CN113805641B - Photonic neural network - Google Patents
Photonic neural network Download PDFInfo
- Publication number
- CN113805641B CN113805641B CN202111116962.XA CN202111116962A CN113805641B CN 113805641 B CN113805641 B CN 113805641B CN 202111116962 A CN202111116962 A CN 202111116962A CN 113805641 B CN113805641 B CN 113805641B
- Authority
- CN
- China
- Prior art keywords
- modulator
- convolution kernel
- array
- light
- neural network
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06E—OPTICAL COMPUTING DEVICES; COMPUTING DEVICES USING OTHER RADIATIONS WITH SIMILAR PROPERTIES
- G06E1/00—Devices for processing exclusively digital data
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Optical Modulation, Optical Deflection, Nonlinear Optics, Optical Demodulation, Optical Logic Elements (AREA)
Abstract
The invention provides a photonic neural network, which comprises a laser array, a modulator array, a first beam combiner, a modulator, a beam splitter, a delay line, a second beam combiner and a detector, wherein the first beam combiner is connected with the modulator; the laser array emits a plurality of laser beams with different wavelengths; the modulator array modulates the laser output by the laser array, and loads the weight of the convolution kernel to the light emitted by the laser array; the first beam combiner combines the modulated multiple laser beams to obtain a first beam; the modulator modulates the first light beam and loads information to be processed on the first light beam; the beam splitter splits a first light beam loaded with information to be processed to obtain multiple light beams with different wavelengths; the delay line enables the multiple beams of light after beam splitting to generate equal differential delay; the second beam combiner combines the multiple beams with the same delay to obtain a second beam; the detector receives the second light beam to realize the dislocation addition of the signals. The number of control circuits is reduced while also eliminating calculation errors due to performance differences.
Description
Technical Field
The invention relates to the field of integrated photons, in particular to a photonic neural network.
Background
In recent years, artificial intelligence technology has penetrated aspects of our social lives. In particular, in the last decade, the operation scale of artificial intelligence technology is larger and larger, and the degree of dependence on hardware is also higher and higher. According to OpenAI statistics of artificial intelligence research organization, floating point calculation amount of the deep artificial neural network in 2012-2020 increases at a striking rate, and the floating point calculation amount is doubled every 3-4 months, so that the increasing rate of Moore's law of an integrated circuit is far exceeded. Meanwhile, moore's law has gradually approached the physical limits of semiconductor technology and the limitations of microelectronic technology fabrication processes, and is facing failure problems. Since more than 90% of calculation amount in the artificial intelligence algorithm is matrix operation, the photon network is very suitable for matrix calculation. Therefore, the development of photonic neural networks is an effective solution for large-scale artificial intelligence operations. However, the on-chip photonic neural network that has been proposed at present is mainly based on a mach-zehnder interferometer (MZI), and when the scale of the convolution kernel (n×n) becomes large, that is, the matrix operation scale becomes large, the number of MZI increases exponentially, which greatly increases the area of the photonic neural network. Meanwhile, each MZI in the photonic neural network needs two electrical ports to control the phase change of the MZI, and a large-scale photonic neural network needs a larger number of electrical ports. In addition, each input end and each output end of the photonic neural network are required to be connected with a modulator and a detector to respectively realize the input of the signal to be processed and the output of the signal after the processing. As the size of the photonic neural network convolution kernel increases, the number of modulators and detectors increases linearly. Since signals between different ports of the photonic neural network are input/output through different modulators/detectors. Therefore, performance differences between modulators/detectors may result in different inputs/outputs occurring after the same signal passes. The effect of this difference will be greater, especially for high precision operations. For example, when the performance difference between modulators/detectors connected to different channels of a photonic neural network is greater than 1%, the accuracy of the operation of the neural network will be less than 6 bits. This greatly limits the computational accuracy of the photonic neural network.
