CN114565091A - Optical neural network device, chip and optical implementation method for neural network calculation - Google Patents
Optical neural network device, chip and optical implementation method for neural network calculation Download PDFInfo
- Publication number
- CN114565091A CN114565091A CN202210140961.7A CN202210140961A CN114565091A CN 114565091 A CN114565091 A CN 114565091A CN 202210140961 A CN202210140961 A CN 202210140961A CN 114565091 A CN114565091 A CN 114565091A
- Authority
- CN
- China
- Prior art keywords
- optical
- optical signals
- module
- paths
- optical signal
- 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.)
- Pending
Links
- 230000003287 optical effect Effects 0.000 title claims abstract description 410
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 78
- 238000000034 method Methods 0.000 title claims abstract description 61
- 238000004364 calculation method Methods 0.000 title abstract description 11
- 239000006185 dispersion Substances 0.000 claims abstract description 75
- 238000012545 processing Methods 0.000 claims abstract description 44
- 238000010606 normalization Methods 0.000 claims description 23
- 230000001934 delay Effects 0.000 claims description 22
- 239000013307 optical fiber Substances 0.000 claims description 20
- 238000001514 detection method Methods 0.000 claims description 8
- 230000008569 process Effects 0.000 description 21
- 239000000835 fiber Substances 0.000 description 11
- 238000006243 chemical reaction Methods 0.000 description 8
- 238000010586 diagram Methods 0.000 description 7
- 238000013461 design Methods 0.000 description 5
- XUIMIQQOPSSXEZ-UHFFFAOYSA-N Silicon Chemical compound [Si] XUIMIQQOPSSXEZ-UHFFFAOYSA-N 0.000 description 4
- 230000008878 coupling Effects 0.000 description 4
- 238000010168 coupling process Methods 0.000 description 4
- 238000005859 coupling reaction Methods 0.000 description 4
- 230000006870 function Effects 0.000 description 4
- 238000000926 separation method Methods 0.000 description 4
- 229910052710 silicon Inorganic materials 0.000 description 4
- 239000010703 silicon Substances 0.000 description 4
- 239000000203 mixture Substances 0.000 description 3
- 238000001228 spectrum Methods 0.000 description 3
- 238000012360 testing method Methods 0.000 description 3
- 108010076504 Protein Sorting Signals Proteins 0.000 description 2
- 238000010521 absorption reaction Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000004422 calculation algorithm Methods 0.000 description 2
- 238000004891 communication Methods 0.000 description 2
- 238000004590 computer program Methods 0.000 description 2
- 230000014509 gene expression Effects 0.000 description 2
- 230000010354 integration Effects 0.000 description 2
- GQYHUHYESMUTHG-UHFFFAOYSA-N lithium niobate Chemical compound [Li+].[O-][Nb](=O)=O GQYHUHYESMUTHG-UHFFFAOYSA-N 0.000 description 2
- 229920000642 polymer Polymers 0.000 description 2
- 238000012549 training Methods 0.000 description 2
- 230000009471 action Effects 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 239000000969 carrier Substances 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 239000011521 glass Substances 0.000 description 1
- 238000011478 gradient descent method Methods 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 239000004065 semiconductor Substances 0.000 description 1
- 238000004611 spectroscopical analysis Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Images
Classifications
-
- 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
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/067—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using optical means
- G06N3/0675—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using optical means using electro-optical, acousto-optical or opto-electronic means
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Theoretical Computer Science (AREA)
- Evolutionary Computation (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Neurology (AREA)
- Optical Communication System (AREA)
Abstract
The embodiment of the application discloses an optical neural network device, a chip and an optical implementation method for neural network calculation, wherein the device comprises: the first modulation module is used for modulating the initial optical signal based on a first voltage to obtain a first optical signal; the dispersion module is used for separating the optical signals with N different wavelengths in the first optical signal on the time domain based on a first time delay to obtain a second optical signal; the optical splitting module is used for splitting the second optical signal into N paths of third optical signals with equal power; the second modulation module is used for modulating the intensities of the N paths of third optical signals respectively based on the N first voltage sets to obtain N paths of fourth optical signals; the inverse dispersion module is used for superposing the fourth optical signals with different wavelengths on a time domain based on a second time delay to obtain N paths of fifth optical signals; and the processing module is used for adjusting the N first voltages and the N first voltage sets based on the N fifth optical signals.
Description
Technical Field
The embodiment of the application relates to the field of photonic integration and high-performance calculation, and relates to but is not limited to an optical neural network device, a chip and an optical implementation method for neural network calculation.
Background
The neural network calculation is one of the main functions of the artificial intelligence chip, and needs to complete a series of complex matrix multiplication operations quickly. The neural network calculations may be implemented using an electronic chip. However, conventional electrical chips are limited in computational efficiency due to the influence of process nodes and material characteristics. The photonic integrated device, particularly the silicon-based photonic integrated device, is rapidly developed in recent years, the silicon-based photonic integrated device has the advantages of high speed, strong anti-interference capability, high integration level, semiconductor process compatibility and the like, and a high-efficiency photonic computing chip can be formed by utilizing the functions of modulation, filtering, beam splitting and the like of the photonic integrated device, so that the efficiency of neural network computing is expected to be greatly improved. That is to say, the efficiency of the neural network calculation is greatly improved when the neural network calculation is implemented by using an optical chip.
The optical chip mainly relies on an optical fiber delay line, a wave combiner and a wave splitter to realize the separation and the addition of signals with different wavelengths in a time domain. However, the optical fiber delay line, the multiplexer and the demultiplexer are disposed, which results in high design complexity, high power consumption and high cost of the optical chip.
Disclosure of Invention
In view of this, the present application provides an optical neural network device, a chip and an optical implementation method of neural network computation.
