Integrated reconfigurable photoelectric computing chip and electronic equipment

Abstract

The invention provides an integrated reconfigurable photoelectric computing chip and electronic equipment. The photoelectric computing chip comprises a physical architecture supporting layer; the optical neural network layer is in supporting connection with the physical architecture supporting layer, the optical neural network layer at least comprises one or a combination of more of a convolution layer, a pooling layer, a full connection layer and a nonlinear layer, and the optical neural network layer converts input data into optical processing signals based on diffraction optics and Fourier optics principles; the electrical neural network layer is connected with the optical neural network layer, converts the optical processing signal into an electrical processing signal based on photoelectric conversion, and completes task calculation according to the electrical processing signal; wherein the convolution layer, the pooling layer and the full connection layer are of a physical mapping structure, and the physical mapping structure comprises a substrate and a photoelectric device arranged on the substrate. The photoelectric computing chip provided by the invention has the characteristics of high parallel computing, integrability, reconfigurability and the like, and the application capability of the photoelectric computing chip in complex tasks can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of optoelectronic computing chips, and more particularly, to an integrated reconfigurable optoelectronic computing chip and an electronic device. Background Art

[0002] Currently, artificial intelligence technology has been widely applied in multiple fields, including but not limited to autonomous driving, medical diagnosis, financial analysis, educational assistance, and security monitoring. With the development of these applications, the computing platforms from terminal devices to central servers have faced a sharp increase in the amount of tasks. This not only increases the model complexity, but also the data to be processed grows exponentially, bringing a large operating pressure to the computing system.

[0003] As the semiconductor industry enters the post-Moore era, traditional electronic chip manufacturing processes are facing challenges of physical limits, and the replacement of processes has become increasingly difficult, making it difficult to improve the performance of electronic chips.

[0004] Optoelectronic computing chips, as a new computing architecture, can combine the advantages of optical computing and analog electronic computing to achieve high-speed interconnection and efficient conversion of optoelectronic signals, with high computing speed and computing power. However, the inventors of the present application found that current optoelectronic computing chips still have problems such as small parallel computing scale, difficulty in integrating complex networks, and poor reconfigurability.

[0005] The content of the background art section is only the technology known to the applicant and does not necessarily represent the prior art in this field. Summary of the Invention

[0006] According to one aspect of the present invention, there is provided an integrated reconfigurable optoelectronic computing chip, including a physical architecture support layer, an optical neural network layer, and an electrical neural network layer. The optical neural network layer is supported and connected to the physical architecture support layer. The optical neural network layer includes at least one or a combination of a convolutional layer, a pooling layer, a fully connected layer, and a non-linear layer. The optical neural network layer converts input data into optical processing signals based on the principles of diffractive optics and Fourier optics. The electrical neural network layer is connected to the optical neural network layer, converts the optical processing signals into electrical processing signals based on optoelectronic conversion, and completes task calculations according to the electrical processing signals. The convolutional layer, the pooling layer, and the fully connected layer are physical mapping structures, and the physical mapping structures include a substrate and optoelectronic devices disposed on the substrate.

[0007] According to some embodiments of the present invention, the electrical neural network layer further includes a photoelectric conversion module, a signal reading module, and an electrical calculation module. The photoelectric conversion module receives the optical processing signal from the optical neural network layer and converts the optical processing signal into an electrical processing signal; the signal reading module reads the electrical processing signal; and the electrical calculation module performs task calculations based on the electrical neural network according to the electrical processing signal from the signal reading module.

[0008] According to some embodiments of the present invention, the optical neural network layer includes at least one convolutional layer.

[0009] According to some embodiments of the present invention, the optical neural network layer includes at least one convolutional layer and at least one pooling layer.

[0010] According to some embodiments of the present invention, the optical neural network layer includes at least one fully connected layer.

[0011] According to some embodiments of the present invention, the optical neural network layer includes at least one non - linear layer and one or more of a convolutional layer, a pooling layer, and a fully connected layer.

[0012] According to some embodiments of the present invention, the optical neural network layer includes at least one convolutional layer, at least one pooling layer, at least one fully connected layer, and at least one non - linear layer.

[0013] According to some embodiments of the present invention, the substrate is a metasurface.

[0014] According to some embodiments of the present invention, the optoelectronic device includes at least one of a diffractive optical element and a modulator.