Disclosure of Invention
The invention provides a photonic neural network, comprising: the system comprises a laser array, a modulator array, a first beam combiner, a modulator, a beam splitter, a delay line, a second beam combiner and a detector;
the laser array is used for emitting a plurality of laser beams with different wavelengths;
the modulator array is used for modulating the laser output by the laser array 1 and loading the weight of the convolution kernel to the light with different wavelengths emitted by the laser array;
the first beam combiner is used for combining the multiple beams of laser beams with the loaded weights to obtain a first beam;
the modulator is used for modulating the first light beam and loading information to be processed on the first light beam;
the beam splitter is used for splitting the first light beam loaded with the information to be processed to obtain multiple light beams with different wavelengths;
the delay line is used for enabling the multiple beams of light after beam splitting to generate equal difference delay;
the second beam combiner is used for combining the multiple beams with the same delay to obtain a second beam;
the detector is used for receiving the second light beam and realizing the dislocation addition of signals.
Optionally, the method further comprises: a phase shifter;
the phase shifter is arranged between the delay line and the second beam combiner;
the phase shifter is used for changing the group refractive index n of the waveguide g A change in the amount of delay in the delay line, i.e. a change in the arithmetic delay, is caused.
Optionally, the method further comprises: optical wire bonding;
the optical wire bond is used to connect the modulator array with the first combiner;
the multiple beams of light modulated by the modulator array enter the first beam combiner through the optical wire bond.
Optionally, a delay tolerance Δτ of the arithmetic delay and a bandwidth BW of the modulator m Inversely proportional;
the length of the delay line is proportional to the speed of light c and the delay tolerance, and is proportional to the group refractive index n of the waveguide g Inversely proportional.
Optionally, the number of weights in the channel number convolution kernel of the laser array 1; for an n convolution kernel, the number of channels is n 2 The method comprises the steps of carrying out a first treatment on the surface of the For an n m convolution kernel, the number of channels is nm, where n and m are integers of any 1 or more.
Optionally, the first beam combiner, the beam splitter or the second beam combiner is composed of: arrayed waveguide grating, multimode interference coupler or broadband directional coupler.
Optionally, for an n×n convolution kernel, the operating bandwidth BW of the modulator array e Operating bandwidth BW with said modulator 4 m The equation is satisfied as follows:
BW m ≥N×n 2 ×BW e wherein, N represents the operation times of the convolution kernel in the artificial intelligent architecture, and N represents the size of the convolution kernel;
for an n×m convolution kernel, the operating bandwidth BW of the modulator array e Operating bandwidth BW with said modulator m The equation is satisfied as follows:
BW m ≥N×n×m×BW e where N represents the number of operations of the convolution kernel in the artificial intelligence architecture and nxm represents the size of the convolution kernel.
Optionally, typical values of the center wavelength interval δλ of the laser array and the modulator array are between 0.4nm and 4.8nm, satisfying the following formula:
δλ<ΔΛ/n 2 where n represents the size of the convolution kernel and Λ represents the wavelength operating range of the modulator.
Optionally, the typical value of the wavelength interval of the laser array is between 0.2nm and 4.8nm, the typical value of the wavelength interval of the first beam combiner is between 0.2nm and 4.8nm, the typical value of the wavelength interval of the beam splitter is between 0.2nm and 4.8nm, and the typical value of the wavelength interval of the second beam combiner is between 0.2nm and 4.8 nm.
Optionally, the laser array and the modulator array are formed from a group iii-v semiconductor material; the first beam combiner, the modulator, the beam splitter, the delay line, the second beam combiner and the detector are made of silicon-based or carbon-based substrate materials.
The photonic neural network provided by the embodiment of the application comprises a laser array, a modulator array, a first beam combiner, a modulator, a beam splitter, a delay line, a second beam combiner and a detector; the laser array is used for emitting a plurality of laser beams with different wavelengths; the modulator array is used for modulating the laser output by the laser array and loading the weight of the convolution kernel to the light with different wavelengths emitted by the laser array; the first beam combiner is used for combining the multiple laser beams with the loaded weights to obtain a first beam; the modulator is used for modulating the first light beam and loading information to be processed on the first light beam; the beam splitter is used for splitting the first light beam loaded with the information to be processed to obtain multiple light beams with different wavelengths; the delay line is used for enabling the multiple beams of light after beam splitting to generate equal difference delay; the second beam combiner is used for combining the multiple beams with the same delay to obtain a second beam; the detector is used for receiving the second light beam and realizing the dislocation addition of signals.
