CN116432726A - Photoelectric hybrid deep neural network operation device and operation method - Google Patents
Photoelectric hybrid deep neural network operation device and operation method Download PDFInfo
- Publication number
- CN116432726A CN116432726A CN202310703089.7A CN202310703089A CN116432726A CN 116432726 A CN116432726 A CN 116432726A CN 202310703089 A CN202310703089 A CN 202310703089A CN 116432726 A CN116432726 A CN 116432726A
- Authority
- CN
- China
- Prior art keywords
- optical
- signals
- signal
- coefficient
- 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.)
- Granted
Links
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 55
- 238000000034 method Methods 0.000 title claims abstract description 55
- 230000003287 optical effect Effects 0.000 claims abstract description 210
- 239000011159 matrix material Substances 0.000 claims abstract description 68
- 230000008569 process Effects 0.000 claims abstract description 33
- 125000004122 cyclic group Chemical group 0.000 claims abstract description 24
- 239000013307 optical fiber Substances 0.000 claims description 51
- 239000000047 product Substances 0.000 claims description 21
- 108010076504 Protein Sorting Signals Proteins 0.000 claims description 17
- 230000000737 periodic effect Effects 0.000 claims description 16
- 239000004065 semiconductor Substances 0.000 claims description 15
- 238000012545 processing Methods 0.000 claims description 12
- 230000001360 synchronised effect Effects 0.000 claims description 12
- 230000004913 activation Effects 0.000 claims description 7
- 239000012467 final product Substances 0.000 claims description 7
- 230000006870 function Effects 0.000 claims description 7
- 230000009466 transformation Effects 0.000 claims description 6
- 238000004891 communication Methods 0.000 claims description 3
- 239000000835 fiber Substances 0.000 claims description 3
- 230000005693 optoelectronics Effects 0.000 claims 1
- 239000010410 layer Substances 0.000 abstract description 7
- 230000003111 delayed effect Effects 0.000 abstract description 5
- 239000002356 single layer Substances 0.000 abstract description 3
- 238000010586 diagram Methods 0.000 description 12
- 238000005516 engineering process Methods 0.000 description 9
- 238000013507 mapping Methods 0.000 description 8
- 238000013473 artificial intelligence Methods 0.000 description 3
- 230000009286 beneficial effect Effects 0.000 description 3
- 239000000969 carrier Substances 0.000 description 2
- 230000010354 integration Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 101100457838 Caenorhabditis elegans mod-1 gene Proteins 0.000 description 1
- 101150110972 ME1 gene Proteins 0.000 description 1
- XUIMIQQOPSSXEZ-UHFFFAOYSA-N Silicon Chemical compound [Si] XUIMIQQOPSSXEZ-UHFFFAOYSA-N 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 239000006185 dispersion Substances 0.000 description 1
- 238000005265 energy consumption Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000010365 information processing Effects 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
- 230000000750 progressive effect Effects 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 229910052710 silicon Inorganic materials 0.000 description 1
- 239000010703 silicon Substances 0.000 description 1
- IRQRBVOQGUPTLG-UHFFFAOYSA-M sodium;3-(n-ethyl-3-methylanilino)-2-hydroxypropane-1-sulfonate Chemical compound [Na+].[O-]S(=O)(=O)CC(O)CN(CC)C1=CC=CC(C)=C1 IRQRBVOQGUPTLG-UHFFFAOYSA-M 0.000 description 1
- 238000012549 training Methods 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06E—OPTICAL COMPUTING DEVICES; COMPUTING DEVICES USING OTHER RADIATIONS WITH SIMILAR PROPERTIES
- G06E3/00—Devices not provided for in group G06E1/00, e.g. for processing analogue or hybrid data
- G06E3/001—Analogue devices in which mathematical operations are carried out with the aid of optical or electro-optical elements
- G06E3/005—Analogue devices in which mathematical operations are carried out with the aid of optical or electro-optical elements using electro-optical or opto-electronic means
-
- 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)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biophysics (AREA)
- Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Biomedical Technology (AREA)
- Computational Linguistics (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- Neurology (AREA)
- Nonlinear Science (AREA)
- Optics & Photonics (AREA)
- Optical Communication System (AREA)
Abstract
The invention discloses a photoelectric mixed deep neural network operation device, which is characterized in that each row of coefficient signals of a coefficient matrix signal are respectively loaded on corresponding sub-modulation optical signals by utilizing a two-dimensional coefficient loading unit to obtain multi-channel coefficient signals, and the multi-channel coefficient optical signals are delayed by an interval delay method to ensure that the obtained multi-channel delay signals can be converged, so that the coefficient matrix signal is loaded on an initial node signal, the purpose of simply and efficiently multiplying the coefficient matrix data on the initial node data is achieved, and the coefficient weighting of a single-layer two-dimensional neural network is realized. The invention also inputs the process node signals into the electro-optical modulator again through the node data source for cyclic operation, so that the process node data can be multiplied by the coefficient matrix data, and the coefficient weighting of the multi-layer two-dimensional neural network is realized through multiple cycles. The invention also discloses an operation method of the photoelectric hybrid deep neural network.
Description
Technical Field
The invention belongs to the field of photon computer systems, and particularly relates to a photoelectric hybrid deep neural network operation device and an operation method.
Background
Artificial intelligence is widely applied to the fields of machine vision, natural language processing, automatic driving and the like nowadays, and because the current electronic chip adopts a classical computer structure which separates a program space from a data space, the data load between a storage unit and a computing unit is unstable, the power consumption is higher, and the training efficiency of a network model is limited.
The common solution is to improve the operation efficiency by improving the integration level of an electronic chip or by in-memory calculation, but is limited by the micro quantum characteristic and the macro high-frequency response characteristic of the electronic chip, and electrons are used as carriers for information processing and transmission to face a series of inherent bottlenecks such as crosstalk, power consumption, time delay and the like, so that the energy consumption of the traditional convolutional neural network is larger and the speed is lower. These technical directions also face significant challenges.
The literature [ Shastri B J, tait a N, ferreira de Lima T, et al Photonics for artificial intelligence and neuromorphic computing Nature Photonics, 2021, 15 (2): 102-114 ] discloses that photonic technologies using photons as information carriers have large bandwidth, low loss, and parallelism, and in contrast, photonic devices are also used to accelerate the process of linear operations due to their higher bandwidth and parallelism. Researchers have been attracted to apply photonic technology in the field of artificial intelligence.