[0015] According to another aspect of the present invention, the present invention provides an electronic device. The electronic device includes the optoelectronic computing chip as described above.

[0016] Advantageous Effects

[0017] The optoelectronic computing chip provided by the present invention includes a physical architecture support layer, an optical neural network layer, and an electrical neural network layer. The physical architecture support layer and the optical neural network layer constitute a physically reconfigurable optical network computing module, and the electrical neural network layer constitutes an electrical network computing module. Thus, the optoelectronic computing chip can integrate the functions of photoelectric conversion and electrical neural network computing, and can achieve parallel computing of large - scale input data. And the optical neural network layer can include one or more combinations of a convolutional layer, a pooling layer, a fully connected layer, and a non - linear layer. By configuring the structure of the optical neural network layer accordingly, a reconfigurable optical neural network layer can be formed. The present invention can perform different algorithm constructions through the reconfigurable optical neural network layer, and can meet the intelligent computing tasks under complex application scenarios.

[0018] The optoelectronic computing chip provided by the present invention has characteristics such as high parallel computing, integrability, and reconfigurability, which can improve the application ability of the optoelectronic computing chip in complex tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0020] Figure 1 Schematic structural diagram of the optoelectronic computing chip showing an embodiment of the present invention;

[0021] Figure 2a Schematic structural diagram of the convolutional layer showing an embodiment of the present invention;

[0022] Figure 2b Schematic calculation diagram of a convolutional layer showing an embodiment of the present invention;

[0023] Figure 2c Another schematic calculation diagram of the convolutional layer showing an embodiment of the present invention;

[0024] Figure 3a Schematic structural diagram of the pooling layer showing an embodiment of the present invention;

[0025] Figure 3b Schematic calculation diagram of a pooling layer showing an embodiment of the present invention;

[0026] Figure 3c Another schematic calculation diagram of the pooling layer showing an embodiment of the present invention;

[0027] Figure 4a Schematic structural diagram of the fully connected layer showing an embodiment of the present invention;

[0028] Figure 4b Schematic calculation diagram of a fully connected layer showing an embodiment of the present invention;

[0029] Figure 4c Another schematic calculation diagram of the fully connected layer showing an embodiment of the present invention;

[0030] Figure 5 Schematic calculation diagram of a non - linear layer showing an embodiment of the present invention;

[0031] Figure 6 Another schematic structural diagram of the optoelectronic computing chip showing an embodiment of the present invention;

[0032] Figure 7 Schematic calculation diagram of a LeNet - 5 - like network showing an embodiment of the present invention.

[0033] Description of Reference Numerals

[0034] Physical Architecture Support Layer 10; Optical Neural Network Layer 20; Electrical Neural Network Layer 30;

[0035] Photoelectric Conversion Module 31; Signal Readout Module 32; Electrical Calculation Module 33;

[0036] Convolutional Layer 21; Pooling Layer 22; Fully Connected Layer 23; Nonlinear Layer 24. Detailed Implementation Manner

[0037] Example embodiments will now be described more fully with reference to the accompanying drawings. However, the example embodiments can be implemented in various forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the concept of the example embodiments to those skilled in the art. Identical reference numerals in the figures denote identical or similar parts, and thus their repetitive description will be omitted.

[0038] The features, structures, or characteristics described may be combined in one or more embodiments in any suitable manner. In the following description, numerous specific details are provided to give a thorough understanding of the embodiments of the present disclosure. However, those skilled in the art will realize that the technical solutions of the present disclosure can be practiced without one or more of these specific details, or may be implemented in other ways, components, materials, devices, etc. In these cases, well-known structures, methods, devices, implementations, materials, or operations will not be shown or described in detail.

[0039] In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but optionally further includes steps or units not listed, or optionally further includes other steps or units inherent to these processes, methods, products, or devices.

[0040] The terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish different objects, rather than to describe a specific order.

[0041] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0042] Currently, the architectures of optoelectronic computing chips can be mainly divided into two types: on-chip and spatial light.