The technical scheme of the embodiment of the application is realized as follows:
in a first aspect, an embodiment of the present application provides an optical neural network device, including:
the first modulation module is used for modulating an initial optical signal based on a first voltage to obtain a first optical signal, wherein the first optical signal comprises parallel optical signals with N different wavelengths, the intensity of the first optical signal is obtained by adding different intensities corresponding to the optical signals with the different wavelengths, and N is an integer greater than 1;
the dispersion module is configured to separate, in a time domain, N types of optical signals with different wavelengths in the first optical signal based on setting different first time delays corresponding to the optical signals with different wavelengths to obtain a second optical signal, where the second optical signal is a serial optical signal including the N types of optical signals with different wavelengths;
the optical splitting module is used for splitting the second optical signal into N paths of third optical signals with equal power;
the second modulation module is used for modulating the intensities of the N paths of third optical signals respectively based on the input N first voltage sets to obtain N paths of fourth optical signals;
the inverse dispersion module is configured to, based on different second time delays corresponding to optical signals with different wavelengths, superimpose the fourth optical signals with different wavelengths on a time domain to obtain N paths of fifth optical signals, where the first time delay is different from the second time delay;
and the processing module is used for adjusting the N first voltages and the N first voltage sets based on the N fifth optical signals.
In a second aspect, an embodiment of the present application provides an optical neural network chip, including the optical neural network device of the claims.
In a third aspect, an embodiment of the present application provides an optical implementation method for neural network computation, where the method includes:
a first modulation module of the optical neural network device modulates an initial optical signal based on a first voltage to obtain a first optical signal, wherein the first optical signal comprises parallel optical signals with N different wavelengths, the intensity of the first optical signal is obtained by adding different intensities corresponding to the optical signals with the different wavelengths, and N is an integer greater than 1;
the dispersion module of the optical neural network device separates N optical signals with different wavelengths in the first optical signal on a time domain based on setting different first time delays corresponding to the optical signals with different wavelengths to obtain a second optical signal, wherein the second optical signal is a serial optical signal comprising N different wavelengths;
the optical splitting module of the optical neural network device splits the second optical signal into N paths of third optical signals with equal power;
the second modulation module of the optical neural network device modulates the intensity of the N paths of third optical signals respectively based on the input N first voltage sets to obtain N paths of fourth optical signals;
the inverse dispersion module of the optical neural network device superposes the fourth optical signals with different wavelengths on a time domain based on setting different second time delays corresponding to the optical signals with different wavelengths to obtain N paths of fifth optical signals, wherein the first time delay is different from the second time delay;
the processing module of the optical neural network device adjusts the N first voltages and the N first voltage sets based on the N fifth optical signals.
In this embodiment of the present application, the dispersion module is configured to separate N paths of optical signals with different wavelengths in the first optical signal in a time domain based on setting that optical signals with different wavelengths correspond to different first time delays, so as to obtain a second optical signal, where the second optical signal is a serial optical signal including N types of different wavelengths; and the inverse dispersion module is configured to, based on that the optical signals with different wavelengths correspond to different second time delays, superimpose the fourth optical signals with different wavelengths on a time domain to obtain N paths of fifth optical signals, where the first time delay is different from the second time delay. Therefore, the signals with different wavelengths can be separated in a time domain based on the dispersion module, and the signals with different wavelengths can be added in the time domain based on the inverse dispersion module, so that the design complexity, the power consumption and the cost of the optical neural network device are effectively reduced.
Drawings
Fig. 1 is a structural composition diagram of an optical neural network device according to an embodiment of the present disclosure;
fig. 2A is a structural composition diagram of a first modulation module according to an embodiment of the present disclosure;
FIG. 2B is a structural assembly diagram of a process module according to an embodiment of the present disclosure;
fig. 3 is a structural composition diagram of an optical neural network device according to an embodiment of the present disclosure;
fig. 4 is a schematic implementation flow diagram of an optical implementation method for neural network computation according to an embodiment of the present application.
Detailed Description
In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, specific technical solutions of the embodiments of the present application will be described in further detail below with reference to the drawings in the embodiments of the present application. The following examples are intended to illustrate the present application but are not intended to limit the scope of the present application.
In the following description, reference is made to "some embodiments" which describe a subset of all possible embodiments, but it is understood that "some embodiments" may be the same subset or different subsets of all possible embodiments, and may be combined with each other without conflict.
In the following description, references to the terms "first \ second \ third" are only to distinguish similar objects and do not denote a particular order, but rather the terms "first \ second \ third" are used to interchange specific orders or sequences, where appropriate, so as to enable the embodiments of the application described herein to be practiced in other than the order shown or described herein.
Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of the present application only and is not intended to be limiting of the application.
Before further detailed description of the embodiments of the present application, terms and expressions referred to in the embodiments of the present application will be described, and the terms and expressions referred to in the embodiments of the present application will be used for the following explanation.
Wavelength Division Multiplexing (WDM): a technology of combining two or more optical carrier signals (carrying various information) with different wavelengths together at a transmitting end through a Multiplexer (also called a Multiplexer, MUX), and coupling the signals to the same optical fiber of an optical line for transmission; at the receiving end, the optical carriers of various wavelengths are separated by a Demultiplexer (also called a Demultiplexer or a Demultiplexer (DMUX)), and then further processed by an optical receiver to recover the original signal. This technique of simultaneously transmitting two or more optical signals of different wavelengths in the same optical fiber is called wavelength division multiplexing.
A Photodetector (PD) for converting the optical signal into an electrical signal.
In various embodiments of the present application, the adjustment of the input value of each neural network layer in the neural network is implemented by using the adjustment of the N first voltages, and the adjustment of the weight value in the neural network is implemented by using the adjustment of the N first voltage sets. Because the optical signals with different wavelengths are separated in the time domain by using the dispersion module and the optical signals with different wavelengths are superposed in the time domain by using the inverse dispersion module in the embodiment of the application, the optical neural network device provided by the embodiment of the application effectively reduces required optical fiber delay lines, wave combiners and wave splitters, and further can reduce the design complexity, power consumption and cost of the optical chip when the optical neural network device provided by the embodiment of the application is used for forming the optical chip.
Fig. 1 is a diagram illustrating a structural configuration of an optical neural network device according to an embodiment of the present application, where the optical neural network device 100 includes: a first modulation module 110, a dispersion module 120, a light splitting module 130, a second modulation module 140, an inverse dispersion module 150 and a processing module 160; wherein,
a first modulation module 110, configured to modulate an initial optical signal based on a first voltage to obtain a first optical signal, where the first optical signal includes parallel optical signals with N different wavelengths, intensities of the first optical signal are obtained by adding different intensities corresponding to the optical signals with different wavelengths, and N is an integer greater than 1;
the dispersion module 120 is configured to separate, in a time domain, N optical signals with different wavelengths in the first optical signal based on setting that optical signals with different wavelengths correspond to different first time delays, so as to obtain a second optical signal, where the second optical signal is a serial optical signal including the N optical signals with different wavelengths;
in some embodiments, the dispersive module comprises a dispersive optical fiber and/or a dispersive grating.