According to one or more technical schemes provided by the embodiment of the application, information to be processed can be loaded on a plurality of light beams through a modulator, and signals on the plurality of light beams are received through a detector, so that dislocation addition is completed. The method avoids a plurality of control circuits needed by a plurality of modulators and a plurality of detectors, reduces the number of the control circuits, and simultaneously eliminates calculation errors caused by performance differences among the modulators and performance differences among the detectors.
Drawings
Further details, features and advantages of the present disclosure are disclosed in the following description of exemplary embodiments, with reference to the following drawings, wherein:
FIG. 1 shows a schematic diagram of a photonic neural network in accordance with exemplary embodiments of the present invention;
FIG. 2 shows a convolution kernel diagram in accordance with an exemplary embodiment of the present disclosure;
fig. 3 shows a schematic diagram of a misalignment addition calculation according to an exemplary embodiment of the invention.
Detailed Description
Embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. While certain embodiments of the present disclosure have been shown in the accompanying drawings, it is to be understood that the present disclosure may be embodied in various forms and should not be construed as limited to the embodiments set forth herein, but are provided to provide a more thorough and complete understanding of the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are for illustration purposes only and are not intended to limit the scope of the present disclosure.
It should be understood that the various steps recited in the method embodiments of the present disclosure may be performed in a different order and/or performed in parallel. Furthermore, method embodiments may include additional steps and/or omit performing the illustrated steps. The scope of the present disclosure is not limited in this respect.
The term "including" and variations thereof as used herein are intended to be open-ended, i.e., including, but not limited to. The term "based on" is based at least in part on. The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments. Related definitions of other terms will be given in the description below. It should be noted that the terms "first," "second," and the like in this disclosure are merely used to distinguish between different devices, modules, or units and are not used to define an order or interdependence of functions performed by the devices, modules, or units.
It should be noted that references to "one", "a plurality" and "a plurality" in this disclosure are intended to be illustrative rather than limiting, and those of ordinary skill in the art will appreciate that "one or more" is intended to be understood as "one or more" unless the context clearly indicates otherwise.
The names of messages or information interacted between the various devices in the embodiments of the present disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.
An embodiment of the present invention provides a photonic neural network, as shown in fig. 1, including: a laser array 1, a modulator array 2, a first beam combiner 3, a modulator 4, a beam splitter 5, a delay line 6, a second beam combiner 7 and a detector 8.
The laser array 1 is used for emitting a plurality of laser beams with different wavelengths, and the laser array part is used as a light source of a photonic neural network, wherein the number of the laser beams is equal to the number of weights in a convolution kernel.
The modulator array 2 is used for modulating the laser output by the laser array 1, and loading the weight of the convolution kernel onto the light with different wavelengths emitted by the laser array 1, so as to realize a matrix network in the photonic neural network. Wherein the use of the laser array 1 and the modulator array 2 may reduce the overall volume of the device. Illustratively, the modulator array loads the weights in the convolution kernel onto the laser light emitted from the laser array, and each beam of light modulated by the modulator array 2 includes one weight in the convolution kernel. The weights in the convolution kernels may be loaded regularly according to the wavelength of the light beam emitted by the laser array 1, for example, the corresponding convolution kernel weights may be loaded on the light beam in the order from the large wavelength to the small wavelength or from the small wavelength to the large wavelength, and the specific method for loading the weights is not limited in this embodiment.
To ensure that each beam of light emitted in the laser array 1 is modulated, the number of channels of the laser array 1 and the number of channels of the modulator array 2 are equal.
The modulator array 2 may be any modulator array capable of modulating light, including but not limited to an electro-absorption modulator, an electro-optical modulator, a thermo-optical modulator, an acousto-optic modulator, and the like, and the present embodiment does not limit the kind of the modulator array 2.
The first beam combiner 3 is configured to combine the multiple laser beams loaded with the weights to obtain a first beam. After passing through the first beam combiner 3, the multiple lasers with the convolution kernel weight information are combined into a first beam, which can be modulated by a modulator, and the information to be processed is loaded on the beam. Modulation of each beam by multiple modulators is avoided because the modulator performance differences introduce errors. The first beam combiner 3 is used to combine a plurality of light beams, and illustratively, the first beam combiner includes, but is not limited to, an arrayed waveguide grating, a multimode interference coupler, a cascaded broadband directional coupler, and the like, and the present embodiment is not limited to a specific type of the first beam combiner 3. In order to ensure that each beam is combined, the number of channels of the first beam combiner 3 is identical to the number of channels of the laser array 1.