[0043] On-chip optoelectronic computing chips can use Mach-Zehnder modulators or microring modulators as reconfigurable optical processing units, and can implement optoelectronic computing of various neural network architectures (such as convolutional neural networks and fully connected neural networks, etc.). However, the inventors found that although on-chip optoelectronic computing chips are superior to traditional electrical chips in terms of computing speed, their computing scale is relatively small, the excitation power consumption is large, and it is difficult to support large-scale parallel computing tasks, resulting in the limitation of the expansion of their application scenarios.

[0044] On the other hand, spatial light computing chips can provide higher-dimensional computing capabilities. For example, current methods such as diffraction neural networks and spatial convolutions can parallelly process larger-scale data such as images. However, the inventors also found that these network architectures are usually limited to separate spatial convolutions or fully connected operations, with a single network architecture and unable to integrate complex computing tasks. In addition, due to the large volume of spatial light computing systems, they do not have the conditions for integration, and the non-reconfigurability of optical components also greatly limits their application scope.

[0045] According to one aspect of the present invention, the present invention provides an integrated reconfigurable optoelectronic computing chip. Figure 1 The structural schematic diagram of the optoelectronic computing chip showing the embodiments of the present invention.

[0046] According to an exemplary embodiment, as Figure 1 shown, the optoelectronic computing chip may include a physical architecture support layer 10, an optical neural network layer 20, and an electrical neural network layer 30.

[0047] The physical architecture support layer 10 is supported and connected to the optical neural network layer 20, and can provide a supporting force for the optical neural network layer 20.

[0048] Exemplarily, as Figure 1 shown, the physical architecture support layer 10 may be arranged in a semi-surrounding manner on the outer periphery of the optical neural network layer 20. Of course, the physical architecture support layer 10 may also be arranged in a full-surrounding manner on the outer periphery of the optical neural network layer 20. Or, the physical architecture support layer 10 may also be arranged in other forms, as long as it can provide a supporting force for the optical neural network layer 20, and the present invention does not limit this.

[0049] The physical architecture support layer 10 and the optical neural network layer 20 together constitute a physically reconfigurable optical network computing module for receiving and processing input data.

[0050] Exemplarily, the input data includes but is not limited to image data, and the present invention does not limit this.

[0051] According to an exemplary embodiment, the optical neural network layer 20 may at least include one or a combination of a convolutional layer, a pooling layer, a fully connected layer, and a non-linear layer.

[0052] For example, the combination method in the optical neural network layer 20 can be customized according to actual needs, and the number of layers in the optical neural network layer 20 can be customized according to actual needs (for example, the number of layers in the optical neural network layer 20 can be adjusted by adjusting training parameters and data sets).

[0053] According to an exemplary embodiment, the convolutional layer, the pooling layer, and the fully connected layer are physical mapping structures, and the physical mapping structures may include a substrate and optoelectronic devices disposed on the substrate, such that the optical neural network layer 20 can convert input data into optical processing signals based on diffraction optics and Fourier optics principles.

[0054] Optionally, the substrate may be a metasurface.

[0055] Optionally, the optoelectronic device at least includes one of a diffractive optical element and a modulator.

[0056] For example, the optoelectronic device includes, but is not limited to, active or passive optoelectronic components such as diffractive optical elements and liquid crystal phase modulators. For example, the optoelectronic device can be fabricated by CMOS (Complementary Metal-Oxide-Semiconductor) technology.

[0057] Figure 2a Schematic diagram showing the structure of the convolutional layer according to an embodiment of the present invention; Figure 2b Schematic diagram showing a calculation of the convolutional layer according to an embodiment of the present invention; Figure 2c Schematic diagram showing another calculation of the convolutional layer according to an embodiment of the present invention.

[0058] As Figure 2a shown, the convolutional layer may include three sub-convolutional layers (such as convolutional layer 1-1, convolutional layer 1-2, and convolutional layer 1-3). The three sub-convolutional layers are disposed on the metasurface, and the three sub-convolutional layers are cascaded through a three-layer network structure to jointly form one convolutional layer. The convolutional layer can implement feature extraction of input data.

[0059] As Figure 2b shown, the convolutional layer can perform a convolution operation on a single input data (such as image data) through m*n convolutional kernels (such as 200*200), and the calculation method of all convolutional kernels is a parallel calculation method.