Here, the dispersion is a phenomenon in which the complex color light is decomposed into monochromatic light to form a spectrum. Dispersion can be achieved using instruments such as prisms or gratings that act as a dispersive system. For example, after entering the prism, the polychromatic light is respectively dispersed and forms a spectrum when leaving the prism because the prism has different refractive indexes for light of various frequencies and the propagation directions of the various colored light are deflected to different degrees.
Optical fibers are short for optical fibers, and are fibers made of glass or plastic that can be used as a light conducting means. The generation of the optical fiber dispersion is based on two factors, one is that the optical signal entering the optical fiber is not monochromatic light (the light emitted by the light source is not monochromatic or the modulation signal has a certain bandwidth); the second is the dispersion of optical signals by optical fibers. In the embodiment of the present application, the initial light source is a light source including different wavelengths, and the light sources of different wavelengths may be separated in a time domain by using fiber dispersion.
In an implementation process, parameters of the dispersion fiber may be determined according to actual needs, so that the dispersion fiber is used to implement the separation of the optical signals with N different wavelengths in the first optical signal in the time domain.
A grating is an optical device consisting of a large number of parallel slits of equal width and equal spacing, also called a diffraction grating. An optical element that can disperse (resolve into a spectrum) light using the principle of multi-slit diffraction.
In an implementation process, parameters of the dispersion grating may be determined according to actual needs, so as to implement, by using a dispersion fiber, separation of optical signals of N different wavelengths in the first optical signal in a time domain.
For example, a dispersive fiber and/or a dispersive grating may be used such that the first time delays produced by optical signals of different wavelengths are 0, Δ t, … …, (N-1) Δ t, respectively, to generate a serial pulsed optical signal. The implementation method comprises the following steps: the delay of the first optical signal in the first optical signals is set as an initial value (e.g. 0), and on the basis of the initial value, the first delays of other first optical signals in the N first optical signals are sequentially increased according to a preset delay interval (Δ t) (the delay to the nth path of the first optical signal is (N-1) Δ t), so that the delay interval of each optical signal in the N first optical signals is Δ t.
An optical splitting module 130, configured to split the second optical signal into N paths of third optical signals with equal power;
in practice, the spectroscopy module 130 is a passive device. The light splitting module may be composed of entrance and exit slits, a mirror, a multimode interferometer, and a dispersion element.
The second modulation module 140 is configured to modulate the intensities of the N paths of third optical signals respectively based on the input N first voltage sets, so as to obtain N paths of fourth optical signals;
in the implementation process, the input end of the second modulator module 140 is connected to the output end of the optical splitter module, and since the optical splitter splits the second optical signal into N paths of third optical signals with equal power, each third optical signal needs to be input to a corresponding second modulation module, and all the N second modulation modules exist; the output of each second modulator module 140 is connected to a corresponding inverse dispersion module 150. The second modulation module 140 is configured to modulate intensities of the N paths of third optical signals output by the optical splitting module 130 according to the input N first voltage sets, so as to obtain N paths of fourth optical signals. Here, the N fourth optical signals are serial pulse optical signals subjected to intensity modulation.
In an implementation, the second modulator module 140 may include an array of N modulators, where the N modulators are capable of modulating the intensities of the N third optical signals respectively; the second modulator module 140 may also include an integrated device capable of directly modulating the intensity of the N third optical signals.
In an implementation process, each of the N modulators may specifically include: a lithium niobate electro-optic modulator, a polymer electro-optic modulator, a silicon-based electro-optic modulator, an electro-absorption modulator, an integrated electro-optic modulator, or a spatial light modulator.
Here, the second modulator module 140 includes N voltage input ports, which are connected to the processing module 160, and N dynamically varying first voltage sets are provided by the processing module 160 to be applied to the N modulators of the second modulator module 140, respectively. It can be understood that, by adjusting each voltage value in the N first voltage sets, the weight value in the neural network can be adjusted.
It can be understood that since the N third optical signals can be regarded as a sequence of the N optical signals, a voltage sequence (i.e., the first voltage set) needs to be applied to each modulator included in the second modulation module. And under the action of the voltage sequence, the modulation of a plurality of weight values is realized, so that when the weight number is fixed, the optical neural network device provided by the embodiment of the application can reduce the number of required optical modulators. Here, the application interval of each voltage in the voltage sequence is related to the delay interval between each of the N first optical signals.
The inverse dispersion module 150 is configured to, based on that optical signals with different wavelengths are set to correspond to different second time delays, superimpose the fourth optical signals with different wavelengths on a time domain to obtain N paths of fifth optical signals, where the first time delay is different from the second time delay;
in some embodiments, the inverse dispersion module comprises an inverse dispersion fiber and/or an inverse dispersion grating.
Here, the inverse dispersion fiber and the dispersion fiber are fibers of different parameters; inverse dispersion gratings and dispersion gratings are also gratings of different parameters.
In the implementation process, parameters of the inverse dispersion optical fiber can be determined according to actual needs, so that the inverse dispersion optical fiber is utilized to realize that each path of fourth optical signal is superposed on a time domain to obtain N paths of fifth optical signals.
In the implementation process, parameters of the inverse dispersion grating can be determined according to actual needs, so that each path of fourth optical signal is superimposed on a time domain by using the inverse dispersion optical fiber, and N paths of fifth optical signals are obtained.
For example, the parameters of the inverse dispersion fiber and/or the inverse dispersion grating may be adjusted such that the second time delays generated by the optical signals with different wavelengths in each of the fourth optical signals are (N-1) Δ t, (N-2) Δ t, … …, and 0, respectively. And by combining the time delay difference generated by the dispersion module, the optical signals with different wavelengths in each path of fourth optical signal can be superposed on the time domain to obtain N paths of fifth optical signals. Here, since the delay caused by the different wavelengths is opposite to the dispersion module, the serial signals with different wavelengths can be superimposed on the time domain.
A processing module 160, configured to adjust the N first voltages and the N first voltage sets based on the N fifth optical signals.