The literature [ Huang C, fujisawa S, de Lima T F, et al A silicon photonic-electronic neural network for fibre nonlinearity accounting Nature Electronics, 2021, 4 (11): 837-844 ] discloses the combination of photonic technology and traditional neural networks, is hopeful to fully exert the advantages of the two technologies, breaks through the bottleneck of the technology development of the traditional electronic neural network with high power consumption, long delay and limited speed, and solves the technical problem of the limitation of the traditional electronic technology.
The literature [ Xu X, tan M, corcoran B, et al, "11 TOOS photonic convolutional accelerator for optical neural networks," Nature, vol. 589, no. 7840, pp. -51, 2021. ]) discloses to provide a method for realizing convolution operation of a signal to be convolved and a fully connected feedforward neural network based on a dispersion technology, the operation speed is close to that of the latest chip based on the existing electronic technology, the power consumption of the scheme is greatly reduced, and a reliable basis is provided for the photon neural network to be put into practical use. However, the photonic neural network is limited by the problem of limited integration of the photonic technology, is mainly concentrated on single-layer operation at present, and cannot support a large-scale deep neural network.
Therefore, it is highly desirable to design a method for realizing multi-layer operation by cyclic operation, so as to support a large-scale deep neural network.
Disclosure of Invention
The invention discloses a photoelectric hybrid deep neural network operation device which can realize coefficient weighting of a two-dimensional neural network and can efficiently realize operation of a multi-layer neural network through cyclic operation.
The embodiment of the invention provides a photoelectric hybrid deep neural network operation device, which comprises:
the multi-wavelength light source is used for outputting multi-wavelength optical carrier signals;
the wavelength division switching unit is used for modulating the multi-wavelength optical carrier signal into a periodic frequency hopping optical signal by controlling a switching time sequence;
the electro-optical modulator is used for loading the initial node signal on the periodic frequency hopping signal to obtain a modulated optical signal;
the two-dimensional coefficient loading unit is used for splitting the modulated optical signals into multi-path sub-modulated optical signals, respectively loading each row of coefficient signals of the coefficient matrix signals on the multi-path sub-modulated optical signals to obtain multi-path coefficient optical signals, and delaying the multi-path coefficient optical signals at equal intervals to obtain multi-path delay optical signals;
the optical fiber delay unit is used for respectively carrying out time alignment on a plurality of coefficient frequency components of each path of delay optical signal and then converging to obtain an initial product optical signal sequence;
the node data source is used for outputting an initial node signal, converting the process node data into a process node signal, and inputting the process node signal into the electro-optical modulator again for cyclic operation again;
and the output unit is used for converting the initial product optical signal sequence into process node data to complete one-time cyclic operation, and converting the final product optical signal sequence obtained after the cyclic operation is finished into final node data to complete the operation of the deep neural network.
Further, the two-dimensional coefficient loading unit comprises a beam splitter, a plurality of sub-modulators, a first optical fiber delay array and a beam combiner;
the beam splitter is used for dividing the received modulated optical signal into multiple sub-modulated optical signals;
each sub-modulator is used for loading data of the coefficient matrix signals onto corresponding modulated optical signals to obtain corresponding coefficient optical signals;
the first optical fiber delay array consists of a plurality of optical fibers with sequentially increased lengths, and each path of coefficient optical signals is transmitted to the beam combiner through the corresponding optical fiber, so that each path of coefficient optical signals is subjected to equidistant delay to obtain multipath delay coefficient signals;
the beam combiner is used for combining the multipath delay coefficient signals into one path to obtain a two-dimensional modulation optical signal.
Further, each sub-modulator is configured to load data of the coefficient matrix signal onto a corresponding modulated optical signal to obtain a corresponding coefficient optical signal, and includes:
each sub-modulated optical signal comprises a modulated frequency hopping component with different frequencies, and each modulated frequency hopping component is loaded with a corresponding code element of a node signal;
each code element of the row coefficient of the coefficient matrix signal is loaded on the corresponding modulation frequency hopping component in the corresponding sub-modulation optical signal to obtain a plurality of coefficient frequency hopping components, and the corresponding coefficient optical signal is constructed based on the plurality of coefficient frequency hopping components, so that the multiplication of the node signal and the row coefficient of the corresponding coefficient matrix signal is realized.
Further, the optical fiber delay unit comprises a first wavelength division multiplexer, a second optical fiber delay array and a first wavelength division multiplexer;
the first wavelength division demultiplexer is used for dividing the two-dimensional modulation optical signal into a plurality of optical signals with the same frequency, and each optical signal with the same frequency is a set of coefficient frequency hopping components with the same frequency in the two-dimensional modulation optical signal;
the second optical fiber delay array consists of a plurality of optical fibers with gradually increased lengths, and the time of each optical signal with the same frequency is aligned through the optical fibers with different lengths, so that a plurality of coefficient frequency components of each coefficient optical signal are aligned in time, and the frequency components after being aligned in time are output to the first wavelength division multiplexer;
the first wavelength division multiplexer is used for converging the plurality of time-aligned coefficient frequency components to obtain initial product optical signals of the corresponding coefficient optical signals, and constructing an initial product optical signal sequence based on the initial product optical signals respectively corresponding to the plurality of paths of coefficient optical signals.
Further, the output unit comprises a photoelectric detector and an acquisition processing unit;
the photoelectric detector is used for converting an initial product optical signal sequence into an initial analog electric signal and converting a final product optical signal obtained after the cyclic operation is finished into a final analog electric signal;
the acquisition processing unit is used for converting the received initial analog electric signals into initial digital electric signals, and respectively carrying out nonlinear transformation on elements of the initial digital electric signals through nonlinear activation functions to obtain process node data;
the acquisition processing unit is also used for converting the received final analog electric signals into final digital electric signals, and respectively carrying out nonlinear transformation on elements of the final analog electric signals through nonlinear activation functions to obtain final node data.