[0060] Taking a single convolutional kernel as an example, the image data is Fourier-transformed through convolutional layer 1-1, then undergoes spatial propagation, and frequency-domain convolution is performed at convolutional layer 1-2 by setting the convolutional kernel template. After that, it undergoes spatial propagation again, and an inverse Fourier transform is performed at convolutional layer 1-3, and finally, the feature image (i.e., the eigenvalue light field) obtained after being processed by the convolutional layer is output.

[0061] As Figure 2c shown, the calculation principle of this convolutional layer can be that the spatial light field information can perform the calculation of the diffracted light field through the phase and amplitude modulation of the physical mapping structure (metasurface and optoelectronic devices) of this convolutional layer.

[0062] Figure 3a The structure diagram of the pooling layer according to an embodiment of the present invention is shown; Figure 3b A calculation diagram of the pooling layer according to an embodiment of the present invention is shown; Figure 3c Another calculation diagram of the pooling layer according to an embodiment of the present invention is shown.

[0063] As Figure 3a shown, this pooling layer can be set on the metasurface, and through a layer of network structure, it can achieve feature compression of the eigenvalue light field.

[0064] As shown in 3b, each convolutional kernel can correspond to a pooling filter, and its feature compression scale and compression ratio can correspond to the size of the output feature matrix. Exemplarily, the feature compression scale and compression ratio can be custom-set according to the actual needs of the user.

[0065] The pooling layer can adopt the method of average pooling and be adjusted according to the characteristics of the optoelectronic devices. As an embodiment, as Figure 3b shown, if the size of the pooling filter is 10*10, the eigenvalue light field obtained by the convolutional layer can obtain 400 pooled value light fields of 20*20 after being pooled by this pooling layer.

[0066] As Figure 3c shown, the calculation principle of this pooling layer can be that the spatial light field information can perform the calculation of the diffracted light field through the phase and amplitude modulation of the physical mapping structure (metasurface and optoelectronic devices) of this pooling layer.

[0067] Figure 4a The structure diagram of the fully connected layer according to an embodiment of the present invention is shown; Figure 4b A calculation diagram of the fully connected layer according to an embodiment of the present invention is shown; Figure 4c Another calculation diagram of the fully connected layer according to an embodiment of the present invention is shown.

[0068] As Figure 4aAs shown, the fully connected layer can be disposed on the metasurface. The fully connected layer can include two sub fully connected layers (such as fully connected layer 2-1 and fully connected layer 2-2), and the fully connected layer can be disposed after the convolutional layer 1-3. After the input optical field passes through the two sub fully connected layers, the output optical field can be obtained, and the output optical field is sent to the optoelectronic conversion module.

[0069] As Figure 4c shown, the calculation principle of the fully connected layer can be that the spatial optical field information can perform the calculation of the diffracted optical field through the phase and amplitude modulation of the physical mapping structure (metasurface and optoelectronic device) of the fully connected layer.

[0070] Figure 5 Fig. shows a calculation schematic diagram of the non-linear layer according to an embodiment of the present invention.

[0071] As Figure 5 shown, after the input optical field passes through the non-linear layer, the non-linear layer can perform non-linear modulation on the amplitude of the input optical field, and a characteristic image with light intensity change can be obtained.

[0072] According to an example embodiment, the non-linear layer includes but is not limited to a liquid crystal modulator, a phase change material modulator, etc. The non-linear layer can be disposed after any layer of the convolutional layer, the pooling layer, and the fully connected layer to perform corresponding optical field calculations.

[0073] The calculation principle of the non-linear layer can be that after the spatial optical field information passes through the physical mapping structure (such as modulators such as a liquid crystal modulator or a phase change material modulator) of the non-linear layer, the refractive index or absorption coefficient of the optical field at different positions can be modulated, so that the amplitude of the output optical field can be changed, and the network depth can be increased and the complexity of the task can be improved.

[0074] According to an example embodiment, the electrical neural network layer 30 is connected to the optical neural network layer 20, converts the optical processing signal into an electrical processing signal based on optoelectronic conversion, and completes task calculation according to the electrical processing signal.

[0075] For example, as Figure 1 shown, the electrical neural network layer 30 is disposed on the bottom side of the optical neural network layer 20, receives the optical processing signal output by the optical neural network layer 20, and converts the optical processing signal into an electrical processing signal. The electrical neural network layer 30 can obtain a final calculation result based on the calculation of the electrical neural network for the electrical processing signal.