In this embodiment of the present application, the dispersion module is configured to separate N paths of optical signals with different wavelengths in the first optical signal in a time domain based on setting that optical signals with different wavelengths correspond to different first time delays, so as to obtain a second optical signal, where the second optical signal is a serial optical signal including N types of different wavelengths; and the inverse dispersion module is configured to, based on different second time delays corresponding to optical signals with different wavelengths, superimpose the fourth optical signals with different wavelengths on a time domain to obtain N paths of fifth optical signals, where the first time delay is different from the second time delay. Therefore, the signals with different wavelengths can be separated in a time domain based on the dispersion module, and the signals with different wavelengths can be added in the time domain based on the inverse dispersion module, so that the design complexity, the power consumption and the cost of the optical neural network device are effectively reduced.
In some embodiments, as shown in fig. 2A, the first modulation module 110 includes:
the optical generation submodule 111 is used for generating parallel initial optical signals with N different wavelengths;
in practice, the output of the light generating submodule is connected to the input of the first modulation submodule.
The light generation sub-module 111 may generate optical signals of N different wavelengths. Optical signals of different wavelengths may be used to carry information of the test set.
In practical applications, the light generating sub-module 111 may include an array of N lasers, where the N lasers are capable of generating and outputting N kinds of single-frequency continuous light with different wavelengths; the light generating sub-module 111 may also comprise an integrated device capable of directly generating and inputting N optical signals of different wavelengths.
In some embodiments, the light generation sub-module 111 comprises: the optical frequency comb is used for generating an initial optical signal with N wavelengths; and the wave splitter is used for splitting the initial optical signal with the N wavelengths into N parallel initial optical signals with different wavelengths.
Here, an optical frequency comb, referred to as an optical frequency comb, is composed of a series of discrete, equally spaced, regular optical pulse trains, and is capable of simultaneously providing several to several tens of different frequency components within one frequency band. Its advantages are high number of comb lines, wide frequency range, and fixed frequency interval.
In the implementation process, the output end of the optical frequency comb is connected with the input end of the wave splitter, one path of initial optical signals with N wavelengths can be generated based on the optical frequency comb, and then the N paths of parallel initial optical signals with different wavelengths are effectively obtained through the wave splitter.
The first modulation submodule 112 is configured to modulate the intensity of each path of the parallel initial optical signals respectively based on the input N first voltages, so as to obtain N paths of initial optical signals with different intensities;
the input of the first modulation submodule 112 is connected to the output of the light generation submodule, and the output of the first modulation submodule is connected to the combiner submodule. The first modulation submodule 112 may be configured to modulate, according to the input N first voltages, intensities of the N paths of optical signals with different wavelengths output by the optical generation submodule, respectively, so as to obtain N paths of initial optical signals with different intensities. Here, the N initial optical signals are parallel optical signals.
In some embodiments, the first modulation submodule 112 may include an array of N modulators capable of modulating the intensity of N optical signals of different wavelengths, respectively; the first modulation submodule may also include an integrated device capable of directly modulating the intensity of the N optical signals of different wavelengths. In an implementation process, each of the N modulators may specifically include: a lithium niobate electro-optic modulator, a polymer electro-optic modulator, a silicon-based electro-optic modulator, an electro-absorption modulator, an integrated electro-optic modulator, or a spatial light modulator.
In implementation, the first modulation submodule 112 includes N voltage input ports, the N voltage input ports are connected to the processing module, and the processing module provides N dynamically-changing first voltages, which are respectively applied to the N modulators of the first modulation submodule. It is understood that the adjustment of the input values of the neural network layers in the neural network may be achieved by the adjustment of the values of the N first voltages.
And a combining submodule 113, configured to combine the N initial optical signals with different intensities into the first optical signal.
Here, the input of the combining submodule 113 is connected to the output of the first modulation submodule, and the output of the combining submodule is connected to the input of the dispersion module. The combining sub-module 113 may be configured to combine N initial optical signals with different intensities into one parallel first optical signal including N different wavelengths.
In an embodiment of the present application, the first modulation module includes: the optical generation submodule is used for generating parallel initial optical signals with N different wavelengths; the first modulation submodule is used for respectively modulating the intensity of each path of parallel initial optical signal based on the input N first voltages to obtain N paths of initial optical signals with different intensities; and the combining submodule is used for combining the N paths of initial optical signals with different intensities into the first optical signal. Based on the first modulation submodule, the debugging of the intensity of each path of the parallel initial optical signal can be realized; the first optical signal can be effectively obtained by utilizing the combining submodule.
In an embodiment of the present application, the light generation sub-module includes: the optical frequency comb is used for generating an initial optical signal with N wavelengths; and the wave splitter is used for splitting the initial optical signal with the N wavelengths into N parallel initial optical signals with different wavelengths. Thus, N paths of parallel initial optical signals with different wavelengths can be effectively obtained.
In some embodiments, as shown in fig. 2B, the processing module 160 includes:
the nonlinear sub-module 161 is configured to perform nonlinear processing on each of the N fifth optical signals, so as to obtain N fifth optical signals subjected to nonlinear processing;
the photoelectric detection submodule 162 is configured to convert the N paths of fifth optical signals that have been subjected to the nonlinear processing into N paths of electrical signals;
in the implementation process, the input end of the photodetection sub-module 162 is connected to the output end of the non-linear sub-module 161, and is configured to convert the N paths of non-linearly processed fifth optical signals into N paths of electrical signals; and the output terminal of the photodetection sub-module 162 is connected to the input terminal of the normalization sub-module 163. The photo-detection sub-module 162 may include N photo-detectors for measuring the intensity of each of the N fifth optical signals.
In some embodiments, the photo-detection module 162 is capable of detecting the optical power incident on its face and converting this change in optical power into a corresponding electrical current, i.e., the photo-detection module 162 sends N electrical signals carrying the optical intensity of the optical signal to the normalization submodule 163. In practical application, the photoelectric detection module may comprise a PIN tube or an APD tube.
The normalization submodule 163 is configured to perform normalization processing on the N paths of electrical signals, and use N paths of results of the normalization processing as values of the N first voltages; generating N first voltages and applying the N first voltages to the modulation submodule;
in implementation, an input of the normalization sub-module 163 is connected to an output of the photo-detection module 162, and an output of the normalization sub-module 163 is connected to an input of the processing sub-module 164. A normalization module 163, configured to perform normalization processing on the N paths of electrical signals, and take N paths of results of the normalization processing as values of the N first voltages; n first voltages are generated and applied to the first modulation submodule 112. It will be appreciated that the voltage value corresponding to the light intensity measured by the photodetection module is different from the relationship between voltage and light intensity in the first modulation submodule 112, and therefore requires the intervention of a normalization module. Here, the value of the first voltage is connected to the first modulation submodule 112 in a feedback manner, which can be understood by analogy as the output port of the previous roll-up layer and the input port of the next roll-up layer in the neural network.
