CN111680796A - On-chip photonic neural network device based on cascade optical coupler, chip and application method thereof - Google Patents

On-chip photonic neural network device based on cascade optical coupler, chip and application method thereof Download PDF

Info

Publication number
CN111680796A
CN111680796A CN202010605656.1A CN202010605656A CN111680796A CN 111680796 A CN111680796 A CN 111680796A CN 202010605656 A CN202010605656 A CN 202010605656A CN 111680796 A CN111680796 A CN 111680796A
Authority
CN
China
Prior art keywords
output
neural network
repeating unit
optical
optical coupler
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
Application number
CN202010605656.1A
Other languages
Chinese (zh)
Other versions
CN111680796B (en
Inventor
罗章
赖明澈
翦杰
王子聪
孙岩
黎渊
张建民
徐金波
董德尊
熊泽宇
欧洋
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
National University of Defense Technology
Original Assignee
National University of Defense Technology
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by National University of Defense Technology filed Critical National University of Defense Technology
Priority to CN202010605656.1A priority Critical patent/CN111680796B/en
Publication of CN111680796A publication Critical patent/CN111680796A/en
Application granted granted Critical
Publication of CN111680796B publication Critical patent/CN111680796B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

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
    • G06N3/08—Learning methods
    • 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
    • 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

Abstract

The invention discloses an on-chip photonic neural network device based on a cascade optical coupler, a chip and an application method thereof0The multi-layer neuron comprises M repeating units, each repeating unit comprises multiple amplitude modulation phase modulators and a MIMO optical coupler, and the multiple amplitude modulation phase modulators of any M-th repeating unit are used for respectively modulating Nm‑1After amplitude modulation and phase modulation are carried out on the optical signals input by the path, N is obtained through the MIMO optical couplermA channel output, the detection unit includes NMA path detector. The all-optical photonic neural network deep learning system can realize the large-bit-width input all-optical photonic neural network deep learning system in the on-chip integrated system on the basis of the cascade optical coupler, and has the advantages of extremely low calculation overhead and little work in the forward reasoning process after the training and learning are finishedThe consumption is low, and the system on chip is easy to integrate and supports large-bit-width data input.