[0076] Optionally, as Figure 1 shown, the electrical neural network layer 30 can further include an optoelectronic conversion module 31, a signal reading module 32, and an electrical calculation module 33.

[0077] The optoelectronic conversion module 31 receives the optical processing signal from the optical neural network layer 20 and converts the optical processing signal into an electrical processing signal. The signal reading module 32 reads the electrical processing signal. The electrical computing module 33 performs task calculations based on the electrical neural network according to the electrical processing signal from the signal reading module 32.

[0078] For example, the optoelectronic conversion module 31 can efficiently receive the optical processing signal from the optical neural network layer 20 and quickly and accurately convert it into an electrical processing signal, which can ensure the integrity and low loss of the signal during transmission. The signal reading module 32 can accurately read the electrical processing signal output by the optoelectronic conversion module 31, and through amplification and filtering processing, can ensure the high-fidelity and low-noise characteristics of the signal in subsequent processing. And the electrical computing module 33 can efficiently calculate the electrical processing signal provided by the signal reading module 32 based on the electrical neural network, which can support complex task processing, etc.

[0079] Through the above embodiments, the optoelectronic computing chip provided by the present invention includes a physical architecture support layer, an optical neural network layer, and an electrical neural network layer. The physical architecture support layer and the optical neural network layer constitute a physically reconfigurable optical network computing module, and the electrical neural network layer constitutes an electrical network computing module, so that the optoelectronic computing chip can integrate the functions of optoelectronic conversion and electrical neural network computing, and can achieve parallel computing of large-scale input data. And the optical neural network layer can include one or more combinations of a convolutional layer, a pooling layer, a fully connected layer, and a non-linear layer. By configuring the structure of the optical neural network layer accordingly, a reconfigurable optical neural network layer can be formed. The present invention can perform different algorithm constructions through the reconfigurable optical neural network layer, and can meet the intelligent computing tasks under complex application scenarios.

[0080] The optoelectronic computing chip provided by the present invention has characteristics such as high parallel computing, integrability, and reconfigurability, which can improve the application ability of the optoelectronic computing chip in complex tasks.

[0081] Optionally, the optical neural network layer 20 includes at least one convolutional layer.

[0082] As an example embodiment, the optical neural network layer 20 includes a convolutional layer. After the input data passes through this convolutional layer, the optical neural network layer 20 can extract the eigenvalue optical field and send the eigenvalue optical field to the electrical neural network layer 30, so that the electrical neural network layer 30 can perform subsequent calculation operations according to the eigenvalue optical field.

[0083] Optionally, the optical neural network layer 20 includes at least one convolutional layer and at least one pooling layer.

[0084] As another exemplary embodiment, the optical neural network layer 20 includes a convolutional layer and a pooling layer. After the input data passes through the convolutional layer, the convolutional layer can extract an eigenvalue optical field, and the pooling layer then performs feature compression on the eigenvalue optical field to obtain a compressed optical field, and sends the compressed optical field to the electrical neural network layer 30. So that the electrical neural network layer 30 can perform subsequent calculation operations based on the compressed optical field.

[0085] Optionally, the optical neural network layer 20 includes at least one fully connected layer.

[0086] As an exemplary embodiment, the optical neural network layer 20 includes a fully connected layer. After the input data passes through the fully connected layer, the optical neural network layer 20 can obtain an eigenvalue optical field and a corresponding output classification value, and send the eigenvalue optical field and the corresponding output classification value to the electrical neural network layer 30. So that the electrical neural network layer 30 can perform subsequent calculation operations based on the eigenvalue optical field and the corresponding output classification value.

[0087] Optionally, the optical neural network layer 20 includes at least one non-linear layer and one or more of a convolutional layer, a pooling layer, and a fully connected layer.

[0088] As an exemplary embodiment, the non-linear layer can cooperate with other structural layers (such as convolutional layers, pooling layers, and fully connected layers, etc.) so that the optical neural network layer 20 can obtain the amplitude of the corresponding output optical field, and send the amplitude of the output optical field to the electrical neural network layer 30. So that the electrical neural network layer 30 can perform subsequent calculation operations based on the amplitude of the output optical field.

[0089] Optionally, the optical neural network layer 20 includes at least one convolutional layer, at least one pooling layer, at least one fully connected layer, and at least one non-linear layer.