A processing submodule 164, configured to determine a value of each voltage in the N first voltage sets based on N results of the normalization processing; n first voltage sets are generated and applied across the modulation module.
The input of the processing submodule 164 is connected to the output of the normalisation module 163 and the output of the processing submodule 164 is connected to the second modulation module 150. The processing submodule 164 is configured to determine a value of each voltage in the N first voltage sets based on the N normalization processing results; n first voltage sets are generated and applied to the second modulation module 150. Here, the values of the voltages in the set of first voltages are fed back to the second modulation module 150, and the adjustment of the values of the voltages in the set of first voltages in a loop can be understood as an adjustment of the weight values of the data to be subjected to convolution calculation in a loop in the neural network.
In some embodiments, the processor sub-module 164 is configured to compare the N-way results of the normalization process with a first data set; the first data set represents a training set corresponding to information carried by N paths of optical signals with different wavelengths; and adjusting the value of each voltage in the N first voltage sets according to the comparison result and by combining a gradient descent algorithm. Here, the information carried by the N optical signals with different wavelengths may correspond to a test set of the neural network, and a training set corresponding to the test set is stored in the storage unit of the processing sub-module 164. In practical application, the N-way results of the normalization processing are compared with the first data set, and the values of the voltages in the N first voltage sets are continuously adjusted by combining algorithms such as gradient descent and the like until an optimal solution of the weight value is obtained. Here, the idea of the gradient descent method is to solve the minimum value in the direction of the gradient descent.
In implementation, the processing sub-module 164 includes an Application Specific Integrated Circuit (ASIC).
In the embodiment of the application, the adjustment of the input value of each neural network layer in the neural network is realized by using the adjustment of the N first voltages, and the adjustment of the weight value in the neural network is realized by using the adjustment of the N first voltage sets.
The embodiments of the present application will be described in more detail with reference to specific application scenarios. In this embodiment, as shown in fig. 3, the optical neural network device 100 includes a first modulation module 110, a dispersion module 120, a light splitting module 130, a second modulation module 140, an inverse dispersion module 150, and a processing module 160, wherein,
the first modulation module 110 includes a light generation sub-module 111 (optical frequency comb, wavelength splitter), a first modulation sub-module 112, i.e., n modulators (modulator 1, modulator 2, … …, modulator n) and a combiner sub-module 113 (combiner);
the light splitting module 130 is a light splitter;
the second modulation module 140, i.e., n modulators (modulator W1, modulator W2 … …, modulator Wn);
the processing module 160 includes: a non-linear submodule 161, i.e. n non-linear elements (non-linear element 1, non-linear element 2, … …, non-linear element n); a photodetector submodule 162, i.e. n photodetectors (photodetector 1, photodetector 2, … …, photodetector n); n normalization submodules 163, i.e., n conversion chips (conversion chip 1, conversion chip 2, … …, conversion chip n); the processing submodule 162 is an application specific processor chip ASIC.
The separation of the different inputs in the time domain can be achieved by means of the dispersive module (dispersive element) 120;
the addition of the different inputs in the time domain can be achieved with the inverse dispersion module (inverse dispersion element) 160.
The optical neural network device is implemented according to the following specific principle:
using optical frequency to comb onFrequency domain generating n-wavelength multi-wavelength light source with respective wavelengths of lambda1,λ2,……,λnThe multi-wavelength light source is connected to a wavelength division multiplexer (DMUX0) for dividing the input signal into λ1,λ2,……,λnN parallel optical signals. Each optical signal is connected with a first modulation submodule 112, that is, n parallel optical signals are correspondingly connected with n modulators (modulator 1, modulator 2, … …, modulator n), and the n parallel optical signals are modulated to generate parallel pulse optical signals X with different intensities1,X2,……,Xn。
Finally, the beams are combined by a combiner (MUX0)113 to generate the beam with intensity (X)1+X2+ … … + Xn).
The pulse light signal sequence passes through the dispersion element 120, and different delays are generated due to different wavelengths, so that lambda is converted1,λ2,……,λnThe multi-wavelength pulse signals are separated in time domain to generate serial pulse light signal sequence (X)1,X2,……,Xn). Then, the serial pulse optical signal is divided into n channels with equal power by the optical splitter 130, each channel is connected with one second modulation module 140, that is, n channels are correspondingly connected with n modulators (modulator W1, modulator W2 … …, modulator Wn), and each modulator pair sequentially pairs the serial pulse optical signal (X)1,X2,……,Xn) The modulation weights are (W11, W12, … …, W1n), (W21, W22, … …, W2n), … …, (Wn1, Wn2, … …, Wnn), and so on.
Generated by a W1 modulator (W11. X)1,W12·X2,……,W1n·Xn) The serial pulse optical signal of/n is generated by a Wn modulator (Wn 1. X)1,Wn2·X2,……,Wnn·Xn) Serial pulse light signal of/n, and so on.
Each of the n modulators (modulator W1, modulator W2 … …, modulator Wn) is connected to an inverse dispersion element 150, whose delay due to the wavelength difference is opposite to that of the dispersion element, so that the serial signals of different wavelengths are added again, and the output intensities are respectively (A), (B), (C), (D), and D), a) and D) in a) an optical) A) and (D)W11·X1+W12·X2+……W1n·Xn)/n,……,(Wn1·X1+Wn2·X2+……Wnn·Xn) The pulse optical signals of/n generate n paths of parallel optical signals.