The modulator 4 is arranged to modulate the first light beam and to load the information to be processed onto the first light beam. The information to be processed may be information such as an image or a sound, and the type of the information to be processed is not limited. The information to be processed is loaded onto the first light beam by the modulator 4, each light beam combining the first light beam is loaded with the information to be processed due to the independent propagation principle of the light, and multiplication of the convolution kernel weight and the information to be processed is completed. Illustratively, the modulator 4 includes, but is not limited to, a Mach-Zehnder modulator, where the requirement for the modulator is that the bandwidth be large, and that it be capable of covering all wavelengths in the laser array, and the specific type of modulator 4 is not limited herein.
In some alternative embodiments, when modulating the information to be processed onto the first light beam, each light beam constituting the first light beam is modulated with the entire information to be processed. The information loaded on each beam at this time is the product of the convolution kernel weight of the beam loaded at the modulator array 2 and the multiplication of the total information to be processed. In the convolutional neural network, only the multiplication result of the weight loaded by each beam of light and the information to be processed in a specific position is needed to be added, and the positions of the information to be processed corresponding to each beam of light are different, so that the information loaded on each beam of light is needed to be added in a staggered manner.
The beam splitter 5 is used for splitting a first light beam loaded with information to be processed to obtain multiple light beams with different wavelengths. The first beam after passing through the modulator 4 comprises each beam emitted from the laser array 1, wherein each beam performs a multiplication of the convolution kernel weights carried by the beam with the information to be processed. The first light beam is split by the beam splitter 5, and in order to ensure that the split light and the light passing through the first beam combiner 3 are in one-to-one correspondence, the number of channels of the beam splitter 5 is equal to that of the channels of the first beam combiner 3. Illustratively, the beam splitter 5 may be an arrayed waveguide grating, a multimode interference coupler, a cascaded broadband directional coupler, or the like, and the specific type of the beam splitter 5 is not limited herein.
The delay line 6 is used for generating an equal delay for the multiple beams of light after beam splitting, wherein the equal delay is different in delay amount of each beam of light, but the delay amount is in an equal-difference array and is used for misplacement of information loaded on each beam of light, so that information detected by the detector at the same moment is exactly the information in the same convolution kernel. In order to ensure that each beam of light has a delay, the number of delay lines is equal to the number of channels of the laser array 1. The structure of the delay line may be a single-mode waveguide or a multi-mode waveguide, and the width and the height of the waveguide may be designed according to the requirements of the photonic neural network, and the structure, the width and the height of the delay line are not limited herein. The waveguide structure of the delay line may include a bar waveguide, a ridge waveguide, a slab waveguide, a step waveguide, or a taper waveguide, for example, and is not limited herein.
The second beam combiner 7 is used for combining the multiple beams with the same delay to obtain a second beam; an equal delay is created between the multiple beams that combine the second beam. Illustratively, the second beam combiner 7 includes, but is not limited to, an arrayed waveguide grating, a multimode interference coupler, a cascaded broadband directional coupler, or the like, and the kind of the beam combiner is not limited herein.
The detector 8 is arranged to receive the second light beam and to effect a staggered addition of the signals. Most light with different delayed wavelengths in the second light beam irradiates the detector 8, and the detector performs addition operation on the light signals received at each moment, and because equal difference time delay is generated between the light with different wavelengths, the detector can perform addition operation of information in a convolution kernel, and the function of the whole convolution neural network is realized. The detector may be a photodetector, for example, and the type of detector is not limited herein.
Illustratively, as shown in fig. 2 and 3, the misalignment addition process of the exemplary embodiment of the present invention is as follows: fig. 2 shows a portion of information to be processed, and the thickened block in fig. 2 shows a convolution kernel in the convolutional neural network, the numbers in the convolution kernel merely indicating the position of the portion of information in the convolution kernel, and not the information included in the portion. FIG. 3 shows a process of photonic neural network misalignment addition, λ 1 -λ 9 The nine light beams with different wavelengths are sequentially represented, the detector 8 only receives information on a plurality of columns at the same time and performs addition operation on the received information, the vertical frame in fig. 3 corresponds to the convolution kernel in fig. 2, the detector 8 receives information in the vertical frame in fig. 3 at a certain time and adds the received information, and therefore addition operation on the content in the convolution kernel is completed, and the function of the convolution neural network is achieved. Numbers in fig. 3The word indicates only the correspondence between the position and the information in the convolution kernel in fig. 2, and does not indicate the information included in the portion.
