of the first photodiode, and the positive terminal of the second photodiode is grounded.

[0007] Optionally, the signal receiving module is coupled to the connection node between the positive terminal of the first photodiode and the negative terminal of the second photodiode; the signal receiving module includes a signal amplifier, the input terminal of which is coupled to the connection node between the positive terminal of the first photodiode and the negative terminal of the second photodiode to receive the current signal at the connection node, and the signal amplifier converts the received current signal into a voltage signal and amplifies it.

[0008] Optionally, the signal amplifier is a transimpedance amplifier (TIA).

[0009] Optionally, the signal receiving module further includes a capacitor, one end of which is connected to the connection node, and the other end of which is connected to the input terminal of the signal amplifier.

[0010] Optionally, the positive voltage applied to the negative terminal of the first photodiode is selected from 1V to 3V.

[0011] Optionally, the first optical input signal and the second optical input signal are coherent.

[0012] This invention also provides a photoelectric neural network, including the nonlinear activation unit as described above.

[0013] Optionally, the photoelectric neural network further includes: an optical modulation unit, used to modulate an intrinsic reference optical signal according to an electrical signal to be processed to obtain an optical modulation signal; and an optical matrix calculation unit, coupled to the optical modulation unit to receive the optical modulation signal and output a first optical input signal and a second optical input signal based on the intrinsic reference optical signal and the optical modulation signal; wherein, when the light intensity difference between the light intensity of the first optical input signal and the light intensity of the second optical input signal is less than or equal to a preset threshold, the current signal between the positive terminal of the first photodiode and the negative terminal of the second photodiode remains unchanged; when the light intensity difference between the light intensity of the first optical input signal and the light intensity of the second optical input signal is greater than the preset threshold, the current signal between the positive terminal of the first photodiode and the negative terminal of the second photodiode is linearly correlated with the light intensity difference.

[0014] Optionally, the optical matrix calculation unit includes: a first linear operation subunit, used to perform linear operations on the optical modulation signal to obtain a first signal; and a coherent mixing subunit, used to perform coherent mixing processing on the first signal and the intrinsic reference optical signal to obtain the first optical input signal and the second optical input signal.

[0015] Optionally, the optical matrix calculation unit includes: a second linear operation subunit, used to perform linear operations on the optical modulation signal to obtain a second signal; wherein, the intrinsic reference optical signal is input as the first optical input signal to the first photodiode of the nonlinear activation unit, and the second signal is input as the second optical input signal to the second photodiode of the nonlinear activation unit; when the light intensity of the first optical input signal is greater than or equal to the light intensity of the second optical input signal, the current signal between the positive terminal of the first photodiode and the negative terminal of the second photodiode is negatively linearly correlated with the light intensity of the second optical input signal; when the light intensity of the first optical input signal is less than the light intensity of the second optical input signal, the current signal between the positive terminal of the first photodiode and the negative terminal of the second photodiode remains unchanged.

[0016] Optionally, the optical matrix calculation unit includes: a third linear operation subunit, used to perform linear operations on the optical modulation signal to obtain a third signal; wherein the third signal is input as the first optical input signal to the first photodiode of the nonlinear activation unit, and the intrinsic reference light signal is input as the second optical input signal to the second photodiode of the nonlinear activation unit; wherein, when the light intensity of the first optical input signal is greater than or equal to the light intensity of the second optical input signal, the current signal between the positive terminal of the first photodiode and the negative terminal of the second photodiode is positively linearly correlated with the light intensity of the first optical input signal; when the light intensity of the first optical input signal is less than the light intensity of the second optical input signal, the current signal between the positive terminal of the first photodiode and the negative terminal of the second photodiode remains unchanged.

[0017] Optionally, the intensity of the intrinsic reference optical signal remains unchanged.

[0018] Optionally, the photoelectric neural network further includes: a light source for generating the intrinsic reference optical signal; and an electrical processing unit for processing the data to be processed into the electrical signal to be processed.

[0019] This invention also provides a photonic chip that integrates one or more photoelectric neural networks as described above, wherein the output electrical signal of the previous photoelectric neural network in the multi-layer photoelectric neural network serves as the input electrical signal of the next photoelectric neural network.