Description

On-chip photonic neural network device based on cascade optical coupler, chip and application method thereof
Technical Field
The invention relates to a neural network deep learning technology, in particular to an on-chip photonic neural network device based on a cascade optical coupler, a chip and an application method thereof, which are used for realizing an all-optical photonic neural network deep learning system with large bit width input in an on-chip integrated system on the basis of the cascade optical coupler.
Background
Artificial intelligence is a new technology science for researching and developing intelligent theory, method technology and application system for simulating extension and extension people. On one hand, the development of algorithm, computing power and big data drives artificial intelligence from three dimensions to make breakthrough in various aspects of computing intelligence, perception intelligence and cognitive intelligence; on the other hand, artificial intelligence is rapidly fused with various industries, so that the artificial intelligence can help the traditional industries to upgrade and transform, improve quality and enhance efficiency, and bring up a brand new industrial wave in the global scope.
The mainstream technology of artificial intelligence today is deep neural networks for big data processing. The deep neural network is a computational network model which is proposed by the inspiration of the human brain signal processing process, and since the calculation model is proposed, the deep neural network is widely applied to the fields of computer vision, voice recognition, natural semantic understanding, man-machine game and the like, and the expression of machine learning in the related fields is greatly improved.
In the last decade, the computing power of the neural network accelerator is greatly improved through the continuous improvement of a computing architecture, various technical routes including a convolver structure, a multi-tree structure, a matrix multiplier structure and the like are continuously developed, and deep learning is promoted to obtain one achievement with milestone significance and another achievement. However, the current computing hardware of the electric domain computing system has a large gap with the brain in terms of power consumption and performance when processing the neural network. Taking an artificial intelligence program Alphago which is provided by Google and takes a neural network accelerator TPU with a matrix multiplier structure as hardware basis and is based on a deep learning algorithm as an example, the Alphago needs a computing system with the power consumption of 20MW and the volume of 30m3(ii) a The human brain running the same calculation task has the power consumption of only 20W and the volume of only 0.0015m3. Under the current electric domain computing system architecture, the hardware performance of the neural network cannot be fully utilized. Although the electric domain computing system has been improved in computing architecture according to the characteristics of the neural network, the electric domain computing system is essentially centered on data processing and computing, and a processor reads data from a memory, performs computing processing, and then writes the data back to the memory. The neural network algorithm is centered on data flow, and data operation of the neural network algorithm is completed in the neural network along with the data flow. The implementation of neural network algorithms based on data processing-centered hardware necessitates the meshing and storage of a large number of data streams and the movement to an arithmetic processing unit for arithmetic processing when necessary. On one hand, this requires massive storage, and on the other hand, the storage of data and the operational processing of data are separated from each other, which in turn results in the movement of massive data. It is this architectural conflict that causes the electrical domain computing hardware to behave inefficiently when running neural network algorithms.
While computer domain scholars continuously improve computing architecture to improve the computing efficiency of the neural network, optical domain scholars also continuously propose novel schemes to realize neural network computing. Photon computing research has also made great progress for many years, but is limited by the physical nature of zero resting mass of photons, and has not made exciting breakthrough in realizing optical caching. Without optical information storage, traditional digital logic computers cannot be finally realized based on photonic devices, and the development of photonic computing systems has been delayed due to the lack of optical storage. But the architecture of the neural network algorithm with data flow as the center can be well matched with the photon computing system. On the other hand, each neuron of the neural network algorithm can use the input data immediately without caching, so that the defect that the photon computing network has no caching function is not important any more.
In 2017, a photon calculation team of MIT in the united states has first proposed an all-optical photonic neural network, which is formed by cascaded Mach-Zehnder interference devices (MZI) with additional phase shift function, and can realize arbitrary linear transformation and output of 4 paths of input signals. The team completes linear transformation by the network and completes nonlinear transformation by a computer, and the two are matched to realize a first-level 4-channel neuron layer. By multiplexing the all-optical photonic neural network, a 4-layer neural network is realized, and the recognition accuracy of 4 vowels exceeds 90%. The first reported all-optical photonic neural network capable of working practically has poor realizability, on one hand, the all-optical photonic neural network does not realize complete neural network operation in an optical domain, but more seriously, the photonic network is based on an MZI device array, when the input bit width is high and the number of network layers is large, a large number of MZI devices are required to be used, the network scale is huge, and the process preparation difficulty is large. This is because the realization of the all-optical photonic neural network is a mechanical simulation of electronic logic computation, and does not make good use of photonic characteristics.
The us UCLA team in 2019 proposed a diffraction depth neural network based on a multilayer diffraction film structure. The design utilizes the expansibility, the locality and the coherence of a diffraction light field to complete the convolution linear transformation between two layers of neurons, utilizes the projection rate and the phase shift of a diffraction film as learning parameters, fully utilizes the physical characteristics of the light field, and has concise physical realization of the neurons and good realizability. The group successfully realizes the identification of the 28 pixel multiplied by 28 pixel picture by a convolution neural network formed by 5 layers of diffraction films, and the identification success rate is up to 93%. Meanwhile, the scheme is used as a complete optical system, external light source transmission imaging is used as input, and the optical neural network system almost does not consume power in the recognition process after learning is completed. The scheme promotes the all-optical photonic neural network to make a big step towards practical application, and attracts wide attention with excellent power consumption performance. However, the implementation principle is natural two-dimensional convolution-like calculation, and the method cannot be applied to a non-two-dimensional convolution neural network algorithm, and in practical application, the convolution neural network only accounts for less than 10% of the application of all neural networks. On the other hand, the scheme is based on the space optical principle, and is difficult to be integrated into a system on a chip in a dimensionality reduction mode. The chip is the mainstream way for realizing the miniaturization and integration of the photon computing system, and the on-chip integration of the all-optical photon neural network as a special photon computing system is also an important development direction of the all-optical photon neural network.
Disclosure of Invention
The technical problems to be solved by the invention are as follows: aiming at the problems in the prior art, the invention provides an on-chip photonic neural network device based on a cascade optical coupler, a chip and an application method thereof, and the invention can realize an all-optical photonic neural network deep learning system with large bit width input in an on-chip integrated system on the basis of the cascade optical coupler, has extremely low computation overhead and low power consumption in the forward reasoning process after training and learning are finished, and is easy to integrate and supports large bit width data input.
In order to solve the technical problems, the invention adopts the technical scheme that:
a photonic neural network device on a chip based on a cascade optical coupler comprises an all-optical photonic neural network, wherein the all-optical photonic neural network comprises an amplitude modulation unit, a multilayer neuron and a detection unit, and the amplitude modulation unit comprises N0The multi-layer neuron comprises M repeating units which are sequentially cascaded in a layered mode, each repeating unit comprises a plurality of amplitude modulation phase modulators and a multi-end input multi-end output optical coupler, and the amplitude modulation phase modulators of any mth repeating unit are used for respectively modulating Nm-1After amplitude modulation and phase modulation are carried out on the optical signals input by the path, N is obtained through a multi-end input multi-end output optical couplermAn output optical signal, saidThe detecting unit comprises N and M repeating unitsMThe paths output optical signals to the detectors in one-to-one correspondence.
Optionally, the mimo-mimo optical coupler is a star coupler or a multimode interference coupler.
Optionally, a coherent light source is connected to an input end of the amplitude modulation unit, and the coherent light source includes M output optical paths.
Optionally, the coherent light source includes a laser and a single-ended input multi-ended output optical coupler, an output end of the laser is connected to an input of the single-ended input multi-ended output optical coupler, and an output end of the single-ended input multi-ended output optical coupler includes N0An output optical path.
In addition, the invention also provides an all-optical photonic neural network chip based on the cascade optical coupler, which comprises a chip body, wherein the chip body comprises an amplitude modulation unit, a multilayer neuron and a detection unit, and the amplitude modulation unit comprises N0The multi-layer neuron comprises M repeating units which are sequentially cascaded in a layered mode, each repeating unit comprises a plurality of amplitude modulation phase modulators and a multi-end input multi-end output optical coupler, and the amplitude modulation phase modulators of any mth repeating unit are used for respectively modulating Nm-1After amplitude modulation and phase modulation are carried out on the optical signals input by the path, N is obtained through a multi-end input multi-end output optical couplermThe detection unit comprises N of the Mth repeating unitMThe paths output optical signals to the detectors in one-to-one correspondence.
Optionally, the mimo-mimo optical coupler is a star coupler or a multimode interference coupler.
In addition, the invention also provides an application method of the on-chip photonic neural network device based on the cascade optical coupler, and the step of carrying out classification reasoning by the photonic neural network device comprises the following steps:
1) splitting the laser output into N via a single-ended input multi-ended output optical coupler0All-coherent light source, and N0All the path coherent light sources are I0；
2) Will N0The path of all-coherent light source is loaded to an amplitude modulator in an amplitude modulation unit in an intensity modulation mode to carry out amplitude modulation to form optical input excitation, and the complex amplitude of the ith path of excitation
Figure BDA0002560997540000031
Is composed of
Figure BDA0002560997540000032
Wherein
Figure BDA0002560997540000033
Representing normalized ith input data with a maximum value of 1, N0Optical input excitation of the light path
Figure BDA0002560997540000034
3)N0Optical excitation of a(0)Respectively entering an amplitude modulation phase modulator of a first repeating unit in the multilayer neuron, performing amplitude and phase modulation, outputting to a multi-end input multi-end output optical coupler of the first repeating unit, and outputting N after coherent superposition through the multi-end input multi-end output optical coupler1The complex amplitude of the channel signal, i-th output signal is
Figure BDA0002560997540000035
N1Optical output is noted
Figure BDA0002560997540000041
Any m-th repeating unit that follows: n is a radical ofm-1Road light signal
Figure BDA0002560997540000042
The input amplitude modulation phase modulator is subjected to amplitude and phase modulation and then output to the multi-end input multi-end output optical coupler of the mth repeating unit, so that N is outputmRoad signal
Figure BDA0002560997540000043
Multiple-input multiple-output optical coupler of final Mth repeating unitN of outputMRoad light signal
Figure BDA0002560997540000044
Directly captured by the photodetector in the detection unit to output NMRoad light intensity
Figure BDA0002560997540000045
Optionally, the functional expression of the amplitude and phase modulation in step 3) is as follows:
Figure BDA0002560997540000046
in the above formula, the first and second carbon atoms are,
Figure BDA0002560997540000047
is the output of the k-th way of the mth repeating unit,
Figure BDA0002560997540000048
transformation matrix N for mth repeating unitm×Nm-1The k-th path of (a) is converted into the matrix elements of the i-th path, j is an imaginary unit,
Figure BDA0002560997540000049
is the output of the ith path of the (m-1) th repeating unit,
Figure BDA00025609975400000410
the amplitude attenuation factor of the ith path of the mth stage,
Figure BDA00025609975400000411
an additional phase for the ith path of the mth stage.
Optionally, the NMRoad light intensity
Figure BDA00025609975400000412
Arbitrary ith output light intensity IiThe function expression of (a) is as follows:
Figure BDA00025609975400000413
in the above formula, the first and second carbon atoms are,
Figure BDA00025609975400000414
is composed of
Figure BDA00025609975400000415
The amount of conjugation of (a) to (b),
Figure BDA00025609975400000416
is the output of the ith path of the mth repeating unit.
Optionally, before the step 1), a step of training the photonic neural network device is further included:
s1) randomly selecting n samples from the training set as a sample group, wherein each sample in the sample group
Figure BDA00025609975400000417
With corresponding label
Figure BDA00025609975400000418
Wherein xiAll of N representing sample x0The ith component of the data, where yiAll of N in total representing sample yMThe ith component in the data, each component y of the tag yiOnly the component value of the position corresponding to the classification set where the sample is located is 1, and the other components are 0;
s2) setting initial parameters for amplitude and phase modulation in the all-optical photonic neural network, inputting each sample in the sample group into the all-optical photonic neural network to output a reasoning result, comparing the reasoning result with a label y to obtain an accuracy rate evaluation, and finishing training and exiting if the accuracy rate meets the requirement; otherwise, skipping to execute the next step;
s3) output light intensity to all-optical photon neural network
Figure BDA00025609975400000419
Push button
Figure BDA00025609975400000420
After normalization, the method is recorded
Figure BDA00025609975400000421
Then
Figure BDA00025609975400000422
According to a preset back propagation deviation function
Figure BDA00025609975400000423
Is propagated backward from the M-th repeating unit, and during the backward propagation, the amplitude attenuation factor in the M-th repeating unit is adjusted according to a preset gradient
Figure BDA00025609975400000424
Additional phase
Figure BDA00025609975400000425
Jumping to perform step S1);
said back propagation deviation function
Figure BDA00025609975400000426
The function expression of the initial value of (a) is as follows:
Figure BDA0002560997540000051
in the above formula, the first and second carbon atoms are,
Figure BDA0002560997540000052
is an initial value of a back-propagation deviation function, siThe ith path of normalized result, y, representing the output light intensity of the all-optical photon neural networkiAll of N in total representing sample yMThe ith component, s, in the individual datakThe k path normalization result, y, of the output light intensity of the all-optical photon neural network is representedkAll of N in total representing sample yMThe kth component of the data, n representing the number of samples in the sample group,
Figure BDA0002560997540000053
represents the conjugate of the output of the ith path of the mth repeating unit;
said back propagation deviation function
Figure BDA0002560997540000054
The recursion of the back propagation is shown by the following equation:
Figure BDA0002560997540000055
in the above formula, the first and second carbon atoms are,
Figure BDA0002560997540000056
representing the back propagation bias function propagated to the m-1 st repeating unit,
Figure BDA0002560997540000057
the amplitude attenuation factor of the kth path of the mth stage,
Figure BDA0002560997540000058
is the additional phase of the kth path of the mth stage, j is an imaginary unit,
Figure BDA0002560997540000059
transformation matrix N for mth repeating unitm×Nm-1The ith path of (a) is converted to the conjugate of the matrix element of the kth path,
Figure BDA00025609975400000510
the amplitude attenuation factor of the ith path of the mth stage,
Figure BDA00025609975400000511
representing a back propagation bias function propagated to the mth repeating unit;
amplitude attenuation factor in mth repeating unit
Figure BDA00025609975400000512
The gradient function expression of (a) is as follows:
Figure BDA00025609975400000513
in the above formula, the first and second carbon atoms are,
Figure BDA00025609975400000514
represents the amplitude attenuation factor of the ith path of the mth stage
Figure BDA00025609975400000515
Re represents the operation of the real part,
Figure BDA00025609975400000516
represents the conjugate of the output of the ith path of the mth repeating unit,
Figure BDA00025609975400000517
representing a back propagation bias function propagated to the mth repeating unit;
additional phase in mth repeating unit
Figure BDA00025609975400000518
The gradient function expression of (a) is as follows:
Figure BDA00025609975400000519
in the above formula, the first and second carbon atoms are,
Figure BDA00025609975400000520
additional phase representing ith path of mth stage
Figure BDA00025609975400000521
And Im denotes the imaginary part taking operation,
Figure BDA00025609975400000522
represents the conjugate of the output of the ith path of the mth repeating unit,
Figure BDA00025609975400000523
representing a counter-propagating deviation propagating to the m-th repeating unitA function.
Compared with the prior art, the on-chip photonic neural network device based on the cascade optical coupler has the following advantages: the cascade-based optical coupler comprises an all-optical photonic neural network, wherein the all-optical photonic neural network comprises an amplitude modulation unit 1, a multilayer neuron 2 and a detection unit 3, and the amplitude modulation unit 1 comprises N0The multi-layer neuron 2 comprises M repeating units which are sequentially cascaded in a layered mode, each repeating unit comprises a plurality of amplitude modulation phase modulators and a multi-end input multi-end output optical coupler, and the amplitude modulation phase modulators of any mth repeating unit are used for respectively modulating Nm-1After amplitude modulation and phase modulation are carried out on the optical signals input by the path, N is obtained through a multi-end input multi-end output optical couplermThe detection unit 3 comprises N of the Mth repeating unitMThe invention can realize the all-optical photon neural network deep learning system with large bit width input in the on-chip integrated system based on the cascade optical coupler, has extremely low calculation cost and low power consumption in the forward reasoning process after the training and learning are finished, and the on-chip system is easy to integrate and supports the large bit width data input.
Drawings
Fig. 1 is a schematic structural diagram of a photonic neural network device according to an embodiment of the present invention.
Fig. 2 is a schematic diagram of an optical path of a single-ended input multi-ended output optical coupler according to an embodiment of the present invention.
Fig. 3 is a schematic diagram of an optical path of the mimo optical coupler according to an embodiment of the present invention.
Fig. 4 is a schematic structural diagram of a training principle of a photonic neural network device according to an embodiment of the present invention.
Detailed Description
As shown in fig. 1, this embodiment provides an on-chip photonic neural network device based on a cascade optical coupler, which includes an all-optical photonic neural network, where the all-optical photonic neural network includes an amplitude modulation unit 1, a multilayer neuron 2, and a detection unit 3, and the amplitude modulation unit 1 includes N0With one-to-one correspondence of input optical signalsThe multi-layer neuron 2 comprises M repeating units which are sequentially cascaded in a layered mode, each repeating unit comprises a plurality of amplitude modulation phase modulators and a multi-end input multi-end output optical coupler, and the amplitude modulation phase modulators of any mth repeating unit are used for respectively modulating Nm-1After amplitude modulation and phase modulation are carried out on the optical signals input by the path, N is obtained through a multi-end input multi-end output optical couplermThe detection unit 3 comprises N of the M repeating unitMThe paths output optical signals to the detectors in one-to-one correspondence.
The photonic neural network chip is formed by cascading on-chip optical couplers, and typical Multi-port on-chip optical couplers include, but are not limited to, Star couplers (Star Coupler, fig. 2, 3) and Multi-mode interference couplers (MMI Coupler). The low-linewidth high-power laser source realizes the generation of a multipath coherent light source by splitting a Single-Input-Multi-Output Coupler (SIMO Coupler, fig. 2) through a Single-Input-Multi-Output Coupler, and injects the multipath coherent light source into the all-optical photonic neural network. The multi-bit wide-to-be-classified data is modulated onto each optical signal by using an electrical injection or thermal control amplitude modulator, and the intensity of the optical signal is proportional to the normalized input data. In this embodiment, the multi-input multi-output optical coupler is a star coupler or a multi-mode interference type coupler.
In this embodiment, the input end of the amplitude modulation unit 1 is connected to a coherent light source 4, and the coherent light source 4 includes M output optical paths.
In this embodiment, the coherent light source 4 includes a laser 41 and a single-ended input multi-ended output optical coupler 42, the output end of the laser 41 is connected to the input of the single-ended input multi-ended output optical coupler 42, and the output end of the single-ended input multi-ended output optical coupler 42 includes N0An output optical path.
On the basis, the embodiment further provides an all-optical photonic neural network chip based on the cascade optical coupler, which includes a chip body, the chip body includes an amplitude modulation unit 1, a multilayer neuron 2 and a detection unit 3, the amplitude modulation unit 1 includes N0The multi-layer neuron 2 comprises M repeating units which are sequentially cascaded in a layered manner, wherein each repeating unit comprises a plurality of amplitude modulation phase modulators and a plurality of amplitude modulation phase modulatorsA multi-input multi-output optical coupler, and multiple amplitude modulation and phase modulation devices of any mth repeating unit for respectively modulating Nm-1After amplitude modulation and phase modulation are carried out on the optical signals input by the path, N is obtained through a multi-end input multi-end output optical couplermThe detection unit 3 comprises N of the M repeating unitMThe paths output optical signals to the detectors in one-to-one correspondence.
In summary, in this embodiment, a single photonic device is used to implement linear transformation between input and output signals of two adjacent layers of neurons in an on-chip integrated optical system by using the coherent coupling characteristic of a multi-port on-chip optical coupler, and the input signal of the optical coupler is subjected to amplitude and phase modulation to implement fine tuning of the linear transformation, thereby completing the basic unit function of the neuron; on the basis, a cascade multistage on-chip optical coupler constructs multilayer neurons to form an on-chip integrated all-optical photonic neural network; the single-stage iterative back propagation gradient descent learning algorithm compatible with the complex propagation function is designed based on a standard back propagation gradient descent algorithm framework, and efficient learning convergence of the photonic neural network is achieved. Referring to fig. 4, the photonic neural network system on chip based on the cascade optical coupler is composed of an all-optical photonic neural network chip hardware based on the cascade optical coupler and a learning feedback network software based on a back propagation gradient descent algorithm, which is matched with the all-optical photonic neural network chip hardware. Referring to fig. 4, modulated optical signal a(0)And then the amplitude and phase of the signal are modulated for the second time by a first-stage amplitude modulation phase modulator, and then the signal enters a Multi-Input-Multi-Output Coupler (MIMO Coupler, fig. 3) for coherent superposition. The amplitude modulation phase modulator modulates the amplitude and the phase of an optical signal simultaneously, the modulation depth is determined after learning of a learning feedback network, and the effect of the amplitude modulation phase modulator is to change the coherent coupling characteristic of the optical signal in an optical coupler and obtain different linear transformation outputs. The optical coupler is matched with an amplitude modulation phase modulator at the front end of the optical coupler to form a basic unit of a cascade optical coupler photon neural network, and a layer of neuron functions are realized. Output a of the optocoupler(1)The input of the next layer of neuron is passed through the first-stage amplitude modulation phase modulator and then fed into the next-stage optical coupler. Repeating the above steps to pass through a multi-stage optical couplerThen, the last stage optical coupler outputs optical signal a(M)The light is directly captured by a group of light detectors, and the classification corresponding to the channel with the maximum light intensity is the classification output of the photon neural network, so that the forward propagation reasoning process is completed.
In this embodiment, the step of performing classification inference through the photonic neural network device includes:
1) splitting the laser output into N via a single-ended input multi-ended output optical coupler0All-coherent light source, and N0All the path coherent light sources are I0；
2) Will N0The path of all-coherent light source is loaded to an amplitude modulator in an amplitude modulation unit 1 in an intensity modulation mode for amplitude modulation to form optical input excitation, and the complex amplitude of the ith path of excitation
Figure BDA0002560997540000071
Is composed of
Figure BDA0002560997540000072
Wherein
Figure BDA0002560997540000073
Representing normalized ith input data with a maximum value of 1, N0Optical input excitation of the light path
Figure BDA0002560997540000074
3)N0Optical excitation of a(0)Respectively enter an amplitude modulation phase modulator of a first repeating unit in the multilayer neuron 2 for amplitude and phase modulation, then are output to a multi-end input multi-end output optical coupler of the first repeating unit, and are output with N after coherent superposition through the multi-end input multi-end output optical coupler1The complex amplitude of the channel signal, i-th output signal is
Figure BDA0002560997540000075
N1Optical output is noted
Figure BDA0002560997540000076
Any m-th repeating unit that follows: n is a radical ofm-1Road light signal
Figure BDA0002560997540000077
The input amplitude modulation phase modulator is subjected to amplitude and phase modulation and then output to the multi-end input multi-end output optical coupler of the mth repeating unit, so that N is outputmRoad signal
Figure BDA0002560997540000081
Finally, N of the multi-end input and multi-end output optical coupler output of the Mth repeating unitMRoad light signal
Figure BDA0002560997540000082
Directly captured by the photo detector in the detection unit 3 to output NMRoad light intensity
Figure BDA0002560997540000083
In this embodiment, the functional expression of the amplitude and phase modulation in step 3) is shown as follows:
Figure BDA0002560997540000084
in the above formula, the first and second carbon atoms are,
Figure BDA0002560997540000085
is the output of the k-th way of the mth repeating unit,
Figure BDA0002560997540000086
transformation matrix N for mth repeating unitm×Nm-1The k-th path of (a) is converted into the matrix elements of the i-th path, j is an imaginary unit,
Figure BDA0002560997540000087
is the output of the ith path of the (m-1) th repeating unit,
Figure BDA0002560997540000088
the amplitude attenuation factor of the ith path of the mth stage,
Figure BDA0002560997540000089
an additional phase for the ith path of the mth stage.
In this embodiment, N isMRoad light intensity
Figure BDA00025609975400000810
Arbitrary ith output light intensity IiThe function expression of (a) is as follows:
Figure BDA00025609975400000811
in the above formula, the first and second carbon atoms are,
Figure BDA00025609975400000812
is composed of
Figure BDA00025609975400000813
The amount of conjugation of (a) to (b),
Figure BDA00025609975400000814
is the output of the ith path of the mth repeating unit.
When network training is carried out, the photon neural network chip outputs a light intensity normalization result to be input into the learning feedback network, and the amplitude modulation and phase modulation depth of the front end of each stage of optical coupler of the forward propagation network is adjusted through a gradient descent learning algorithm. In this embodiment, before step 1), the method further includes the step of training the photonic neural network device:
s1) randomly selecting n samples from the training set as a sample group, wherein each sample in the sample group
Figure BDA00025609975400000815
With corresponding label
Figure BDA00025609975400000816
Wherein xiAll of N representing sample x0The ith component of the data, where yiAll of N in total representing sample yMThe ith component in the individual data,each component y of label yiOnly the component value of the position corresponding to the classification set where the sample is located is 1, and the other components are 0;
s2) setting initial parameters for amplitude and phase modulation in the all-optical photonic neural network, inputting each sample in the sample group into the all-optical photonic neural network to output a reasoning result, comparing the reasoning result with a label y to obtain an accuracy rate evaluation, and finishing training and exiting if the accuracy rate meets the requirement; otherwise, skipping to execute the next step;
s3) output light intensity to all-optical photon neural network
Figure BDA00025609975400000817
Push button
Figure BDA00025609975400000818
After normalization, the method is recorded
Figure BDA00025609975400000819
Then
Figure BDA00025609975400000820
According to a preset back propagation deviation function
Figure BDA00025609975400000821
Is propagated backward from the M-th repeating unit, and during the backward propagation, the amplitude attenuation factor in the M-th repeating unit is adjusted according to a preset gradient
Figure BDA00025609975400000822
Additional phase
Figure BDA00025609975400000823
Jumping to perform step S1);
said back propagation deviation function
Figure BDA00025609975400000824
The function expression of the initial value of (a) is as follows:
Figure BDA0002560997540000091
in the above formula, the first and second carbon atoms are,
Figure BDA0002560997540000092
is an initial value of a back-propagation deviation function, siThe ith path of normalized result, y, representing the output light intensity of the all-optical photon neural networkiAll of N in total representing sample yMThe ith component, s, in the individual datakThe k path normalization result, y, of the output light intensity of the all-optical photon neural network is representedkAll of N in total representing sample yMThe kth component of the data, n representing the number of samples in the sample group,
Figure BDA0002560997540000093
represents the conjugate of the output of the ith path of the mth repeating unit;
said back propagation deviation function
Figure BDA0002560997540000094
The recursion of the back propagation is shown by the following equation:
Figure BDA0002560997540000095
in the above formula, the first and second carbon atoms are,
Figure BDA0002560997540000096
representing the back propagation bias function propagated to the m-1 st repeating unit,
Figure BDA0002560997540000097
the amplitude attenuation factor of the kth path of the mth stage,
Figure BDA0002560997540000098
is the additional phase of the kth path of the mth stage, j is an imaginary unit,
Figure BDA0002560997540000099
is the m-th repeat unitElement transformation matrix Nm×Nm-1The ith path of (a) is converted to the conjugate of the matrix element of the kth path,
Figure BDA00025609975400000910
the amplitude attenuation factor of the ith path of the mth stage,
Figure BDA00025609975400000911
representing a back propagation bias function propagated to the mth repeating unit;
amplitude attenuation factor in mth repeating unit
Figure BDA00025609975400000912
The gradient function expression of (a) is as follows:
Figure BDA00025609975400000913
in the above formula, the first and second carbon atoms are,
Figure BDA00025609975400000914
represents the amplitude attenuation factor of the ith path of the mth stage
Figure BDA00025609975400000915
Re represents the operation of the real part,
Figure BDA00025609975400000916
represents the conjugate of the output of the ith path of the mth repeating unit,
Figure BDA00025609975400000917
representing a back propagation bias function propagated to the mth repeating unit;
additional phase in mth repeating unit
Figure BDA00025609975400000918
The gradient function expression of (a) is as follows:
Figure BDA00025609975400000919
in the above formula, the first and second carbon atoms are,
Figure BDA00025609975400000920
additional phase representing ith path of mth stage
Figure BDA00025609975400000921
And Im denotes the imaginary part taking operation,
Figure BDA00025609975400000922
represents the conjugate of the output of the ith path of the mth repeating unit,
Figure BDA00025609975400000923
representing the back propagation bias function propagated to the mth repeating unit.
In this embodiment, the cost function C is defined as:
Figure BDA00025609975400000924
i.e. defining the cost function C as 1/4 of the average of the squared distance of the inference output s of all samples x within the set of samples and the corresponding label y.
Defining a complex variant bias function as:
Figure BDA0002560997540000101
the recursion of the back propagation of the deviation function and the initial values are as indicated before. Finally, the adjustment of each control parameter is completed by the calculated gradient according to a standard gradient descent algorithm, and the learning of the next group of parameters is carried out until convergence.
The photonic neural network device in the embodiment has the following advantages: 1. after training and learning are completed, the calculation overhead of the forward reasoning process is extremely low, and the power consumption is low; 2. the system on chip is easy to integrate; 3. large bit width data entry is supported.
As will be appreciated by one skilled in the art, embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein. The present application is directed to methods, apparatus (systems), and computer program products according to embodiments of the application wherein instructions, which execute via a flowchart and/or a processor of the computer program product, create means for implementing functions specified in the flowchart and/or block diagram block or blocks. These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
The above description is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments, and all technical solutions belonging to the idea of the present invention belong to the protection scope of the present invention. It should be noted that modifications and embellishments within the scope of the invention may occur to those skilled in the art without departing from the principle of the invention, and are considered to be within the scope of the invention.

Claims (10)

1. A photonic neural network device on chip based on cascade optical coupler is characterized by comprising an all-optical photonic neural networkThe all-optical photonic neural network comprises an amplitude modulation unit (1), a multilayer neuron (2) and a detection unit (3), wherein the amplitude modulation unit (1) comprises N0The multi-layer neuron (2) comprises M repeating units which are sequentially cascaded in a layered mode, each repeating unit comprises a plurality of amplitude modulation phase modulators and a multi-end input multi-end output optical coupler, and the amplitude modulation phase modulators of any mth repeating unit are used for respectively modulating N signalsm-1After amplitude modulation and phase modulation are carried out on the optical signals input by the path, N is obtained through a multi-end input multi-end output optical couplermThe detection unit (3) comprises N and M repeating unitsMThe paths output optical signals to the detectors in one-to-one correspondence.
2. The cascaded optical coupler-based on-chip photonic neural network device of claim 1, wherein the multi-input multi-output optical coupler is a star coupler or a multi-mode interference type coupler.
3. The cascade optical coupler-based on-chip photonic neural network device according to claim 1, wherein the coherent light source (4) is connected to the input end of the amplitude modulation unit (1), and the coherent light source (4) comprises M output optical paths.
4. The cascade optical coupler-based on-chip photonic neural network device according to claim 3, wherein the coherent light source (4) comprises a laser (41) and a single-ended input multi-ended output optical coupler (42), an output end of the laser (41) is connected to an input of the single-ended input multi-ended output optical coupler (42), and an output end of the single-ended input multi-ended output optical coupler (42) comprises N0An output optical path.
5. The all-optical photonic neural network chip based on the cascade optical coupler comprises a chip body and is characterized in that the chip body comprises an amplitude modulation unit (1), a multilayer neuron (2) and a detection unit (3), wherein the amplitude modulation unit (1) comprises N0Input lightThe multi-layer neuron (2) comprises M repeating units which are sequentially cascaded in a layered mode, each repeating unit comprises a plurality of amplitude modulation phase modulators and a multi-end input multi-end output optical coupler, and the amplitude modulation phase modulators of any mth repeating unit are used for respectively modulating Nm-1After amplitude modulation and phase modulation are carried out on the optical signals input by the path, N is obtained through a multi-end input multi-end output optical couplermThe detection unit (3) comprises N and M repeating unitsMThe paths output optical signals to the detectors in one-to-one correspondence.
6. The cascade optical coupler-based all-optical photonic neural network chip of claim 5, wherein the multi-input multi-output optical coupler is a star coupler or a multi-mode interference type coupler.
7. The application method of the cascade optical coupler-based on-chip photonic neural network device as claimed in any one of claims 1 to 4, wherein the step of performing classification inference through the photonic neural network device comprises:
1) splitting the laser output into N via a single-ended input multi-ended output optical coupler0All-coherent light source, and N0All the path coherent light sources are I0；
2) Will N0The path of all-coherent light source is loaded to an amplitude modulator in an amplitude modulation unit (1) in an intensity modulation mode for amplitude modulation to form optical input excitation, and the complex amplitude of the ith path of excitation
Figure FDA0002560997530000011
Is composed of
Figure FDA0002560997530000012
Wherein
Figure FDA0002560997530000013
Representing normalized ith input data with a maximum value of 1, N0Optical input excitation of the light path
Figure FDA0002560997530000021
3)N0Optical excitation of a(0)Respectively enter an amplitude modulation phase modulator of a first repeating unit in a multilayer neuron (2) for amplitude and phase modulation, then are output to a multi-end input multi-end output optical coupler of the first repeating unit, and are output N after coherent superposition through the multi-end input multi-end output optical coupler1The complex amplitude of the channel signal, i-th output signal is
Figure FDA0002560997530000022
N1Optical output is noted
Figure FDA0002560997530000023
Any m-th repeating unit that follows: n is a radical ofm-1Road light signal
Figure FDA0002560997530000024
The input amplitude modulation phase modulator is subjected to amplitude and phase modulation and then output to the multi-end input multi-end output optical coupler of the mth repeating unit, so that N is outputmRoad signal
Figure FDA0002560997530000025
Finally, N of the multi-end input and multi-end output optical coupler output of the Mth repeating unitMRoad light signal
Figure FDA0002560997530000026
Directly captured by the photodetector in the detection unit (3) to output NMRoad light intensity
Figure FDA0002560997530000027
8. The method for applying the cascade optical coupler-based on-chip photonic neural network device according to claim 7, wherein the function expression of the amplitude and phase modulation in step 3) is as follows:
Figure FDA0002560997530000028
in the above formula, the first and second carbon atoms are,
Figure FDA0002560997530000029
is the output of the k-th way of the mth repeating unit,
Figure FDA00025609975300000210
transformation matrix N for mth repeating unitm×Nm-1The k-th path of (a) is converted into the matrix elements of the i-th path, j is an imaginary unit,
Figure FDA00025609975300000211
is the output of the ith path of the (m-1) th repeating unit,
Figure FDA00025609975300000212
the amplitude attenuation factor of the ith path of the mth stage,
Figure FDA00025609975300000213
an additional phase for the ith path of the mth stage.
9. The method of claim 8, wherein N is the number of photonic neural network devices on a chip based on cascade optical couplersMRoad light intensity
Figure FDA00025609975300000214
Arbitrary ith output light intensity IiThe function expression of (a) is as follows:
Figure FDA00025609975300000215
in the above formula, the first and second carbon atoms are,
Figure FDA00025609975300000216
is composed of
Figure FDA00025609975300000217
The amount of conjugation of (a) to (b),
Figure FDA00025609975300000218
is the output of the ith path of the mth repeating unit.
10. The method for applying the cascade optical coupler-based on-chip photonic neural network device according to claim 9, wherein the step 1) is preceded by the step of training the photonic neural network device:
s1) randomly selecting n samples from the training set as a sample group, wherein each sample in the sample group
Figure FDA00025609975300000219
With corresponding label
Figure FDA00025609975300000220
Wherein xiAll of N representing sample x0The ith component of the data, where yiAll of N in total representing sample yMThe ith component in the data, each component y of the tag yiOnly the component value of the position corresponding to the classification set where the sample is located is 1, and the other components are 0;
s2) setting initial parameters for amplitude and phase modulation in the all-optical photonic neural network, inputting each sample in the sample group into the all-optical photonic neural network to output a reasoning result, comparing the reasoning result with a label y to obtain an accuracy rate evaluation, and finishing training and exiting if the accuracy rate meets the requirement; otherwise, skipping to execute the next step;
s3) output light intensity to all-optical photon neural network
Figure FDA0002560997530000031
Push button
Figure FDA0002560997530000032
After normalization, the method is recorded
Figure FDA0002560997530000033
Then
Figure FDA0002560997530000034
According to a preset back propagation deviation function
Figure FDA0002560997530000035
Is propagated backward from the M-th repeating unit, and during the backward propagation, the amplitude attenuation factor in the M-th repeating unit is adjusted according to a preset gradient
Figure FDA0002560997530000036
Additional phase
Figure FDA0002560997530000037
Jumping to perform step S1);
said back propagation deviation function
Figure FDA0002560997530000038
The function expression of the initial value of (a) is as follows:
Figure FDA0002560997530000039
in the above formula, the first and second carbon atoms are,
Figure FDA00025609975300000310
is an initial value of a back-propagation deviation function, siThe ith path of normalized result, y, representing the output light intensity of the all-optical photon neural networkiAll of N in total representing sample yMThe ith component, s, in the individual datakThe k path normalization result, y, of the output light intensity of the all-optical photon neural network is representedkAll of N in total representing sample yMData of a personThe k-th component of (a), n represents the number of samples in the sample group,
Figure FDA00025609975300000311
represents the conjugate of the output of the ith path of the mth repeating unit;
said back propagation deviation function
Figure FDA00025609975300000312
The recursion of the back propagation is shown by the following equation:
Figure FDA00025609975300000313
in the above formula, the first and second carbon atoms are,
Figure FDA00025609975300000314
representing the back propagation bias function propagated to the m-1 st repeating unit,
Figure FDA00025609975300000315
the amplitude attenuation factor of the kth path of the mth stage,
Figure FDA00025609975300000316
is the additional phase of the kth path of the mth stage, j is an imaginary unit,
Figure FDA00025609975300000317
transformation matrix N for mth repeating unitm×Nm-1The ith path of (a) is converted to the conjugate of the matrix element of the kth path,
Figure FDA00025609975300000318
the amplitude attenuation factor of the ith path of the mth stage,
Figure FDA00025609975300000319
representing a back propagation bias function propagated to the mth repeating unit;
m-th repeating unit middle vibrationAmplitude attenuation factor
Figure FDA00025609975300000320
The gradient function expression of (a) is as follows:
Figure FDA00025609975300000321
in the above formula, the first and second carbon atoms are,
Figure FDA00025609975300000322
represents the amplitude attenuation factor of the ith path of the mth stage
Figure FDA00025609975300000323
Re represents the operation of the real part,
Figure FDA00025609975300000324
represents the conjugate of the output of the ith path of the mth repeating unit,
Figure FDA00025609975300000325
representing a back propagation bias function propagated to the mth repeating unit;
additional phase in mth repeating unit
Figure FDA00025609975300000326
The gradient function expression of (a) is as follows:
Figure FDA00025609975300000327
in the above formula, the first and second carbon atoms are,
Figure FDA00025609975300000328
additional phase representing ith path of mth stage
Figure FDA00025609975300000329
And Im denotes the imaginary part taking operation,
Figure FDA00025609975300000330
represents the conjugate of the output of the ith path of the mth repeating unit,
Figure FDA0002560997530000041
representing the back propagation bias function propagated to the mth repeating unit.
CN202010605656.1A 2020-06-29 2020-06-29 On-chip photonic neural network device and chip based on cascade optical coupler and application method of device and chip Active CN111680796B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010605656.1A CN111680796B (en) 2020-06-29 2020-06-29 On-chip photonic neural network device and chip based on cascade optical coupler and application method of device and chip

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010605656.1A CN111680796B (en) 2020-06-29 2020-06-29 On-chip photonic neural network device and chip based on cascade optical coupler and application method of device and chip

Publications (2)

Publication Number Publication Date
CN111680796A true CN111680796A (en) 2020-09-18
CN111680796B CN111680796B (en) 2023-07-11

Family

ID=72456870

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010605656.1A Active CN111680796B (en) 2020-06-29 2020-06-29 On-chip photonic neural network device and chip based on cascade optical coupler and application method of device and chip

Country Status (1)

Country Link
CN (1) CN111680796B (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112988113A (en) * 2021-04-29 2021-06-18 中国科学院西安光学精密机械研究所 Photon matrix vector multiplier
CN113298246A (en) * 2021-05-27 2021-08-24 山东云海国创云计算装备产业创新中心有限公司 Data processing method, device and computer readable storage medium
CN117592533A (en) * 2023-10-30 2024-02-23 浙江大学 Method for realizing neural network nonlinear operation by using semiconductor laser

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110247728A (en) * 2019-07-01 2019-09-17 光子算数(北京)科技有限责任公司 A kind of photon neural network chip
US20190294199A1 (en) * 2016-06-02 2019-09-26 Massachusetts Institute Of Technology Apparatus and methods for optical neural network
CN110929864A (en) * 2019-12-05 2020-03-27 北京超放信息技术有限公司 Optical diffraction neural network on-line training method and system
US20200142441A1 (en) * 2018-11-02 2020-05-07 Lightmatter, Inc. Matrix multiplication using optical processing
CN111291881A (en) * 2020-03-05 2020-06-16 西南大学 Mode identification device and method based on cascade VCSELs photonic nerves

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190294199A1 (en) * 2016-06-02 2019-09-26 Massachusetts Institute Of Technology Apparatus and methods for optical neural network
US20200142441A1 (en) * 2018-11-02 2020-05-07 Lightmatter, Inc. Matrix multiplication using optical processing
CN110247728A (en) * 2019-07-01 2019-09-17 光子算数(北京)科技有限责任公司 A kind of photon neural network chip
CN110929864A (en) * 2019-12-05 2020-03-27 北京超放信息技术有限公司 Optical diffraction neural network on-line training method and system
CN111291881A (en) * 2020-03-05 2020-06-16 西南大学 Mode identification device and method based on cascade VCSELs photonic nerves

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
BIN SHI等: "Deep Neural Network Through an InP SOA-Based Photonic Integrated Cross-Connect" *
陈宏伟等: "光子神经网络发展与挑战" *

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112988113A (en) * 2021-04-29 2021-06-18 中国科学院西安光学精密机械研究所 Photon matrix vector multiplier
CN113298246A (en) * 2021-05-27 2021-08-24 山东云海国创云计算装备产业创新中心有限公司 Data processing method, device and computer readable storage medium
CN113298246B (en) * 2021-05-27 2023-02-28 山东云海国创云计算装备产业创新中心有限公司 Data processing method, device and computer readable storage medium
CN117592533A (en) * 2023-10-30 2024-02-23 浙江大学 Method for realizing neural network nonlinear operation by using semiconductor laser

Also Published As

Publication number Publication date
CN111680796B (en) 2023-07-11

Similar Documents

Publication Publication Date Title
Li et al. Class-specific differential detection in diffractive optical neural networks improves inference accuracy
De Marinis et al. Photonic neural networks: A survey
CN111680796B (en) On-chip photonic neural network device and chip based on cascade optical coupler and application method of device and chip
Sui et al. A review of optical neural networks
Cheng et al. Silicon photonics codesign for deep learning
Qu et al. Inverse design of an integrated-nanophotonics optical neural network
CN110929864B (en) Optical diffraction neural network on-line training method and system
WO2022134391A1 (en) Fusion neuron model, neural network structure and training and inference methods therefor, storage medium, and device
CN111478735B (en) General quantum walking simulation chip structure based on integrated optics
CN115545173A (en) Optical modulation neuron for signal processing and all-optical diffraction neural network method
US11600060B2 (en) Nonlinear all-optical deep-learning system and method with multistage space-frequency domain modulation
Gu et al. SqueezeLight: Towards scalable optical neural networks with multi-operand ring resonators
CN114970836B (en) Reservoir neural network implementation method and system, electronic device and storage medium
Zhang et al. Summary of convolutional neural network compression technology
Launay et al. Hardware beyond backpropagation: a photonic co-processor for direct feedback alignment
WO2023109065A1 (en) Method and system for implementing recurrent neural network, and electronic device and storage medium
Kirtas et al. Learning photonic neural network initialization for noise-aware end-to-end fiber transmission
CN115496195A (en) MZI array and FP-SA-based optical pulse neural network implementation device
CN114707629A (en) Matrix calculation method based on light scattering and optical neural network
Li et al. Orbital angular momentum logic gates based on optical diffraction neural network
Li et al. High-efficiency reinforcement learning with hybrid architecture photonic integrated circuit
Zhang et al. Time-stretch optical neural network with time-division multiplexing
Yao et al. Fast differentiable evolution of quantum states under gaussian transformations
Feng et al. Optical Neural Networks for Holographic Image Recognition
US20230316074A1 (en) Framework for simulating general multi-scale problems

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