Each path of signal output by the inverse dispersion single element is connected with a nonlinear submodule 161, namely n paths of parallel optical signals are correspondingly connected with n nonlinear units (nonlinear unit 1, nonlinear unit 2, … …, nonlinear unit n) to generate F ((W11. X. Y. X. Y. X. Y. X. Y. X. Y. X. Y. X. Y. X. Y. X. Y. X. Y. X. Y. X. Y. X. Y. X. Y. X. Y. X. Y. X. Y. X. Y. X. Y. X. Y. X. Y1+W12·X2+……W1n·Xn)/n),……,F((Wn1·X1+Wn2·X2+……Wnn·Xn) N) of parallel optical signals, where F is a nonlinear transformation function required by the neural network. The n parallel optical signals are converted into parallel electrical signals by the photoelectric detection submodule 162, that is, the n photodetectors (photodetector 1, photodetector 2, … …, photodetector n), and the parallel electrical signals enter the n normalization submodules 163, that is, the n conversion chips (conversion chip 1, conversion chip 2, … …, conversion chip n), after normalization processing, the control voltages are respectively sent to each of the n modulators (modulator W1, modulator W2 … …, modulator Wn), so as to generate a new X1,X2,……,XnAnd parallel optical signals. Meanwhile, the processing submodule 162, i.e., an application specific processor chip (ASIC), adjusts voltages applied to the n modulators (modulator W1, modulator W2 … …, modulator Wn) according to the comparison result of the input parallel electrical signal and the learning data, and updates the modulation weights (W11, W12, … …, W1n), (W21, W22, … …, W2n), … …, (Wn1, Wn2, … …, Wnn), respectively.
And circulating the above processes until the comparison result of the parallel electric signals input in the processor chip and the learning data reaches the application requirement.
The embodiment of the present application further provides an optical neural network chip, where the optical neural network chip includes: the optical neural network device 100 provided by the embodiment of the present application.
The optical neural network structure used in the optical neural network chip provided in the above embodiments and the optical neural network device embodiments belong to the same concept, and specific implementation processes thereof are described in the device embodiments in detail and are not described herein again.
The embodiment of the application provides an optical implementation method for neural network calculation. Fig. 4 is a schematic implementation flow diagram of an optical implementation method of neural network computation according to an embodiment of the present application. As shown in fig. 4, the method comprises the steps of:
step S410, a first modulation module of the optical neural network device modulates an initial optical signal based on a first voltage to obtain a first optical signal, where the first optical signal includes parallel optical signals with N different wavelengths, the intensities of the first optical signal are obtained by adding different intensities corresponding to the optical signals with the different wavelengths, and N is an integer greater than 1;
step S420, based on setting different first time delays corresponding to optical signals with different wavelengths, a dispersion module of the optical neural network device separates optical signals with N different wavelengths in the first optical signal in a time domain to obtain a second optical signal, where the second optical signal is a serial optical signal including the N different wavelengths;
step S430, the optical splitting module of the optical neural network device splits the second optical signal into N paths of third optical signals with equal power;
step S440, a second modulation module of the optical neural network device modulates the intensities of the N paths of third optical signals respectively based on the input N first voltage sets to obtain N paths of fourth optical signals;
step S450, based on setting different second time delays corresponding to optical signals with different wavelengths, an inverse dispersion module of the optical neural network device superimposes the fourth optical signals with different wavelengths on a time domain to obtain N paths of fifth optical signals, where the first time delay is different from the second time delay;
step S460, the processing module of the optical neural network device adjusts N first voltages and N first voltage sets based on the N fifth optical signals.
In some embodiments, the dispersion module comprises a dispersive optical fiber and/or a dispersive grating; the inverse dispersion module comprises an inverse dispersion optical fiber and an inverse dispersion grating.
In the embodiment of the present application, based on setting different first time delays corresponding to optical signals with different wavelengths, a dispersion module separates N paths of optical signals with different wavelengths in a first optical signal in a time domain to obtain a second optical signal, where the second optical signal is a serial optical signal including N different wavelengths; and the inverse dispersion module superposes each path of fourth optical signal on a time domain based on setting different second time delays corresponding to the optical signals with different wavelengths to obtain N paths of fifth optical signals, wherein the first time delay is different from the second time delay. Therefore, the signals with different wavelengths can be separated in a time domain based on the dispersion module, and the signals with different wavelengths can be added in the time domain based on the inverse dispersion module, so that the design complexity, the power consumption and the cost of the optical neural network device are effectively reduced.
In some embodiments, the above step S410 "the first modulation module of the optical neural network device modulates the initial optical signal based on the first voltage to obtain the first optical signal", includes the following steps:
step 411, the light generation submodule of the first modulation module generates N paths of parallel initial light signals with different wavelengths;
step 412, the first modulation submodule of the first modulation module modulates the intensity of each path of the parallel initial optical signals respectively based on the input N first voltages to obtain N paths of initial optical signals with different intensities;
step 413, the combining sub-module of the first modulation module combines the N initial optical signals with different intensities into the first optical signal.
In an embodiment of the present application, the first modulation module includes: the light generation submodule generates parallel initial light signals with N different wavelengths; the first modulation submodule respectively modulates the intensity of each path of parallel initial optical signals based on the input N first voltages to obtain N paths of initial optical signals with different intensities; and the combining submodule combines the N paths of initial optical signals with different intensities into the first optical signal. Based on the first modulation submodule, the debugging of the intensity of each path of the parallel initial optical signal can be realized; the first optical signal can be effectively obtained by utilizing the combining submodule.
In some embodiments, the above step 411 "the light generation sub-module generates N parallel initial optical signals of different wavelengths" includes the steps of:
step A, an optical frequency comb of the optical generation sub-module generates an initial optical signal with N wavelengths;
and step B, the wave separator of the optical generation submodule divides the initial optical signal with N wavelengths into N parallel initial optical signals with different wavelengths.
In the embodiment of the application, the optical frequency comb generates an initial optical signal with N wavelengths; and the wave splitter divides the initial optical signal with N wavelengths into N parallel initial optical signals with different wavelengths. Thus, N paths of parallel initial optical signals with different wavelengths can be effectively obtained.
The above description of the apparatus embodiments, similar to the above description of the method embodiments, has similar beneficial effects as the method embodiments. For technical details not disclosed in the embodiments of the apparatus of the present application, reference is made to the description of the embodiments of the method of the present application for understanding.
It should be noted that, in the embodiment of the present application, if the method is implemented in the form of a software functional module and sold or used as a standalone product, the method may also be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the embodiments of the present application may be embodied in the form of a software product, which is stored in a storage medium and includes instructions for causing an electronic device (which may be a mobile phone, a tablet computer, a notebook computer, a desktop computer, etc.) to execute all or part of the methods described in the embodiments of the present application. And the aforementioned storage medium includes: various media capable of storing program codes, such as a usb disk, a removable hard disk, a Read Only Memory (ROM), a magnetic disk, or an optical disk. Thus, embodiments of the present application are not limited to any specific combination of hardware and software.