Through the photonic neural network provided by the embodiment, the extraction of the characteristics of the information to be processed can be completed through one detector and one modulator, so that the function of the convolutional neural network is realized. The problems of large area and low integration level of the photonic neural network caused by a plurality of control loops required by a plurality of modulators are avoided. With only one detector and modulator, systematic errors due to performance differences between individual modulators and/or detectors are also prevented. The photonic neural network with higher integration level and calculation precision can be realized.
In some alternative embodiments, to ensure the accuracy of the resulting equi-delay between the beams, the convolutional neural network may further comprise a phase shifter 9 to prevent information in not one convolution kernel from being illuminated on the detector at the same time; the phase shifter is arranged between the delay line 6 and the second beam combiner 7; the phase shifter 9 is used to change the group refractive index n of the waveguide g Causing a change in the amount of delay in the delay line. The delay amount of each beam can be controlled to be strictly equal by changing the delay amount in the delay line through the phase shifter, and accurate equi-differential delay is generated among the beams. The phase shifter 9 may include a thermo-optic phase shifter, an electro-optic phase shifter, an acousto-optic phase shifter, or the like, and the specific kind of phase shifter is not limited herein.
In some alternative embodiments, the weight of the convolution kernel may also be represented by the intensity of the light beam. Illustratively, the modulator array 2 may load the convolution kernel weight information by varying the intensity of the light beam emitted by the laser array 1. For example, when the first light intensity represents that the convolution kernel weight is 1, the light intensity of the light beam is adjusted to be half of the first light intensity when the weight represented by the light beam is 0.5, and when the weight represented by the light beam is 3, the light intensity of the light beam can be adjusted to be 3 times of the first light intensity. The present description only shows one possible implementation, and the specific modulation scheme of the modulator array 2 is not limited. The convolution kernel weight information may also be represented by changing the phase, polarization state, etc. of the light, for example.
In some alternative embodiments, optical wire bonds 10 may also be included to simplify the installation process of the optical system; an optical wire bond 10 is used to connect the modulator array 2 with the first combiner 3; the multiple beams of light modulated by the modulator array 2 enter the first beam combiner 3 through the optical wire bond 10. Optical wire bonding is an effective way to achieve monolithic integration between devices on substrates of different materials, by 3D printing the waveguide structure to connect the devices on two substrates of different materials together. Compared with other hybrid integration technologies, the scheme has lower cost and higher yield. In the alternative embodiment, the hybrid integration of devices among different substrates can be effectively realized, the manufacturing cost is reduced, and the yield is improved.
In some alternative embodiments, the delay tolerance Δτ of the arithmetic delay is equal to the bandwidth BW of the modulator m Inversely proportional, i.e. Δτ=1/BW m Otherwise, the signals are misplaced during addition, and the recognition accuracy is reduced. To ensure that the delay line 6 provides an accurate amount of delay, the length ΔL of the delay line 6 may be proportional to the speed of light c and the delay tolerance Δτ, and the group index of refraction n of the material used for the delay line 6 g Inversely proportional, i.e. Δl=c Δτ/n g Ensuring that the light can produce accurate delay after passing through the corresponding delay line 6.
In some alternative embodiments, the number of channels of the laser array 1 is equal to the number of weights in the convolution kernel; for an n convolution kernel, the number of channels is n 2 The method comprises the steps of carrying out a first treatment on the surface of the For an n m convolution kernel, the number of channels is nm, where n and m are integers of any 1 or more. For example, for a square convolution kernel, the scale of the convolution kernel is 3 and the number of channels of the laser array is 9. For non-square convolution kernels of unequal length and width, n is 3 and m is 4, then the number of channels of the laser array is 12. Each beam of light emitted by the laser represents a weight in a convolution kernel, which is a square matrix in a convolutional neural network.