[0090] For example, as Figure 6 shown, the optical neural network layer 20 may include two convolutional layers 21, two pooling layers 22, two fully connected layers 23, and one non-linear layer 24.

[0091] As an embodiment, as Figure 7 shown, assume that the input data is an image data with a size of 1000*1000. After the image data passes through the C1 convolutional layer, 25 eigenvalue optical fields (i.e., eigen-optical field matrices) with a size of 200*200 can be obtained. Among them, the convolutional kernel is an array of 25 phase modulation templates with a pixel size of 200*200 (formed by a metasurface device).

[0092] After that, the eigenvalue optical field is input into the S2 pooling layer. Among them, the pooling filter is an array of 400 phase modulation templates with a scale of 10*10 pixels (formed by a metasurface device). The eigenvalue optical field corresponding to each convolution kernel (200*200) is compressed into a compressed optical field with a scale of 400 pixels of 20*20.

[0093] Based on the same mechanism as above, after passing through the second convolutional layer (C3 convolutional layer) and the second pooling layer (S4 pooling layer) again, 160,000 compressed optical fields with a scale of 2*2 pixels can be obtained. During this process, all optical field calculations are parallel calculations.

[0094] After that, the compressed optical field passes through two fully connected layers (F5 fully connected layer and F6 fully connected layer) and a non-linear layer for optical field calculation and feature classification, so that 10 optical field intensities can be output to the electrical neural network layer 30. The electrical neural network layer 30 calculates the received optical field intensities, so that the recognition result can be obtained.

[0095] Through the above embodiments, the present invention can realize the physical mapping of a photoelectric computing chip to a network structure such as a LeNet-5 network structure through the structure reconstruction of the optical neural network layer 20, construct an optical neural network, and thus realize neural network computing.

[0096] According to another aspect of the present invention, the present invention also provides an electronic device. The electronic device includes the photoelectric computing chip as described above.

[0097] Finally, it should be noted that the above are only the preferred embodiments of the present invention and are not used to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, for those skilled in the art, they can still modify the technical solutions of the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An integrated reconfigurable optoelectronic computing chip, characterized in that, Comprising: Physical architecture support layer; Optical neural network layer, which is supported and connected to the physical architecture support layer. The optical neural network layer includes at least one or a combination of a convolutional layer, a pooling layer, a fully connected layer, and a non-linear layer. The optical neural network layer converts input data into an optical processing signal based on the principles of diffractive optics and Fourier optics; Electrical neural network layer, which is connected to the optical neural network layer, converts the optical processing signal into an electrical processing signal based on photoelectric conversion, and completes task calculation according to the electrical processing signal; Among them, the convolutional layer, the pooling layer, and the fully connected layer are physical mapping structures, and the physical mapping structure includes a substrate and optoelectronic devices disposed on the substrate.

2. The optoelectronic computing chip according to claim 1, wherein The electrical neural network layer further includes: Photoelectric conversion module, which receives the optical processing signal from the optical neural network layer and converts the optical processing signal into an electrical processing signal; Signal readout module, which reads the electrical processing signal; Electrical calculation module, which performs task calculation based on the electrical neural network according to the electrical processing signal from the signal readout module.

3. The optoelectronic computing chip according to claim 1, wherein The optical neural network layer includes at least one of the convolutional layers.

4. The optoelectronic computing chip according to claim 1, wherein The optical neural network layer includes at least one of the convolutional layers and at least one of the pooling layers.

5. The optoelectronic computing chip according to claim 1, characterized in that, The optical neural network layer includes at least one of the fully connected layers.

6. The optoelectronic computing chip according to claim 1, wherein The optical neural network layer includes at least one of the non-linear layers and one or more of the convolutional layer, the pooling layer, and the fully connected layer.

7. The optoelectronic computing chip according to claim 1, wherein The optical neural network layer includes at least one of the convolutional layers, at least one of the pooling layers, at least one of the fully connected layers, and at least one of the non-linear layers.

8. The optoelectronic computing chip according to claim 1, wherein The substrate is a metasurface.

9. The optoelectronic computing chip according to claim 1, wherein The optoelectronic device includes at least one of a diffractive optical element and a modulator.

10. An electronic device, characterized in that, Including the optoelectronic computing chip according to any one of claims 1-9.

Images