Further, the wavelength division switching unit comprises a second wavelength division demultiplexer, a wavelength division switching array and a second wavelength division multiplexer;
the second wavelength division multiplexer is used for dividing the multi-wavelength optical carrier signals into a plurality of optical carrier signals, and each optical carrier signal is input to the wavelength division multiplexer through a corresponding semiconductor amplifier;
the wavelength division optical switch array comprises a plurality of semiconductor optical amplifiers, a switching time sequence formed by the switching time of the plurality of semiconductor amplifiers is formed by making the switching time of the different semiconductor amplifiers different, corresponding optical carrier signals are amplified into corresponding optical pulse signals through each semiconductor amplifier, and then the plurality of optical pulse signals are mutually separated at a specified frequency hopping interval in a set period based on the switching time sequence;
the second wavelength division multiplexer is used for combining the plurality of separated optical pulse signals into one path to obtain a periodic frequency hopping optical signal.
Further, the system also comprises a coefficient matrix source, wherein the coefficient matrix source is used for converting a plurality of coefficient matrix data into corresponding coefficient matrix signals, and in each cyclic operation, the designated coefficient matrix signals are input to a two-dimensional coefficient loading unit;
the number of lines of each coefficient matrix data is not greater than the number of wavelengths of the multi-wavelength optical carrier signal.
The system further comprises a synchronous control unit, wherein the output end of the synchronous control unit is respectively in communication connection with the input ends of the wave-splitting switch array, the node data source and the coefficient matrix source;
the synchronous control unit is used for respectively outputting synchronous control signals to the wavelength division optical switch array and the coefficient matrix source, so that the number of formed optical pulse signals is the same as the number of lines of the coefficient matrix data, and the purpose that the number of code elements of the process node signals is the same as the number of optical pulse signals of the periodic frequency hopping optical signals is achieved;
the synchronization control unit is further configured to output a time control signal to the node data source, so that the code element of the process node signal or the initial node signal can be time-synchronized with the corresponding optical pulse signal, and further the code element of the node signal or the initial node signal can be loaded on the corresponding optical pulse signal.
Further, the duration of each optical pulse signal is the same as the duration of the symbol of the node signal and the symbol of the coefficient matrix signal.
The specific embodiment of the invention also provides a photoelectric hybrid deep neural network operation method, which is characterized in that node data and coefficient matrix signals are input into the photoelectric hybrid deep neural network operation device, so that the coefficient matrix signals are loaded on the node signals to obtain final node signals, and the final node signals are converted into final node data, thereby completing the operation of the deep neural network.
Compared with the prior art, the invention has the beneficial effects that:
according to the invention, each row of coefficient signals of the coefficient matrix signals are respectively loaded on the corresponding sub-modulation optical signals by utilizing the two-dimensional coefficient loading unit to obtain multiple paths of coefficient signals, and the multiple paths of coefficient optical signals are delayed by an interval delay method so that the obtained multiple paths of delay signals can be converged, thereby completing the loading of the coefficient matrix signals on the initial node signals, further achieving the purpose of simply and efficiently multiplying the coefficient matrix data on the initial node data, and realizing the coefficient weighting of the single-layer two-dimensional neural network.
The invention also inputs the process node signals into the electro-optical modulator again through the node data source for cyclic operation, so that the process node data can be multiplied by the coefficient matrix data, and the coefficient weighting of the multi-layer two-dimensional neural network is realized through multiple cycles.
Drawings
Fig. 1 is a block diagram of a photoelectric hybrid deep neural network computing device according to an embodiment of the present invention;
fig. 2 is a schematic structural diagram of an optical-electrical hybrid deep neural network computing device according to an embodiment of the present invention;
fig. 3 is a schematic structural diagram of a wave-splitting switch unit according to an embodiment of the present invention;
FIG. 4 is a schematic diagram of a two-dimensional coefficient loading unit according to an embodiment of the present invention;
FIG. 5 is a schematic diagram of an optical fiber delay unit according to an embodiment of the present invention;
fig. 6 is a schematic diagram of a time-frequency mapping relationship of a periodic frequency hopping optical signal output by a wavelength division switching unit according to an embodiment of the present invention;
FIG. 7 is a schematic diagram of a time-frequency mapping relationship of an initial modulated optical signal according to an embodiment of the present invention;
FIG. 8 is a schematic diagram of a time-frequency mapping relationship of a first delay signal transmitted by a first optical fiber in a two-dimensional coefficient loading module according to an embodiment of the present invention;
FIG. 9 is a schematic diagram of a time-frequency mapping relationship of a second delay signal transmitted by a second optical fiber in a two-dimensional coefficient loading module according to an embodiment of the present invention;
FIG. 10 is a schematic diagram of a time-frequency mapping relationship of an N-th delay signal transmitted by an N-th optical fiber in a two-dimensional coefficient loading module according to an embodiment of the present invention;
FIG. 11 is a schematic diagram of a time-frequency mapping relationship of a two-dimensional modulated optical signal according to an embodiment of the present invention;
fig. 12 is a schematic diagram of a time-frequency mapping relationship of an initial product optical signal according to an embodiment of the present invention.
Detailed Description
Aiming at the defects of the prior art, the concept of the invention is to take a periodical frequency hopping optical signal as an optical carrier, carry out full-wavelength delay on a multi-channel coefficient optical signal based on a two-dimensional coefficient loading unit and carry out delay alignment on different coefficient frequency components based on an optical fiber delay array, realize coefficient weighting of a two-dimensional neural network through time multiplexing, and sequentially realize operation of a multi-layer neural network by combining the nonlinear processing capacity of an electric domain and utilizing the dynamic reconfigurable characteristic of a system on the basis. The method is simple and efficient, the system can be flexibly expanded, and the method is suitable for multidimensional deep neural network operation. The following describes a photoelectric hybrid deep neural network computing device provided in an embodiment of the present invention in detail.
The embodiment of the invention provides a photoelectric hybrid deep neural network operation device, shown in fig. 1, comprising: the multi-wavelength light source is used for outputting multi-wavelength optical carrier signals; the wavelength division switching unit is used for modulating the multi-wavelength optical carrier signal into a periodic frequency hopping optical signal by controlling a switching time sequence; the electro-optical modulator is used for loading the initial node signal on the periodic frequency hopping signal to obtain a modulated optical signal; the two-dimensional coefficient loading unit is used for splitting the modulated optical signals into multi-path sub-modulated optical signals, respectively loading each row of coefficient signals of the coefficient matrix signals on the multi-path sub-modulated optical signals to obtain multi-path coefficient optical signals, and delaying the multi-path coefficient optical signals at equal intervals to obtain multi-path delay optical signals; the optical fiber delay unit is used for respectively carrying out time alignment on a plurality of coefficient frequency components of each path of delay optical signal and then converging to obtain an initial product optical signal sequence; the node data source is used for outputting an initial node signal, converting the process node data into a process node signal, and inputting the process node signal into the electro-optical modulator again for cyclic operation again; and the output unit is used for converting the initial product optical signal sequence into process node data to complete one-time cyclic operation, and converting the final product optical signal sequence obtained after the cyclic operation is finished into final node data to complete the operation of the deep neural network.
In a specific embodiment, the present invention provides a deep neural network computing device with photoelectric mixing, as shown in fig. 2, including: the system comprises a multi-wavelength light source, a wave splitting switch unit, an electro-optical modulator, a two-dimensional coefficient loading unit, an optical fiber delay unit, a node data source, a coefficient matrix source, a synchronous control unit and an output unit. The output unit comprises a photoelectric detector and an acquisition processing unit.
The multi-wavelength light source provided by the embodiment of the invention is used for outputting multi-wavelength optical carrier signals of M frequency components, and the multi-wavelength optical signals with equal amplitudes of the frequency components can be expressed as A= [ A, A, A, …, A by using a matrix] T M×1 Wherein M is a positive integer, and A is single-wavelength signal strength.
The wavelength division switching unit provided by the embodiment of the invention is used for modulating the multi-wavelength optical carrier signal into a periodic frequency hopping optical signal by controlling the switching time sequence.
In a specific embodiment, as shown in fig. 3, the wavelength division switching unit includes a second wavelength division multiplexer, a wavelength division switching array, and a second wavelength division multiplexer.
Wherein, the multi-wavelength optical carrier signals are input into the second demultiplexer to obtain M optical carrier signals, the wavelength division optical switch array provided by the embodiment of the invention comprises a plurality of semiconductor optical amplifiers, namely SOA1, SOA2, … and SOAM,the switching time sequence of the semiconductor optical amplifiers is controlled by the synchronous control signals output by the synchronous control unit, so that M semiconductor optical amplifiers are arranged at time intervals
The method comprises the steps of sequentially switching on and off, determining the starting quantity of M semiconductor amplifiers based on synchronous control signals, enabling the quantity of optical carrier signals to be consistent with the number of lines of a coefficient matrix, and enabling the quantity of the optical carrier signals to be consistent with the number of lines of the coefficient matrix as the number of lines of the coefficient matrix determines the number of symbols of the node signals, so that the number of symbols of the node signals can be consistent with the number of the optical carrier signals, and conditions are created for loading the number of symbols of the node signals onto the corresponding optical carrier signals.
The second wavelength division multiplexer is used for combining a plurality of separated optical pulse signals into one path to obtain a periodical frequency hopping optical signal, thereby forming a designated frequency hopping interval in a set period T
Is provided. Periodic frequency hopping optical signal L output by wavelength division optical switch array s As shown in fig. 6, a periodic frequency hopping optical signal L according to an embodiment of the present invention s The method comprises the following steps:
the electro-optical modulator provided by the embodiment of the invention is used for transmitting an initial node signal x= [ x ] 1 ,x 2 ,x 3 ,…,x M ]Is loaded onto the corresponding optical pulse signal of the periodic frequency hopping signal to obtain an initial modulated optical signal composed of a plurality of modulated frequency hopping components, the time-frequency mapping relation of which is shown in fig. 7, the initial modulated optical signal S of the first cyclic operation Mod_1 The method comprises the following steps:
the two-dimensional coefficient loading unit provided by the embodiment of the invention is used for splitting an initial modulation optical signal into a plurality of paths of sub-modulation optical signals, respectively loading each row of coefficient signals of the coefficient matrix signals on the plurality of paths of sub-modulation optical signals to obtain a plurality of paths of coefficient optical signals, and delaying the plurality of paths of coefficient optical signals at equal intervals to obtain a plurality of paths of delay optical signals.
In a specific embodiment, as shown in fig. 4, the two-dimensional coefficient loading unit provided in the specific embodiment of the present invention includes a1×n beam splitter, a plurality of sub-modulators, that is, sub-modulator 1 … sub-modulator N, a first fiber delay array, and an n×1 beam combiner.
Wherein the 1 xn splitter is configured to split the initial modulated optical signal into a plurality of sub-modulated optical signals.
The plurality of sub-modulators provided in this embodiment load each row of coefficient signals of the matrix coefficient signals output by the matrix coefficient source onto a corresponding sub-modulated optical signal to obtain corresponding N coefficient optical signals, where each coefficient optical signal includes a plurality of coefficient frequency hopping components, and each coefficient frequency hopping component is obtained by loading a symbol of each row of coefficient signals of the matrix coefficient signal onto a corresponding modulation frequency hopping component of the corresponding sub-modulated optical signal.
The first optical fiber delay array provided in this embodiment is composed of a plurality of optical fibers with sequentially increased lengths, and an input end of each optical fiber is connected with an output end of a corresponding sub-modulator.
The specific steps for determining the length of each optical fiber provided by the specific embodiment of the invention are as follows: the N coefficient optical signals correspond to the N optical fibers, the first optical fiber is taken as a reference, the coefficient optical signals input to the first optical fiber are loaded with a first row of coefficient signals in the coefficient matrix signals, and the first row of coefficient signals are [ W ] 11 ,W 12 …,W 1M ]The lengths of other sections of optical fibers are sequentially increased
Wherein->
Duration of a single symbol for a node signal，cFor the speed of the light in vacuum,n f for the refractive index of the optical fiber, thereby realizing equal interval delay of each coefficient optical signal>
N-path delay signals are obtained according to the time of the first cyclic operation, and the nth delay signal is S Mod_1_n The method comprises the following steps:
it can be understood that, since each delay signal corresponds to the number of rows of the two-dimensional coefficient matrix one by one, the index of the delay signal and the number of rows of the two-dimensional coefficient matrix are both denoted by N, and the total number is N. As shown in fig. 8, the first delay signal transmitted by the first optical fiber is not delayed; as shown in fig. 9, the second delay signal transmitted by the second optical fiber is delayed as a whole compared with the first delay signal
Time of (2); as shown in FIG. 10, the N-th delay signal transmitted by the N-th optical fiber is delayed by +.>
Is a time of (a) to be used.
The n×1 combiner provided in this embodiment is configured to combine N delay coefficient signals into one path to obtain a two-dimensional modulated optical signal, as shown in fig. 11.
An embodiment of the present invention provides an optical fiber delay unit, as shown in fig. 5, including a first wavelength division multiplexer, a second optical fiber delay array, and a first wavelength division multiplexer.
The first demultiplexer provided in the embodiment of the present invention is configured to decompose a two-dimensional modulated optical signal into a plurality of optical signals with the same frequency, construct a corresponding optical signal with the same frequency based on a plurality of coefficient frequency hopping components with the same frequency in the two-dimensional modulated optical signal, and return to fig. 11, for example, to equalize the frequencies f 1 Ax of (2) 1 w 11 ，Ax 1 w 21 ，…，Ax 1 w N1 The combined frequencies are f 1 Is a single-frequency optical signal.
The second optical fiber delay array provided by the embodiment of the invention consists of a plurality of optical fibers with gradually increased lengths, and the determining mode of the length of each optical fiber of the second optical fiber delay array is the same as that of the length of each optical fiber of the first optical fiber delay array. As shown in FIG. 5, the longer frequencies of the input first optical fiber are f M The same frequency optical signals are sequentially input into corresponding optical fibers based on the time delay from long to short, namely the frequency input into the first optical fiber with the longest time delay is f M Is input with an increased length of
Is the frequency f 1 Is a single-frequency optical signal.
The time of each optical signal with the same frequency is aligned through optical fibers with different lengths, so that a plurality of coefficient frequency components of each coefficient optical signal are aligned in time, and the frequency components after being aligned in time are output to a first wavelength division multiplexer to obtain an initial product optical signal, as shown in fig. 12.
The first wavelength division multiplexer provided by the embodiment of the invention is used for converging the plurality of coefficient frequency components aligned in time to obtain initial product optical signals of the corresponding coefficient optical signals, and constructing initial product optical signal sequences [ S (1), S (2) …, S (N-1) and S (N) ] based on the initial product optical signals respectively corresponding to the plurality of paths of coefficient optical signals.
Returning to fig. 2, the output end of the synchronization control unit provided in the embodiment of the present invention is respectively connected with the input ends of the wavelength division optical switch array, the node data source and the coefficient matrix source in a communication manner.
The synchronous control unit provided by the embodiment of the invention is used for respectively outputting synchronous control signals to the wavelength division optical switch array and the coefficient matrix source, so that the number of formed optical pulse signals is the same as the number of lines of the coefficient matrix data, and the purpose that the number of code elements of the process node signals is the same as the number of optical pulse signals of the periodic frequency hopping optical signals is achieved.
The synchronization control unit provided by the embodiment of the invention is further used for outputting a time control signal to the node data source, so that the code element of the process node signal or the initial node signal can be time-synchronized with the corresponding optical pulse signal and can be loaded on the corresponding optical pulse signal.
The coefficient matrix source provided by the embodiment of the invention is used for converting a plurality of coefficient matrix data into corresponding coefficient matrix signals, and the designated coefficient matrix signals are input to the two-dimensional coefficient loading unit in each cyclic operation.
The output unit provided by the embodiment of the invention is used for converting the initial product optical signal sequence into process node data and converting the final product optical signal sequence obtained after the cyclic operation is finished into final node data so as to finish the operation of the deep neural network.
The output unit provided by the embodiment of the invention comprises a photoelectric detector and an acquisition processing unit; the photoelectric detector is used for converting the initial product optical signal sequence into an initial analog electric signal, wherein an nth element R (n) in the initial analog electric signal is as follows:
it can be understood that, since each element in the initial analog electrical signal corresponds to the number of rows of the two-dimensional coefficient matrix one by one, the index of the initial analog electrical signal and the number of rows of the two-dimensional coefficient matrix are both denoted by N, and the total number is N, and since the rows of the coefficient matrix correspond to the symbols of the node signal one by one, the index of each column of the coefficient matrix and the index of each row of the symbol of the node signal are both M, and the number of symbols of each row of the coefficient signal, the number of modulation frequency hopping components of the first modulated optical signal, and the number of optical pulse signals are all the same as M.
The acquisition processing unit provided by the embodiment of the invention is used for converting the received initial analog electric signal into the initial digital electric signal, and respectively carrying out nonlinear transformation on the elements of the initial digital electric signal through the nonlinear activation function to obtain process node data, thereby completing one-time cyclic operation.
The node data source provided by the embodiment of the invention is used for providing an initial node signal, generating a process node signal from the process node data in series, returning to fig. 2, inputting the process node signal into the electro-optical modulator again, performing cyclic operation again, converting a final product optical signal obtained after the cyclic operation is finished into a final analog electric signal by the photoelectric detector, converting the received final analog electric signal into a final digital electric signal by the acquisition processing unit, and respectively performing nonlinear transformation on elements of the final analog electric signal through a nonlinear activation function to obtain final node data, thereby completing the operation of a multi-layer two-dimensional deep neural network.
The duration of each optical pulse signal provided by the embodiment of the invention is the same as the duration of the code element of the node signal and the code element of the coefficient matrix signal.
The embodiment of the invention also provides a photoelectric hybrid deep neural network operation method, which inputs the node data and the coefficient matrix signals into the photoelectric hybrid deep neural network operation device provided by the embodiment of the invention, so that the coefficient matrix signals are loaded on the node signals to obtain final node signals, and the final node signals are converted into final node data, thereby completing the operation of the deep neural network.
The technical scheme of the invention has the following beneficial effects:
(1) The invention realizes the switch control of the multi-wavelength optical carrier signal based on the semiconductor optical amplifier, and can realize nanosecond switching speed, thereby being beneficial to the improvement of photon operation speed.
(2) According to the invention, based on the two-dimensional coefficient loading module, the simultaneous delay of different frequency signals and the delay alignment of different frequency signals are realized by the optical fiber delay array, the coefficient weighting of the two-dimensional neural network can be realized by combining time multiplexing with frequency multiplexing, and the scheme is simple and efficient.
(3) The invention combines nonlinear processing capability of nonlinear activation function of electric domain, and utilizes dynamic reconfigurable characteristic of system to realize operation of multi-layer neural network in sequence.
It should also be noted that 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 one … …" does not exclude the presence of other like elements in a process, method, article or apparatus that comprises the element.
The embodiments of the present invention are described in a progressive manner, and the same and similar parts of the embodiments are all referred to each other, and each embodiment is mainly described in the differences from the other embodiments. In particular, for system embodiments, since they are substantially similar to method embodiments, the description is relatively simple, as relevant to see a section of the description of method embodiments.
The foregoing description of the preferred embodiments of the invention is not intended to be limiting, but rather is intended to cover all modifications, equivalents, or alternatives falling within the spirit and principles of the invention.
Finally, it should be noted that the above list is only specific embodiments of the present invention. The invention is not limited to the above embodiments, but many variations are possible. All modifications directly derived or suggested to one skilled in the art from the present disclosure should be considered as being within the scope of the present invention.
Claims (10)
1. An optoelectronic hybrid deep neural network computing device, comprising:
the multi-wavelength light source is used for outputting multi-wavelength optical carrier signals;
the wavelength division switching unit is used for modulating the multi-wavelength optical carrier signal into a periodic frequency hopping optical signal by controlling a switching time sequence;
the electro-optical modulator is used for loading the initial node signal on the periodic frequency hopping signal to obtain a modulated optical signal;
the two-dimensional coefficient loading unit is used for splitting the modulated optical signals into multi-path sub-modulated optical signals, respectively loading each row of coefficient signals of the coefficient matrix signals on the multi-path sub-modulated optical signals to obtain multi-path coefficient optical signals, and delaying the multi-path coefficient optical signals at equal intervals to obtain multi-path delay optical signals;
the optical fiber delay unit is used for respectively carrying out time alignment on a plurality of coefficient frequency components of each path of delay optical signal and then converging to obtain an initial product optical signal sequence;
the node data source is used for outputting an initial node signal, converting the process node data into a process node signal, and inputting the process node signal into the electro-optical modulator again for cyclic operation again;
and the output unit is used for converting the initial product optical signal sequence into process node data to complete one-time cyclic operation, and converting the final product optical signal sequence obtained after the cyclic operation is finished into final node data to complete the operation of the deep neural network.
2. The electro-optical hybrid deep neural network computing device of claim 1, wherein the two-dimensional coefficient loading unit comprises a beam splitter, a plurality of sub-modulators, a first fiber delay array, and a beam combiner;
the beam splitter is used for dividing the received modulated optical signal into multiple sub-modulated optical signals;
each modulator is used for loading the data of the coefficient matrix signals onto the corresponding modulated optical signals to obtain corresponding coefficient optical signals;
the first optical fiber delay array consists of a plurality of optical fibers with sequentially increased lengths, and each path of coefficient optical signals is transmitted to the beam combiner through the corresponding optical fiber, so that each path of coefficient optical signals is subjected to equidistant delay to obtain multipath delay coefficient signals;
the beam combiner is used for combining the multipath delay coefficient signals into one path to obtain a two-dimensional modulation optical signal.
3. The electro-optical hybrid deep neural network computing device of claim 2, wherein each sub-modulator is configured to load data of the coefficient matrix signal onto a corresponding modulated optical signal to obtain a corresponding coefficient optical signal, and comprises:
each sub-modulated optical signal comprises a modulated frequency hopping component with different frequencies, and each modulated frequency hopping component is loaded with a corresponding code element of a node signal;
each code element of the row coefficient of the coefficient matrix signal is loaded on the corresponding modulation frequency hopping component in the corresponding sub-modulation optical signal to obtain a plurality of coefficient frequency hopping components, and the corresponding coefficient optical signal is constructed based on the plurality of coefficient frequency hopping components, so that the aim of multiplying the node signal with the row coefficient signal of the corresponding coefficient matrix signal is achieved.
4. The optical-electrical hybrid deep neural network computing device of claim 3, wherein the optical fiber delay unit comprises a first demultiplexer, a second optical fiber delay array, and a first wavelength division multiplexer;
the first wavelength division demultiplexer is used for dividing the two-dimensional modulation optical signal into a plurality of optical signals with the same frequency, and each optical signal with the same frequency is a set of coefficient frequency hopping components with the same frequency in the two-dimensional modulation optical signal;
the second optical fiber delay array consists of a plurality of optical fibers with gradually increased lengths, and the time of each optical signal with the same frequency is aligned through the optical fibers with different lengths, so that a plurality of coefficient frequency components of each coefficient optical signal are aligned in time, and the frequency components after being aligned in time are output to the first wavelength division multiplexer;
the first wavelength division multiplexer is used for converging the plurality of time-aligned coefficient frequency components to obtain initial product optical signals of the corresponding coefficient optical signals, and constructing an initial product optical signal sequence based on the initial product optical signals respectively corresponding to the plurality of paths of coefficient optical signals.
5. The photoelectric hybrid deep neural network computing device of claim 1, wherein the output unit comprises a photodetector and an acquisition processing unit;
the photoelectric detector is used for converting an initial product optical signal sequence into an initial analog electric signal and converting a final product optical signal obtained after the cyclic operation is finished into a final analog electric signal;
the acquisition processing unit is used for converting the received initial analog electric signals into initial digital electric signals, and respectively carrying out nonlinear transformation on elements of the initial digital electric signals through nonlinear activation functions to obtain process node data;
the acquisition processing unit is also used for converting the received final analog electric signals into final digital electric signals, and respectively carrying out nonlinear transformation on elements of the final analog electric signals through nonlinear activation functions to obtain final node data.
6. The photoelectric hybrid deep neural network operation device according to claim 1, wherein the wavelength division switching unit includes a second wavelength division multiplexer, a wavelength division switching array, and a second wavelength division multiplexer;
the second wavelength division multiplexer is used for dividing the multi-wavelength optical carrier signals into a plurality of optical carrier signals, and each optical carrier signal is input to the wavelength division multiplexer through a corresponding semiconductor amplifier;
the wavelength division optical switch array comprises a plurality of semiconductor optical amplifiers, a switching time sequence formed by the switching time of the plurality of semiconductor amplifiers is formed by making the switching time of the different semiconductor amplifiers different, corresponding optical carrier signals are amplified into corresponding optical pulse signals through each semiconductor amplifier, and then the plurality of optical pulse signals are mutually separated at a specified frequency hopping interval in a set period based on the switching time sequence;
the second wavelength division multiplexer is used for combining the plurality of separated optical pulse signals into one path to obtain a periodic frequency hopping optical signal.
7. The photoelectric hybrid deep neural network operation device according to claim 6, further comprising a coefficient matrix source for converting a plurality of coefficient matrix data into corresponding coefficient matrix signals, the specified coefficient matrix signals being input to the two-dimensional coefficient loading unit in each cyclic operation;
the number of lines of each coefficient matrix data is not greater than the number of wavelengths of the multi-wavelength optical carrier signal.
8. The photoelectric hybrid deep neural network computing device according to claim 7, further comprising a synchronization control unit, wherein an output end of the synchronization control unit is respectively in communication connection with input ends of the wavelength division optical switch array, the node data source and the coefficient matrix source;
the synchronous control unit is used for respectively outputting synchronous control signals to the wavelength division optical switch array and the coefficient matrix source, so that the number of formed optical pulse signals is the same as the number of lines of the coefficient matrix data, and the purpose that the number of code elements of the process node signals is the same as the number of optical pulse signals of the periodic frequency hopping optical signals is achieved;
the synchronization control unit is further configured to output a time control signal to the node data source, so that the code element of the process node signal or the initial node signal can be time-synchronized with the corresponding optical pulse signal, and further the code element of the node signal or the initial node signal can be loaded on the corresponding optical pulse signal.
9. The electro-optical hybrid deep neural network operation device of claim 6, wherein the duration of each optical pulse signal is the same as the duration of the symbols of the node signal and the symbols of the coefficient matrix signal.
10. A method for operating a deep neural network by photoelectric mixing, wherein node data and coefficient matrix signals are input into the device for operating a deep neural network by photoelectric mixing according to any one of claims 1 to 9, so that the coefficient matrix signals are loaded on the node signals to obtain final node signals, and the final node signals are converted into final node data, thereby completing the operation of the deep neural network.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310703089.7A CN116432726B (en) | 2023-06-14 | 2023-06-14 | Photoelectric hybrid deep neural network operation device and operation method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310703089.7A CN116432726B (en) | 2023-06-14 | 2023-06-14 | Photoelectric hybrid deep neural network operation device and operation method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116432726A true CN116432726A (en) | 2023-07-14 |
| CN116432726B CN116432726B (en) | 2023-08-25 |
Family
ID=87085862
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310703089.7A Active CN116432726B (en) | 2023-06-14 | 2023-06-14 | Photoelectric hybrid deep neural network operation device and operation method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116432726B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116957031A (en) * | 2023-07-24 | 2023-10-27 | 浙江大学 | Photoelectric computer based on optical multi-neuron activation function module |
| CN117436494A (en) * | 2023-12-18 | 2024-01-23 | 合肥硅臻芯片技术有限公司 | Neural network operation chip and operation method |
Citations (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US5258934A (en) * | 1990-05-14 | 1993-11-02 | California Institute Of Technology | Charge domain bit serial vector-matrix multiplier and method thereof |
| US6100831A (en) * | 1998-06-30 | 2000-08-08 | The United States Of America As Represented By The Secretary Of The Navy | Optoelectronic analog-to-digital converter using wavelength division multiplexing |
| US20110249755A1 (en) * | 2008-12-16 | 2011-10-13 | Youji Shibahara | Moving image coding method, moving image decoding method, moving image coding apparatus, moving image decoding apparatus, program, and integrated circuit |
| US20170351293A1 (en) * | 2016-06-02 | 2017-12-07 | Jacques Johannes Carolan | Apparatus and Methods for Optical Neural Network |
| JP2018205864A (en) * | 2017-05-31 | 2018-12-27 | 日本電信電話株式会社 | Optical neural network learning device |
| CN109871940A (en) * | 2019-01-31 | 2019-06-11 | 清华大学 | A kind of multilayer training algorithm of impulsive neural networks |
| US20190370644A1 (en) * | 2018-06-04 | 2019-12-05 | Lightmatter, Inc. | Convolutional layers for neural networks using programmable nanophotonics |
| CN112424796A (en) * | 2018-06-05 | 2021-02-26 | 光子智能股份有限公司 | Photoelectric computing system |
| JP2021093563A (en) * | 2019-12-06 | 2021-06-17 | ソニーセミコンダクタソリューションズ株式会社 | Solid state image sensor and control method |
| US20210342730A1 (en) * | 2020-05-01 | 2021-11-04 | equal1.labs Inc. | System and method of quantum enhanced accelerated neural network training |
| US20210341765A1 (en) * | 2020-04-29 | 2021-11-04 | Lightelligence, Inc. | Optical modulation for optoelectronic processing |
| CN113792870A (en) * | 2021-10-29 | 2021-12-14 | 广西大学 | All-optical control optical neural network system based on acousto-optic technology |
| EP3940696A1 (en) * | 2020-07-15 | 2022-01-19 | Österreichische Akademie der Wissenschaften | Inpainting based signal gap reconstruction with sparse representation |
| CN114358271A (en) * | 2022-03-18 | 2022-04-15 | 之江实验室 | Time-wavelength interweaving photon neural network convolution acceleration chip |
| CN114424213A (en) * | 2020-06-25 | 2022-04-29 | 普立恩科技有限公司 | Analog hardware implementation of neural networks |
| CN114462587A (en) * | 2022-02-10 | 2022-05-10 | 电子科技大学 | FPGA implementation method for photoelectric hybrid computing neural network |
| CN114584441A (en) * | 2022-03-02 | 2022-06-03 | 中国人民解放军军事科学院战争研究院 | Digital signal modulation identification method based on deep learning |
| CN115167610A (en) * | 2022-09-08 | 2022-10-11 | 之江实验室 | Two-dimensional photon convolution acceleration chip and application system and method |
-
2023
- 2023-06-14 CN CN202310703089.7A patent/CN116432726B/en active Active
Patent Citations (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US5258934A (en) * | 1990-05-14 | 1993-11-02 | California Institute Of Technology | Charge domain bit serial vector-matrix multiplier and method thereof |
| US6100831A (en) * | 1998-06-30 | 2000-08-08 | The United States Of America As Represented By The Secretary Of The Navy | Optoelectronic analog-to-digital converter using wavelength division multiplexing |
| US20110249755A1 (en) * | 2008-12-16 | 2011-10-13 | Youji Shibahara | Moving image coding method, moving image decoding method, moving image coding apparatus, moving image decoding apparatus, program, and integrated circuit |
| US20170351293A1 (en) * | 2016-06-02 | 2017-12-07 | Jacques Johannes Carolan | Apparatus and Methods for Optical Neural Network |
| JP2018205864A (en) * | 2017-05-31 | 2018-12-27 | 日本電信電話株式会社 | Optical neural network learning device |
| US20190370644A1 (en) * | 2018-06-04 | 2019-12-05 | Lightmatter, Inc. | Convolutional layers for neural networks using programmable nanophotonics |
| CN112424796A (en) * | 2018-06-05 | 2021-02-26 | 光子智能股份有限公司 | Photoelectric computing system |
| CN109871940A (en) * | 2019-01-31 | 2019-06-11 | 清华大学 | A kind of multilayer training algorithm of impulsive neural networks |
| JP2021093563A (en) * | 2019-12-06 | 2021-06-17 | ソニーセミコンダクタソリューションズ株式会社 | Solid state image sensor and control method |
| US20210341765A1 (en) * | 2020-04-29 | 2021-11-04 | Lightelligence, Inc. | Optical modulation for optoelectronic processing |
| US20210342730A1 (en) * | 2020-05-01 | 2021-11-04 | equal1.labs Inc. | System and method of quantum enhanced accelerated neural network training |
| CN114424213A (en) * | 2020-06-25 | 2022-04-29 | 普立恩科技有限公司 | Analog hardware implementation of neural networks |
| EP3940696A1 (en) * | 2020-07-15 | 2022-01-19 | Österreichische Akademie der Wissenschaften | Inpainting based signal gap reconstruction with sparse representation |
| CN113792870A (en) * | 2021-10-29 | 2021-12-14 | 广西大学 | All-optical control optical neural network system based on acousto-optic technology |
| CN114462587A (en) * | 2022-02-10 | 2022-05-10 | 电子科技大学 | FPGA implementation method for photoelectric hybrid computing neural network |
| CN114584441A (en) * | 2022-03-02 | 2022-06-03 | 中国人民解放军军事科学院战争研究院 | Digital signal modulation identification method based on deep learning |
| CN114358271A (en) * | 2022-03-18 | 2022-04-15 | 之江实验室 | Time-wavelength interweaving photon neural network convolution acceleration chip |
| CN115167610A (en) * | 2022-09-08 | 2022-10-11 | 之江实验室 | Two-dimensional photon convolution acceleration chip and application system and method |
Non-Patent Citations (6)
| Title |
|---|
| HUANG C等: "A silicon photonic– electronic neural network for fibre nonlinearity compensation", 《NATURE ELECTRONICS》 * |
| LUIZ AUGUSTO MELO PEREIRA等: "Amplified radio-over-fiber system linearization using recurrent neural networks", 《JOURNAL OF OPTICAL COMMUNICATIONS AND NETWORKING》 * |
| SHASTRI B J等: "Photonics for artificial intelligence and neuromorphic computing", 《NATURE PHOTONICS》 * |
| XU X等: "11 TOPS photonic convolutional accelerator for optical neural networks", 《NATURE》 * |
| 尹坤阳;潘伟;谢立东;徐素霞;: "一种基于深度学习的人体交互行为分层识别方法", 厦门大学学报(自然科学版), no. 03 * |
| 杨俊波;李修建;杨建坤;刘菊;苏显渝;徐平;: "自由空间可重构光互连网络的光学矩阵运算", 中国激光, no. 07 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116957031A (en) * | 2023-07-24 | 2023-10-27 | 浙江大学 | Photoelectric computer based on optical multi-neuron activation function module |
| CN116957031B (en) * | 2023-07-24 | 2024-05-24 | 浙江大学 | Photoelectric computer based on optical multi-neuron activation function module |
| CN117436494A (en) * | 2023-12-18 | 2024-01-23 | 合肥硅臻芯片技术有限公司 | Neural network operation chip and operation method |
| CN117436494B (en) * | 2023-12-18 | 2024-02-23 | 合肥硅臻芯片技术有限公司 | Neural network operation chip and operation method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116432726B (en) | 2023-08-25 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN116432726B (en) | Photoelectric hybrid deep neural network operation device and operation method | |
| US11373089B2 (en) | Serialized electro-optic neural network using optical weights encoding | |
| CN114819132B (en) | Photon two-dimensional convolution acceleration method and system based on time-wavelength interleaving | |
| CN114358271A (en) | Time-wavelength interweaving photon neural network convolution acceleration chip | |
| CN115169542B (en) | Two-dimensional photon convolution acceleration system and device for convolution neural network | |
| CN114815959B (en) | Photon tensor calculation acceleration method and device based on wavelength division multiplexing | |
| CN110351000B (en) | All-optical serial-parallel conversion system based on wavelength division multiplexing technology | |
| JP4700753B2 (en) | Multi-data rate compatible ultra-fast self-adaptive all-optical data packet rate doubling method | |
| CN117077750B (en) | Parallel photon convolution operation chip and system | |
| CN115167610B (en) | Two-dimensional photon convolution acceleration chip and application system and method | |
| CN115222035B (en) | Photon neural network convolution acceleration chip | |
| EP1562070A1 (en) | Optical digital-to-analog conversion | |
| CN111458953A (en) | Optical analog-to-digital conversion architecture based on photon parallel sampling and implementation method thereof | |
| CN117784313A (en) | Two-dimensional photon convolution operation chip and system based on cyclic array waveguide grating | |
| US11934943B1 (en) | Two-dimensional photonic neural network convolutional acceleration chip based on series connection structure | |
| CN116029350B (en) | Two-dimensional photon coherent convolution acceleration chip based on time interleaving and application system thereof | |
| US8089381B2 (en) | Photonic D/A | |
| CN113325917A (en) | Light computing device, system and computing method | |
| WO2006091351A1 (en) | Matrix time division multiplex (mtdm) interrogation | |
| CN116316007A (en) | Ultra-high-speed arbitrary waveform generator and generation method based on synthesized dimension | |
| CN115130666A (en) | Two-dimensional photon convolution acceleration method and system | |
| CN110995354B (en) | Optical assembly and optical module | |
| CN116484931B (en) | Photon matrix multiplication operation device and operation method for neural network | |
| CN117709423B (en) | Deep neural network photon acceleration chip and operation system thereof | |
| CN111245553A (en) | Forming photon-assisted optical serial-parallel conversion system and optical communication apparatus employing the same |
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 |