[0020] Compared with the prior art, the technical solution of the embodiments of the present invention has the following beneficial effects:

[0021] This invention provides a nonlinear activation unit, specifically comprising: a photodetector module, the photodetector module including a first photodiode and a second photodiode connected in series, the first photodiode being used to receive a first optical input signal and perform photoelectric conversion, and the second photodiode being used to receive a second optical input signal and perform photoelectric conversion; and a signal receiving module, the signal receiving module being coupled to the positive terminal of the first photodiode and the negative terminal of the second photodiode to receive a current signal between the positive terminal of the first photodiode and the negative terminal of the second photodiode; wherein, the positive terminal of the first photodiode is connected to the negative terminal of the second photodiode; a positive voltage is applied to the negative terminal of the first photodiode, and the positive terminal of the second photodiode is grounded.

[0022] In this embodiment of the invention, a first photodiode and a second photodiode connected in series receive a first optical input signal and a second optical input signal, respectively. A positive voltage is applied to the negative terminal of the first photodiode, and the positive terminal of the second photodiode is grounded. This enables the realization of a photoelectric nonlinear activation function. Specifically, when the intensity difference between the light intensity of the first optical input signal and the light intensity of the second optical input signal is less than or equal to a preset threshold (e.g., 0), the current signal between the positive terminal of the first photodiode and the negative terminal of the second photodiode remains unchanged (e.g., always 0). When the intensity difference is greater than the preset threshold, the current signal between the positive terminal of the first photodiode and the negative terminal of the second photodiode is linearly correlated with the intensity difference (e.g., positively linearly correlated). This relationship is equivalent to realizing the nonlinear activation function ReLU. The nonlinear activation unit implemented using the above scheme can perform photoelectric nonlinear activation function calculations with fast calculation speed and low energy consumption.

[0023] Furthermore, in this embodiment of the invention, a photoelectric neural network is provided, including the aforementioned nonlinear activation unit and an optical matrix calculation unit (for implementing linear operations). Since the nonlinear activation unit provided in this embodiment of the invention has the same process requirements as existing photonic chips for implementing optical matrix calculations, cascaded linear and nonlinear operations can be performed simultaneously on the photonic chip without the need for additional devices (e.g., analog-to-digital converters) specifically for performing nonlinear operations in the electrical domain. This is beneficial for improving computing power and reducing energy consumption while simultaneously increasing the integration level of the photoelectric neural network. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the structure of a nonlinear activation unit in an embodiment of the present invention;

[0025] Figure 2 yes Figure 1 A schematic diagram of the first specific implementation of the nonlinear activation unit;

[0026] Figure 3 yes Figure 2 A schematic diagram of the nonlinear activation function implemented by the nonlinear activation unit in the diagram;

[0027] Figure 4 yes Figure 1 A schematic diagram of a second specific implementation of the nonlinear activation unit;

[0028] Figure 5 yes Figure 4 A schematic diagram of the nonlinear activation function implemented by the nonlinear activation unit in the diagram;

[0029] Figure 6 yes Figure 1 A schematic diagram of a third specific implementation of the nonlinear activation unit;

[0030] Figure 7 yes Figure 6 A schematic diagram of the nonlinear activation function implemented by the nonlinear activation unit in the diagram;

[0031] Figure 8 This is a schematic diagram of the structure of a photoelectric neural network in an embodiment of the present invention. Detailed Implementation

[0032] As mentioned earlier, nonlinear operations are the source of the powerful expressive capabilities of artificial neural networks. They can accelerate the convergence speed of the network and improve the accuracy of recognition, making them an indispensable component of neural networks.

[0033] In existing technologies, photonic AI chips often perform linear operations in the optical domain, and then use additional devices / equipment to perform linear operations in the electrical domain. Specifically, linear operations in the optical domain (such as matrix multiplication) are usually implemented using Mach-Zehnder interferometer (MZI) arrays, micro-ring resonator (MRR) arrays, etc.; nonlinear operations in the electrical domain require the use of digital chips or analog-to-digital converters (A / D) devices, which not only require additional power consumption, but also have limited computing speed in the electrical domain, which may become the performance bottleneck of the entire neural network.

[0034] The inventors of this invention have discovered through research that the main reasons for the significant bottleneck in realizing photoelectric nonlinear activation function operations include: (1) the generation of optical nonlinear effects generally requires high optical power; (2) in the photonic AI chip architecture, the flexibility requirements for nonlinear activation functions are high, and the control of existing optical nonlinear effects is difficult, which cannot meet the need for such flexibility; (3) the integration of nonlinear activation units and existing linear operation units on photonic AI chips also presents problems such as process compatibility and device consistency.

[0035] To address the aforementioned technical problems, this invention provides a nonlinear activation unit, specifically comprising: a photodetector module, which includes a first photodiode and a second photodiode connected in series; the first photodiode receiving a first optical input signal and performing photoelectric conversion, and the second photodiode receiving a second optical input signal and performing photoelectric conversion; and a signal receiving module coupled to the anode of the first photodiode and the cathode of the second photodiode to receive a current signal between the anode of the first photodiode and the cathode of the second photodiode; wherein the anode of the first photodiode is connected to the cathode of the second photodiode; a positive voltage is applied to the cathode of the first photodiode, and the anode of the second photodiode is grounded.

[0036] As described above, this embodiment of the invention utilizes a first photodiode and a second photodiode connected in series to receive a first optical input signal and a second optical input signal, respectively. A positive voltage is applied to the negative terminal of the first photodiode, and the positive terminal of the second photodiode is grounded. This enables the realization of a photoelectric nonlinear activation function. Specifically, when the intensity difference between the first and second optical input signals is less than or equal to a preset threshold (e.g., 0), the current signal between the positive and negative terminals of the first and second photodiodes remains unchanged (e.g., always 0). When the intensity difference is greater than the preset threshold, the current signal between the positive and negative terminals of the first and second photodiodes is linearly correlated with the intensity difference (e.g., positively linearly correlated). This relationship is equivalent to realizing the nonlinear activation function ReLU. The nonlinear activation unit implemented using this scheme can perform photoelectric nonlinear activation function calculations with fast processing speed and low energy consumption.

[0037] To make the above-mentioned objectives, features and beneficial effects of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0038] Reference Figure 1 , Figure 1 This is a schematic diagram of the structure of a nonlinear activation unit in an embodiment of the present invention.

[0039] exist Figure 1 In this context, the nonlinear activation unit 10 may include a photoelectric detection module 101 and a signal receiving module 102.

[0040] The photoelectric detection module 101 may include a first photodiode 1011 and a second photodiode 1012 connected in series. The first photodiode 1011 is used to receive a first optical input signal and perform photoelectric conversion, and the second photodiode 1012 is used to receive a second optical input signal and perform photoelectric conversion.

[0041] Furthermore, the first optical input signal and the second optical input signal are coherent (for example, they can be optical input signals obtained by processing optical signals from the same light source separately). This can help improve the signal strength of both the processed and received signals.

[0042] In this configuration, the anode of the first photodiode 1011 is connected to the cathode of the second photodiode 1012. A positive voltage is applied to the cathode of the first photodiode 1011, while the anode of the second photodiode 1012 is grounded.

[0043] In some non-limiting embodiments, the positive voltage applied to the negative terminal of the first photodiode 1011 can be selected from 1V to 3V, for example, 2V.

[0044] In a specific implementation, a balanced photodetector containing two photodiodes connected in series can be used to receive the first optical input signal and the second optical input signal and perform photoelectric conversion respectively.

[0045] The signal receiving module 102 is coupled to the positive terminal of the first photodiode 1011 and the negative terminal of the second photodiode 1012 to receive the current signal between the positive terminal of the first photodiode 1011 and the negative terminal of the second photodiode 1012.

[0046] In some non-limiting embodiments, the anode of the first photodiode 1011 and the cathode of the second photodiode 1012 can be connected via a connection node Q. The signal receiving module 102 is electrically connected to the connection node Q to receive a current signal at the connection node Q. The current signal received by the signal receiving module 102 can be used to obtain nonlinear calculation results.

[0047] The signal receiving module 102 may include a signal amplifier 1021. The input terminal of the signal amplifier 1021 is coupled to the connection node Q between the positive terminal of the first photodiode 1011 and the negative terminal of the second photodiode 1012 to receive the current signal at the connection node Q. The signal amplifier 1021 converts the received current signal into a voltage signal and amplifies it.

[0048] The signal amplifier 1021 can be a trans-impedance amplifier (TIA). The signal amplifier 1021 can also be other signal processing devices that can achieve the same or similar functions, and the embodiments of the present invention do not limit this.

[0049] Furthermore, the signal receiving module 102 may also include a capacitor 1022, one end of which is connected to the connection node Q, and the other end of which is connected to the input terminal of the signal amplifier 1021.