In some alternative embodiments, for an n×n convolution kernel, the following formula is satisfied between the operating bandwidth BWe of modulator array 2 and the operating bandwidth BWm of modulator 4:
BWm≥N×n 2 x BWe, wherein,n represents the number of operations of the convolution kernel in the artificial intelligence calculation process, and N represents the size of the convolution kernel.
For an n×m convolution kernel, the operating bandwidth BW of the modulator array 2 e And the operating bandwidth BW of modulator 4 m The equation is satisfied as follows:
BW m ≥N×n×m×BW e where N represents the number of operations of the convolution kernel in the artificial intelligence architecture and nxm represents the size of the convolution kernel.
Because the artificial intelligent model is a convolution kernel and performs convolution operation step by step with all information in the picture, the loading frequency of the information to be processed on the chip is higher than the change frequency of the weight in the convolution kernel. If the loading frequency of the information to be processed is smaller than the weight change frequency, the accuracy is reduced or the bandwidth is wasted in the working process, namely the working bandwidth is smaller than the actual bandwidth of the device.
In some alternative embodiments, typical values for the center wavelength interval δλ of the laser array 1 and modulator array 2 are between 0.4nm and 4.8nm, with specific magnitudes satisfying the following formula:
δλ<ΔΛ/n 2 where n represents the size of the convolution kernel and Λ represents the wavelength operating range of modulator 4.
If the wavelength interval is too large, the working bandwidth of the modulator 4 is required to be wide enough, otherwise, the information to be processed is loaded on different wavelengths and different occurs, and the network accuracy is reduced.
In some alternative embodiments, the typical value of the wavelength interval of the laser array 1 is between 0.2nm and 4.8nm, the typical value of the wavelength interval of the first beam combiner 3 is between 0.2nm and 4.8nm, the typical value of the wavelength interval of the beam splitter 5 is between 0.2nm and 4.8nm, and the typical value of the wavelength interval of the second beam combiner 7 is between 0.2nm and 4.8 nm.
The typical values described above relate to the current fabrication capabilities of the device, and these values are determined based on the current process fabrication capabilities. If the wavelength interval is too large, the working bandwidth of the modulator 4 is required to be sufficiently wide, otherwise, the loading of the information to be processed onto different wavelengths occurs with different results, and the network accuracy is reduced.
In order to ensure the accuracy of the operation result of the photonic neural network, the output optical power among the channels of the laser array 1 and the characteristics of the modulator array 2 need to be tested and calibrated in advance. The operating bandwidth of the detector 8 is equal to the operating bandwidth of the modulator 4, and the wavelength operating range of the detector 8 should be greater than or equal to the wavelength operating range of the modulator 4. The center wavelength of the laser array 1 coincides with the center wavelength of the modulator array 2. The center wavelengths of the first beam combiner 3, the beam splitter 5, and the second beam combiner 7 coincide with the center wavelength of the laser array.
In some alternative embodiments, the laser array 1 and the electroabsorption modulator array 2 may be fabricated from a III-V semiconductor material; the III-V semiconductor material is gallium arsenic or indium phosphorus or a III-V semiconductor material which is lattice matched with one of the gallium arsenic or indium phosphorus. The modulator 4, delay line 6, detector 8 and phase shifter 9 may be fabricated from silicon-based or carbon-based substrate materials. This example only represents an alternative implementation and is not limited to the manner and materials of fabrication of the optoelectronic device.
In some alternative embodiments, the optical devices of two different substrates may be connected together in a hybrid integrated manner to form a photonic neural network. This embodiment expresses only one alternative implementation, and does not limit the way in which the optical neural network is integrated.
In summary, the photonic neural network in the embodiment of the invention only needs a group of modulators/detectors to respectively realize the input and output of signals, thereby improving the calculation accuracy of the photonic neural network. Meanwhile, a matrix operation part in the photonic neural network is realized by a scheme of combining a modulator and a delay line, and the number of electrical ports is not more than 1/n of a matrix structure formed by MZIs; the whole area of the network is not more than 1/n of the photonic neural network formed by the MZIs.
Although embodiments of the present invention have been described in connection with the accompanying drawings, various modifications and variations may be made by those skilled in the art without departing from the spirit and scope of the invention, and such modifications and variations are within the scope of the invention as defined by the appended claims.