Correspondingly, the present application provides a storage medium, on which a computer program is stored, and the computer program is executed by a processor to implement the steps in the optical implementation method of neural network computation provided in the above embodiments.
Here, it should be noted that: the above description of the storage medium embodiment is similar to the description of the method embodiment described above, with similar beneficial effects as the method embodiment. For technical details not disclosed in the embodiments of the storage medium of the present application, reference is made to the description of the embodiments of the method of the present application for understanding.
It should be appreciated that reference throughout this specification to "one embodiment" or "an embodiment" means that a particular feature, structure or characteristic described in connection with the embodiment is included in at least one embodiment of the present application. Thus, the appearances of the phrases "in one embodiment" or "in an embodiment" in various places throughout this specification are not necessarily all referring to the same embodiment. Furthermore, the particular features, structures, or characteristics may be combined in any suitable manner in one or more embodiments. It should be understood that, in the various embodiments of the present application, the sequence numbers of the above-mentioned processes do not mean the execution sequence, and the execution sequence of each process should be determined by its function and inherent logic, and should not constitute any limitation to the implementation process of the embodiments of the present application. The above-mentioned serial numbers of the embodiments of the present application are merely for description, and do not represent the advantages and disadvantages of the embodiments.
It should be noted that, in this document, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of other like elements in a process, method, article, or apparatus that comprises the element.
In the several embodiments provided in the present application, it should be understood that the disclosed apparatus and method may be implemented in other ways. The above-described device embodiments are merely illustrative, for example, the division of the unit is only a logical functional division, and there may be other division ways in actual implementation, such as: multiple units or components may be combined, or may be integrated into another system, or some features may be omitted, or not implemented. In addition, the coupling, direct coupling or communication connection between the components shown or discussed may be through some interfaces, and the indirect coupling or communication connection between the devices or units may be electrical, mechanical or other forms.
The units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units; can be located in one place or distributed on a plurality of network units; some or all of the units can be selected according to actual needs to achieve the purpose of the solution of the embodiment.
In addition, all functional units in the embodiments of the present application may be integrated into one processing unit, or each unit may be separately regarded as one unit, or two or more units may be integrated into one unit; the integrated unit can be realized in a form of hardware, or in a form of hardware plus a software functional unit.
Those of ordinary skill in the art will understand that: all or part of the steps for realizing the method embodiments can be completed by hardware related to program instructions, the program can be stored in a computer readable storage medium, and the program executes the steps comprising the method embodiments when executed; and the aforementioned storage medium includes: various media that can store program codes, such as a removable Memory device, a Read Only Memory (ROM), a magnetic disk, or an optical disk.
Alternatively, the integrated units described above in the present application may be stored in a computer-readable storage medium if they are implemented in the form of software functional modules and sold or used as independent products. Based on such understanding, the technical solutions of the embodiments of the present application may be embodied in the form of a software product, which is stored in a storage medium and includes instructions for causing an electronic device (which may be a mobile phone, a tablet computer, a notebook computer, a desktop computer, etc.) to execute all or part of the methods described in the embodiments of the present application. And the aforementioned storage medium includes: various media that can store program code, such as removable storage devices, ROMs, magnetic or optical disks, etc.
The methods disclosed in the several method embodiments provided in the present application may be combined arbitrarily without conflict to obtain new method embodiments.
Features disclosed in several of the product embodiments provided in the present application may be combined in any combination to yield new product embodiments without conflict.
The features disclosed in the several method or device embodiments provided in the present application may be combined in any combination to arrive at a new method or device embodiment without conflict.
The above description is only for the embodiments of the present application, but the scope of the present application is not limited thereto, and any person skilled in the art can easily conceive of changes or substitutions within the technical scope of the present application, and shall be covered by the scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.
Claims (10)
1. An optical neural network device, the device comprising:
the first modulation module is used for modulating an initial optical signal based on a first voltage to obtain a first optical signal, wherein the first optical signal comprises parallel optical signals with N different wavelengths, the intensity of the first optical signal is obtained by adding different intensities corresponding to the optical signals with the different wavelengths, and N is an integer greater than 1;
the dispersion module is configured to separate, in a time domain, N types of optical signals with different wavelengths in the first optical signal based on setting different first time delays corresponding to the optical signals with different wavelengths to obtain a second optical signal, where the second optical signal is a serial optical signal including the N types of optical signals with different wavelengths;
the optical splitting module is used for splitting the second optical signal into N paths of third optical signals with equal power;
the second modulation module is used for modulating the intensities of the N paths of third optical signals respectively based on the input N first voltage sets to obtain N paths of fourth optical signals;
the inverse dispersion module is configured to, based on different second time delays corresponding to optical signals with different wavelengths, superimpose the fourth optical signals with different wavelengths on a time domain to obtain N paths of fifth optical signals, where the first time delay is different from the second time delay;
and the processing module is used for adjusting the N first voltages and the N first voltage sets based on the N paths of fifth optical signals.
2. The apparatus of claim 1, wherein the first modulation module comprises:
the optical generation submodule is used for generating N paths of parallel initial optical signals with different wavelengths;
the first modulation submodule is used for respectively modulating the intensity of each path of parallel initial optical signals based on the input N first voltages to obtain N paths of initial optical signals with different intensities;
and the combining submodule is used for combining the N paths of initial optical signals with different intensities into the first optical signal.
3. The apparatus of claim 2, wherein the light generation sub-module comprises:
the optical frequency comb is used for generating an initial optical signal with N wavelengths;
and the wave splitter is used for splitting the initial optical signal with the N wavelengths into N parallel initial optical signals with different wavelengths.
4. The apparatus of claim 1, wherein the dispersive module comprises a dispersive optical fiber and/or a dispersive grating; the inverse dispersion module comprises an inverse dispersion optical fiber and an inverse dispersion grating.