[0050] In this embodiment of the invention, a first photodiode and a second photodiode connected in series receive a first optical input signal and a second optical input signal, respectively. A positive voltage is applied to the negative terminal of the first photodiode, and the positive terminal of the second photodiode is grounded. This enables the realization of a photoelectric nonlinear activation function. Specifically, when the intensity difference between the light intensity of the first optical input signal and the light intensity of the second optical input signal is less than or equal to a preset threshold (e.g., 0), the current signal between the positive terminal of the first photodiode and the negative terminal of the second photodiode remains unchanged (e.g., always 0). When the intensity difference is greater than the preset threshold, the current signal between the positive terminal of the first photodiode and the negative terminal of the second photodiode is linearly correlated with the intensity difference (e.g., positively linearly correlated). This relationship is equivalent to realizing the nonlinear activation function ReLU. The nonlinear activation unit implemented using the above scheme can perform photoelectric nonlinear activation function calculations with fast calculation speed and low energy consumption.

[0051] Reference Figure 2 , Figure 2 yes Figure 1 A schematic diagram of the first specific implementation of the nonlinear activation unit.

[0052] Figure 2 In this process, the first optical input signal received by the first photodiode 1011 can be one of the two optical signals output by the upper-level signal processing unit, and the second optical input signal received by the second photodiode 1012 can be the other of the two optical signals output by the upper-level signal processing unit.

[0053] The upper-level signal processing unit may be an optical matrix calculation unit 20.

[0054] In this embodiment, the optical matrix calculation unit 20 may include a first linear operation subunit 201 and a coherent mixing subunit 202. The first linear operation subunit 201 can be used to perform linear operations (e.g., matrix multiplication) on the optical modulation signal to obtain a first signal. The coherent mixing subunit 202 can be used to perform coherent mixing processing on the first signal and the intrinsic reference optical signal, outputting two optical signals. One of the two optical signals is used as the first optical input signal and input to the first photodiode 1011, and the other is used as the second optical input signal and input to the second photodiode 1012.

[0055] The first linear operation subunit 201 can use a Mach-Zehnder interferometer (MZI) array, a microring resonator (MRR) array, etc., to implement optical linear operations, but is not limited to these. The coherent mixing subunit 202 can use a multi-mode inferometer (MMI), a directional coupler (DC), a Mach-Zehnder interferometer, a microring resonator, etc., to implement coherent signal mixing processing, but is not limited to these.

[0056] Combination Figure 3 , Figure 3 yes Figure 2 A schematic diagram of the nonlinear activation function implemented by the nonlinear activation unit in the diagram.

[0057] exist Figure 3 In the figure, the horizontal axis represents the light intensity difference P1 between the light intensity of the first optical input signal and the light intensity of the second optical input signal, and the vertical axis represents the voltage signal output by the nonlinear activation unit 10, which can specifically be the voltage signal U1 output by the output terminal of the signal amplifier 1021 in the nonlinear activation unit 10 (i.e., the nonlinear calculation result).

[0058] Specifically, when the light intensity difference P1 is less than or equal to a preset threshold (in this embodiment, the preset threshold is set to 0), the current signal between the positive terminal of the first photodiode 1011 and the negative terminal of the second photodiode 1012 remains unchanged (in this embodiment, the current signal is kept at 0); when the light intensity difference P1 is greater than the preset threshold, the current signal between the positive terminal of the first photodiode 1011 and the negative terminal of the second photodiode 1012 is positively linearly correlated with the light intensity difference P1.

[0059] The light intensity difference P1 refers to the difference between the light intensity of the first optical input signal and the light intensity of the second optical input signal.

[0060] Reference Figure 4 , Figure 4 yes Figure 1 A schematic diagram of a second specific implementation of the nonlinear activation unit.

[0061] Figure 4 In this process, the first optical input signal received by the first photodiode 1011 can be an intrinsic reference optical signal, and the second optical input signal received by the second photodiode 1012 can be an optical signal output by the signal processing unit of the previous layer.

[0062] The optical signal output by the upper-level signal processing unit can be the second signal output by the optical matrix calculation unit 20; the intrinsic reference optical signal can be the optical signal emitted by the light source that is coherent with the second optical input signal.

[0063] It should be noted that, in one specific embodiment, the intrinsic reference optical signal can be emitted by the light source and directly input to the first photodiode 1011.

[0064] In another specific embodiment, the intrinsic reference optical signal can be emitted by a light source, then passed through the optical matrix calculation unit 20, and then input to the first photodiode 1011 (e.g., ...). Figure 4 As shown, in this case, the optical matrix calculation unit 20 can directly output the intrinsic reference optical signal without processing it.

[0065] Furthermore, the intensity of the intrinsic reference optical signal can remain constant, that is, a fixed intensity can be used, thereby improving the stability and intensity of the first optical input signal.

[0066] In this embodiment, the optical matrix calculation unit 20 may include a second linear operation subunit 203, which can be used to perform linear operations on the optical modulation signal to obtain the second signal. For a detailed implementation of the second linear operation subunit 203, please refer to... Figure 2 The description of the first linear operation subunit 201 is omitted here.

[0067] Combination Figure 5 , Figure 5 yes Figure 4 A schematic diagram of the nonlinear activation function implemented by the nonlinear activation unit in the diagram.

[0068] exist Figure 5 In the figure, the horizontal axis represents the light intensity P2 of the second optical input signal, and the vertical axis represents the voltage signal output by the nonlinear activation unit 10. Specifically, it can be the voltage signal U2 output by the output terminal of the signal amplifier 1021 in the nonlinear activation unit 10 (i.e., the nonlinear operation result).

[0069] Specifically, when the light intensity of the first optical input signal is greater than or equal to the light intensity of the second optical input signal, the current signal between the positive terminal of the first photodiode 1011 and the negative terminal of the second photodiode 1012 is negatively linearly correlated with the light intensity P2 of the second optical input signal; when the light intensity of the first optical input signal is less than the light intensity P2 of the second optical input signal, the current signal between the positive terminal of the first photodiode 1011 and the negative terminal of the second photodiode 1012 remains unchanged (in this embodiment, the current signal remains 0).

[0070] Reference Figure 6 , Figure 6 yes Figure 1 A schematic diagram of the third specific implementation of the nonlinear activation unit.

[0071] Figure 6 In this process, the first optical input signal received by the first photodiode 1011 can be the optical signal output by the signal processing unit of the previous layer, and the second optical input signal received by the second photodiode 1012 can be the intrinsic reference optical signal.

[0072] The optical signal output by the upper-level signal processing unit can be the third signal output by the optical matrix calculation unit 20; the intrinsic reference optical signal can be an optical signal emitted by the light source that is coherent with the first optical input signal.

[0073] It should be noted that, in one specific embodiment, the intrinsic reference optical signal can be emitted by the light source and directly input to the second photodiode 1012.

[0074] In another specific embodiment, the intrinsic reference optical signal can be emitted by a light source, then passed through the optical matrix calculation unit 20, and then input to the second photodiode 1012 (e.g., ...). Figure 6 As shown, in this case, the optical matrix calculation unit 20 can directly output the intrinsic reference optical signal without processing it.

[0075] Furthermore, the intensity of the intrinsic reference optical signal can remain constant, that is, a fixed intensity can be used, thereby improving the stability and intensity of the second optical input signal.

[0076] In this embodiment, the optical matrix calculation unit 20 may include a third linear operation subunit 204, which can be used to perform linear operations on the optical modulation signal to obtain the third signal. For a detailed implementation of the third linear operation subunit 204, please refer to... Figure 2 The description of the first linear operation subunit 201 is omitted here.

[0077] Combination Figure 7 , Figure 7 yes Figure 6 A schematic diagram of the nonlinear activation function implemented by the nonlinear activation unit in the diagram.

[0078] exist Figure 7 In the figure, the horizontal axis represents the light intensity P3 of the first optical input signal, and the vertical axis represents the voltage signal output by the nonlinear activation unit 10. Specifically, it can be the voltage signal U3 output by the output terminal of the signal amplifier 1021 in the nonlinear activation unit 10 (i.e., the nonlinear operation result).

[0079] Specifically, when the light intensity P3 of the first optical input signal is greater than or equal to the light intensity of the second optical input signal, the current signal between the positive terminal of the first photodiode 1011 and the negative terminal of the second photodiode 1012 is positively linearly correlated with the light intensity P3 of the first optical input signal; when the light intensity P3 of the first optical input signal is less than the light intensity of the second optical input signal, the current signal between the positive terminal of the first photodiode 1011 and the negative terminal of the second photodiode 1012 remains unchanged (in this embodiment, the current signal remains 0).

[0080] Reference Figure 8 , Figure 8 This is a schematic diagram of the structure of a photoelectric neural network in an embodiment of the present invention.

[0081] The photoelectric neural network 1 may include a nonlinear activation unit 10, and may also include an optical matrix calculation unit 20 and an optical modulation unit 30.

[0082] The optical modulation unit 30 is used to modulate the intrinsic reference optical signal according to the electrical signal to be processed to obtain an optical modulation signal; the optical matrix calculation unit 20 is coupled to the optical modulation unit 30 to receive the optical modulation signal, and outputs the first optical input signal and the second optical input signal based on the intrinsic reference optical signal and the optical modulation signal.

[0083] Furthermore, the optical matrix calculation unit 20 may include a first linear operation subunit 201 and a coherent mixing subunit 202. The first linear operation subunit 201 may be used to perform linear operations (e.g., matrix multiplication) on the optical modulation signal to obtain a first signal, and the coherent mixing subunit 202 may be used to perform coherent mixing processing on the first signal and the intrinsic reference optical signal to output the first optical input signal and the second optical input signal.

[0084] In a specific implementation, the optical matrix calculation unit 20 may also include only the first linear operation subunit 201, wherein the first signal output by the first linear operation subunit 201 is used as the first optical input signal, and the intrinsic reference light signal is used as the second optical input signal.

[0085] Furthermore, the photoelectric neural network 1 may also include a light source 40 and an electrical processing unit 50, wherein the light source 40 is used to generate the intrinsic reference light signal, and the electrical processing unit 50 is used to process the data to be processed into the electrical signal to be processed.

[0086] The photonic neural network 1 described in this embodiment of the invention can be used in application fields such as image processing, image recognition, speech recognition, gene sequencing, quantum communication, or quantum computing.

[0087] This invention also provides a photonic chip, which may integrate one or more of the aforementioned photoelectric neural networks 1, wherein the output electrical signal of the previous photoelectric neural network in the multi-layer photoelectric neural network serves as the input electrical signal of the next photoelectric neural network.

[0088] In this embodiment of the invention, since the nonlinear activation unit 10 and the optical matrix calculation unit 20 can be implemented using the same process, cascaded linear and nonlinear operations can be performed simultaneously on the photonic chip without the need for additional devices (such as analog-to-digital converters) to convert analog signals into digital signals and then perform nonlinear processing in the electrical domain. This overcomes the performance bottleneck of existing optoelectronic neural network architectures in handling nonlinear calculations on digital chips, effectively reducing energy consumption, increasing the computational speed of neural network algorithms, and improving the integration density of photonic chips. Furthermore, in practical applications, multi-layer optoelectronic neural networks can be integrated onto a small number of photonic chips according to application requirements, thereby significantly reducing the computational cost of optoelectronic neural networks.

[0089] It should be understood that "multiple" in the embodiments of this application refers to two or more.

[0090] The descriptions of "first," "second," etc., appearing in the embodiments of this application are for illustrative purposes and to distinguish the objects being described. They have no order and do not indicate any special limitation on the number of devices in the embodiments of this application, nor do they constitute any limitation on the embodiments of this application.

[0091] While the present invention has been disclosed above, it is not limited thereto. Any person skilled in the art can make various modifications and alterations without departing from the spirit and scope of the invention; therefore, the scope of protection of the present invention should be determined by the scope defined in the claims.

Claims

1. A nonlinear activation unit, characterized in that, include: A photoelectric detection module includes a first photodiode and a second photodiode connected in series. The first photodiode is used to receive a first optical input signal and perform photoelectric conversion, and the second photodiode is used to receive a second optical input signal and perform photoelectric conversion. A signal receiving module is coupled to the positive terminal of the first photodiode and to the negative terminal of the second photodiode to receive the current signal between the positive terminal of the first photodiode and the negative terminal of the second photodiode. The positive terminal of the first photodiode is connected to the negative terminal of the second photodiode. A positive voltage is applied to the negative terminal of the first photodiode, and the positive terminal of the second photodiode is grounded.

2. The nonlinear activation unit as described in claim 1, characterized in that, The signal receiving module is coupled to the connection node between the positive terminal of the first photodiode and the negative terminal of the second photodiode. The signal receiving module includes: A signal amplifier is provided, wherein the input terminal of the signal amplifier is coupled to the connection node between the positive terminal of the first photodiode and the negative terminal of the second photodiode to receive the current signal at the connection node, and the signal amplifier converts the received current signal into a voltage signal and amplifies it.

3. The nonlinear activation unit as described in claim 2, characterized in that, The signal amplifier is a transimpedance amplifier (TIA).

4. The nonlinear activation unit as described in claim 2, characterized in that, The signal receiving module further includes: A capacitor, one end of which is connected to the connection node, and the other end of which is connected to the input terminal of the signal amplifier.

5. The nonlinear activation unit as described in any one of claims 1 to 4, characterized in that, The positive voltage applied to the negative terminal of the first photodiode is selected from 1V to 3V.

6. The nonlinear activation unit as described in any one of claims 1 to 4, characterized in that, The first optical input signal and the second optical input signal are coherent.

7. A photoelectric neural network, characterized in that, It includes the nonlinear activation unit as described in any one of claims 1 to 6.

8. The photoelectric neural network as described in claim 7, characterized in that, The photoelectric neural network also includes: An optical modulation unit is used to modulate an intrinsic reference optical signal according to an electrical signal to be processed, so as to obtain an optical modulation signal; An optical matrix calculation unit is coupled to the optical modulation unit to receive the optical modulation signal, and outputs the first optical input signal and the second optical input signal based on the intrinsic reference light signal and the optical modulation signal; Wherein, when the light intensity difference between the light intensity of the first optical input signal and the light intensity of the second optical input signal is less than or equal to a preset threshold, the current signal between the positive terminal of the first photodiode and the negative terminal of the second photodiode remains unchanged. When the light intensity difference between the light intensity of the first optical input signal and the light intensity of the second optical input signal is greater than the preset threshold, the current signal between the positive terminal of the first photodiode and the negative terminal of the second photodiode is linearly correlated with the light intensity difference.

9. The photoelectric neural network as described in claim 8, characterized in that, The optical matrix calculation unit includes: The first linear operation subunit is used to perform linear operations on the optical modulation signal to obtain a first signal; The coherent mixing subunit is used to perform coherent mixing processing on the first signal and the intrinsic reference optical signal to obtain the first optical input signal and the second optical input signal.

10. The photoelectric neural network as described in claim 8, characterized in that, The optical matrix calculation unit includes: The second linear operation subunit is used to perform linear operations on the optical modulation signal to obtain a second signal; Wherein, the intrinsic reference optical signal is input to the first photodiode of the nonlinear activation unit as the first optical input signal, and the second signal is input to the second photodiode of the nonlinear activation unit as the second optical input signal; When the light intensity of the first optical input signal is greater than or equal to the light intensity of the second optical input signal, the current signal between the positive terminal of the first photodiode and the negative terminal of the second photodiode is negatively linearly correlated with the light intensity of the second optical input signal. When the light intensity of the first optical input signal is less than the light intensity of the second optical input signal, the current signal between the positive terminal of the first photodiode and the negative terminal of the second photodiode remains unchanged.

11. The photoelectric neural network as described in claim 8, characterized in that, The optical matrix calculation unit includes: The third linear operation subunit is used to perform linear operations on the optical modulation signal to obtain a third signal; The third signal is input to the first photodiode of the nonlinear activation unit as the first optical input signal, and the intrinsic reference optical signal is input to the second photodiode of the nonlinear activation unit as the second optical input signal. Wherein, when the light intensity of the first optical input signal is greater than or equal to the light intensity of the second optical input signal, the current signal between the positive terminal of the first photodiode and the negative terminal of the second photodiode is positively linearly correlated with the light intensity of the first optical input signal; When the light intensity of the first optical input signal is less than the light intensity of the second optical input signal, the current signal between the positive terminal of the first photodiode and the negative terminal of the second photodiode remains unchanged.

12. The photoelectric neural network according to any one of claims 8 to 10, characterized in that, The intensity of the intrinsic reference optical signal remains unchanged.

13. The photoelectric neural network according to any one of claims 8 to 10, characterized in that, The photoelectric neural network also includes: A light source, used to generate the intrinsic reference optical signal; An electrical processing unit is used to process the data to be processed into the electrical signal to be processed.

14. A photonic chip, characterized in that, The photonic chip integrates one or more photoelectric neural networks as described in any one of claims 8 to 13; In this multilayer photoelectric neural network, the output electrical signal of the previous photoelectric neural network layer serves as the input electrical signal of the next photoelectric neural network layer.

Images