Claims (10)
1. A photonic neural network, comprising:
the device comprises a laser array (1), a modulator array (2), a first beam combiner (3), a modulator (4), a beam splitter (5), a delay line (6), a second beam combiner (7) and a detector (8);
the laser array (1) is used for emitting a plurality of laser beams with different wavelengths;
the modulator array (2) is used for modulating the laser output by the laser array (1) and loading the weight of the convolution kernel to the laser with different wavelengths emitted by the laser array (1);
the first beam combiner (3) is used for combining the multiple laser beams with the loaded weights to obtain a first beam;
the modulator (4) is used for modulating the first light beam and loading information to be processed on the first light beam;
the beam splitter (5) is used for splitting the first light beam loaded with the information to be processed to obtain multiple light beams with different wavelengths;
the delay line (6) is used for enabling the multiple beams of split light to generate equal difference delay;
the second beam combiner (7) is used for combining the multiple beams with the same delay to obtain a second beam;
the detector (8) is used for receiving the second light beam and realizing the dislocation addition of signals.
2. The photonic neural network of claim 1, further comprising: a phase shifter (9);
the phase shifter (9) is arranged between the delay line (6) and the second beam combiner (7);
the phase shifter (9) is used for changing the group refractive index n of the waveguide g Causing a change in the amount of delay in the delay line.
3. The photonic neural network of claim 1, further comprising: an optical wire bond (10);
the optical wire bond (10) is used for connecting the modulator array (2) and the first beam combiner (3);
the multiple beams of light modulated by the modulator array (2) enter the first beam combiner (3) through the optical wire bond (10).
4. The photonic neural network of claim 1,
the delay tolerance Deltaτ of the equi-delay and the bandwidth BW of the modulator m Inversely proportional;
the length of the delay line (6) is proportional to the speed of light c and the delay tolerance Deltaτ, and is proportional to the group refractive index n of the waveguide g Inversely proportional.
5. The photonic neural network according to claim 1, characterized in that the number of channels of the laser array (1) is equal to the number of weights in the convolution kernel; for an n convolution kernel, the number of channels is n 2 The method comprises the steps of carrying out a first treatment on the surface of the For an n m convolution kernel, the number of channels is nm, where n and m are integers of any 1 or more.
6. The photonic neural network according to claim 1, characterized in that the first beam combiner (3), the beam splitter (5) or the second beam combiner (7) is constituted by:
arrayed waveguide gratings, multimode interference couplers or broadband directional couplers.
7. The photonic neural network according to claim 1, characterized in that for a convolution kernel of nxn, the operating bandwidth BW of the modulator array (2) e And the operating bandwidth BW of the modulator (4) m The equation is satisfied as follows:
BW m ≥N×n 2 ×BW e wherein, N represents the operation times of a convolution kernel in the artificial intelligent architecture, and N multiplied by N represents the size of the convolution kernel;
for an n×m convolution kernel, the operating bandwidth BW of the modulator array (2) e And the operating bandwidth BW of the modulator (4) m The equation is satisfied as follows:
BW m ≥N×n×m×BW e where N represents the number of operations of the convolution kernel in the artificial intelligence architecture and nxm represents the size of the convolution kernel.
8. The photonic neural network according to claim 1, characterized in that the typical value of the center wavelength interval δλ of the laser array (1) and the modulator array (2) is between 0.4nm and 4.8nm, satisfying the following formula:
δλ<ΔΛ/n 2 where n represents the size of the convolution kernel and Λ represents the wavelength operating range of the modulator (4).
9. The photonic neural network according to claim 1, characterized in that the typical value of the wavelength interval of the laser array (1) is between 0.2nm-4.8 nm; typical values of the wavelength interval of the first beam combiner (3) are between 0.2nm and 4.8 nm; typical values of the wavelength interval of the beam splitter (5) are between 0.2nm and 4.8 nm; the wavelength interval of the second beam combiner (7) has a typical value between 0.2nm and 4.8 nm.