5. The apparatus of claim 2, wherein the processing module comprises:
the nonlinear sub-module is used for respectively carrying out nonlinear processing on each optical signal in the N paths of fifth optical signals to obtain N paths of fifth optical signals which are subjected to nonlinear processing;
the photoelectric detection sub-module is used for converting the N fifth optical signals which are subjected to the nonlinear processing into N electric signals;
the normalization submodule is used for performing normalization processing on the N paths of electric signals and taking N paths of results of the normalization processing as values of the N first voltages; generating N first voltages and applying the N first voltages to the modulation submodule;
the processing submodule is used for determining the value of each voltage in the N first voltage sets based on N paths of results of normalization processing; n first voltage sets are generated and applied across the modulation module.
6. An optical neural network chip comprising the optical neural network device of any one of claims 1 to 5.
7. An optical implementation method of neural network computing, the method comprising:
a first modulation module of the optical neural network device modulates an initial optical signal based on a first voltage to obtain a first optical signal, wherein the first optical signal comprises parallel optical signals with N different wavelengths, the intensity of the first optical signal is obtained by adding different intensities corresponding to the optical signals with the different wavelengths, and N is an integer greater than 1;
the dispersion module of the optical neural network device separates N optical signals with different wavelengths in the first optical signal on a time domain based on setting different first time delays corresponding to the optical signals with different wavelengths to obtain a second optical signal, wherein the second optical signal is a serial optical signal comprising N different wavelengths;
the optical splitting module of the optical neural network device splits the second optical signal into N paths of third optical signals with equal power;
the second modulation module of the optical neural network device modulates the intensity of the N paths of third optical signals respectively based on the input N first voltage sets to obtain N paths of fourth optical signals;
the inverse dispersion module of the optical neural network device superposes the fourth optical signals with different wavelengths on a time domain based on setting different second time delays corresponding to the optical signals with different wavelengths to obtain N paths of fifth optical signals, wherein the first time delay is different from the second time delay;
the processing module of the optical neural network device adjusts the N first voltages and the N first voltage sets based on the N fifth optical signals.
8. The method of claim 7, wherein the first modulation module of the optical neural network device modulates the initial optical signal based on the first voltage to obtain a first optical signal, comprising:
the optical generation submodule of the first modulation module generates N paths of parallel initial optical signals with different wavelengths;
the first modulation submodule of the first modulation module modulates the intensity of each path of parallel initial optical signals respectively based on the input N first voltages to obtain N paths of initial optical signals with different intensities;
and the combining submodule of the first modulation module combines the N paths of initial optical signals with different intensities into the first optical signal.
9. The method of claim 8, wherein the light generation sub-module generates N parallel primary optical signals of different wavelengths, comprising:
an optical frequency comb of the optical generation submodule generates an initial optical signal with N wavelengths;
and the wave separator of the light generation submodule divides the initial optical signal with N wavelengths into N parallel initial optical signals with different wavelengths.
10. The method of claim 7, wherein the dispersive module comprises a dispersive optical fiber and/or a dispersive grating; the inverse dispersion module comprises an inverse dispersion optical fiber and an inverse dispersion grating.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210140961.7A CN114565091A (en) | 2022-02-16 | 2022-02-16 | Optical neural network device, chip and optical implementation method for neural network calculation |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210140961.7A CN114565091A (en) | 2022-02-16 | 2022-02-16 | Optical neural network device, chip and optical implementation method for neural network calculation |
Publications (1)
Publication Number | Publication Date |
---|---|
CN114565091A true CN114565091A (en) | 2022-05-31 |
Family
ID=81714465
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210140961.7A Pending CN114565091A (en) | 2022-02-16 | 2022-02-16 | Optical neural network device, chip and optical implementation method for neural network calculation |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114565091A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115130666A (en) * | 2022-08-31 | 2022-09-30 | 之江实验室 | Two-dimensional photon convolution acceleration method and system |
CN115508958A (en) * | 2022-10-08 | 2022-12-23 | 深圳中科天鹰科技有限公司 | Photon chip based on optical neural network |
-
2022
- 2022-02-16 CN CN202210140961.7A patent/CN114565091A/en active Pending
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115130666A (en) * | 2022-08-31 | 2022-09-30 | 之江实验室 | Two-dimensional photon convolution acceleration method and system |
CN115508958A (en) * | 2022-10-08 | 2022-12-23 | 深圳中科天鹰科技有限公司 | Photon chip based on optical neural network |
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 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2022016894A1 (en) | Photonic neural network | |
CN111882052B (en) | Photon convolution neural network system | |
CN114565091A (en) | Optical neural network device, chip and optical implementation method for neural network calculation | |
CN112232503B (en) | Computing device, computing method, and computing system | |
US10908634B1 (en) | Systems and methods for parallel photonic computing | |
CN116348886A (en) | Coherent photon computing architecture | |
CN112308224B (en) | Optical neural network device, chip and optical implementation method for neural network calculation | |
CN106656878B (en) | Apparatus and method for processing optical signals | |
Jiang et al. | Photonic convolution neural network based on interleaved time-wavelength modulation | |
CN110908428A (en) | Parallel optical computing system for efficiently realizing large-scale matrix operation | |
CN112101540B (en) | Optical neural network chip and calculation method thereof | |
US20240063936A1 (en) | Systems and methods for utilizing photonic degrees of freedom in a photonic processor | |
CN115222035B (en) | Photon neural network convolution acceleration chip | |
US20230351167A1 (en) | Frequency multiplexed photonic neural networks | |
CN116484931B (en) | Photon matrix multiplication operation device and operation method for neural network | |
Shi et al. | Image classification with a 3-layer soa-based photonic integrated neural network | |
Shi et al. | Parallel photonic convolutional processing on-chip with cross-connect architecture and cyclic AWGs | |
CN112232487B (en) | Optical neural network chip and calculation method thereof | |
Zhang et al. | Time-stretch optical neural network with time-division multiplexing | |
Mourgias-Alexandris et al. | Neuromorphic computing through photonic integrated circuits | |
Pappas et al. | A TeraFLOP Photonic Matrix Multiplier using Time-Space-Wavelength Multiplexed AWGR-based Architectures | |
Pappas et al. | A 160 TOPS Multi-dimensional AWGR-based accelerator for Deep Learning | |
WO2023162857A1 (en) | Optical integrated circuit, information processing device, and computation method | |
Nakajima et al. | Densely parallelized photonic tensor processor on hybrid waveguide/free-space-optics | |
US20230376740A1 (en) | Multicolor tunable reservoir computing method and system |
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 |