10. The photonic neural network according to any one of claims 1 to 9, characterized in that the laser array (1) and the modulator array (2) are of a group iii-v semiconductor material; the first beam combiner (3), the modulator (4), the beam splitter (5), the delay line (6), the second beam combiner (7) and the detector (8) are made of silicon-based or carbon-based substrate materials.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111116962.XA CN113805641B (en) | 2021-09-23 | 2021-09-23 | Photonic neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111116962.XA CN113805641B (en) | 2021-09-23 | 2021-09-23 | Photonic neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113805641A CN113805641A (en) | 2021-12-17 |
| CN113805641B true CN113805641B (en) | 2023-08-01 |
Family
ID=78896480
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111116962.XA Active CN113805641B (en) | 2021-09-23 | 2021-09-23 | Photonic neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113805641B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114819132B (en) * | 2022-06-27 | 2022-10-11 | 之江实验室 | Photon two-dimensional convolution acceleration method and system based on time-wavelength interleaving |
| CN115222035B (en) * | 2022-09-20 | 2022-12-30 | 之江实验室 | Photon neural network convolution acceleration chip |
| CN115508958B (en) * | 2022-10-08 | 2024-05-24 | 深圳中科天鹰科技有限公司 | Photonic chip based on optical neural network |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111860822A (en) * | 2020-07-20 | 2020-10-30 | 联合微电子中心有限责任公司 | All-optical nonlinear activation function implementation method and device of optical neural network |
| CN112001487A (en) * | 2020-07-20 | 2020-11-27 | 联合微电子中心有限责任公司 | Photon neural network |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7512573B2 (en) * | 2006-10-16 | 2009-03-31 | Alcatel-Lucent Usa Inc. | Optical processor for an artificial neural network |
| WO2019014345A1 (en) * | 2017-07-11 | 2019-01-17 | Massachusetts Institute Of Technology | Optical ising machines and optical convolutional neural networks |
-
2021
- 2021-09-23 CN CN202111116962.XA patent/CN113805641B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111860822A (en) * | 2020-07-20 | 2020-10-30 | 联合微电子中心有限责任公司 | All-optical nonlinear activation function implementation method and device of optical neural network |
| CN112001487A (en) * | 2020-07-20 | 2020-11-27 | 联合微电子中心有限责任公司 | Photon neural network |
Non-Patent Citations (1)
| Title |
|---|
| 基于优化神经网络的光纤激光器的最优设计;申玉霞;李飞;;激光杂志(第02期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113805641A (en) | 2021-12-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113805641B (en) | Photonic neural network | |
| WO2022016894A1 (en) | Photonic neural network | |
| CN109639359B (en) | Photonic neural network convolution layer chip based on micro-ring resonator | |
| US11269179B2 (en) | Photonic filter bank system and method of use | |
| US12032210B2 (en) | Method of optical modulation for photonic computing | |
| US12085790B2 (en) | Photonic bandgap phase modulator, optical filter bank, photonic computing system, and methods of use | |
| KR20230129447A (en) | Balanced photonic architecture for matrix computation | |
| JP4436451B2 (en) | Optical signal amplification 3-terminal device | |
| CN112101540B (en) | Optical neural network chip and calculation method thereof | |
| CN115222035B (en) | Photon neural network convolution acceleration chip | |
| JP2022159140A (en) | Ultra wide band optical processor | |
| CN116739063A (en) | Neural network accelerator based on multimode interferometer and coherent detection | |
| CN116822601A (en) | Matrix operation accelerator combining wavelength division multiplexing and MZI cascade network | |
| CN110989102A (en) | Silicon-based WDM optical transmission device based on VCSEL array hybrid integration and optical fiber vertical packaging | |
| CN103370112A (en) | Laser light source output apparatus and laser output system | |
| CN112232487B (en) | Optical neural network chip and calculation method thereof | |
| JP6126521B2 (en) | Multilevel light intensity modulator | |
| CN113630207A (en) | Photon neural network chip | |
| WO2023091675A1 (en) | Optical neural network with gain from parity time optical couplers | |
| CN115508958B (en) | Photonic chip based on optical neural network | |
| US11500410B1 (en) | System and method for parallel photonic computation | |
| WO2023086516A1 (en) | Photonic ising compute engine with an optical phased array | |
| CN114217293A (en) | Light-operated multi-beam forming network chip and network | |
| CN116679504A (en) | Optical phased array with high emergent light power | |
| Al-Qadasi et al. | Scaling up silicon photonic-based accelerators: challenges and opportunities, and roadmapping with silicon photonics 2.0 |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |