CN114325932B - On-chip integrated all-optical neural network optical computing chip - Google Patents
On-chip integrated all-optical neural network optical computing chip Download PDFInfo
- Publication number
- CN114325932B CN114325932B CN202210102732.6A CN202210102732A CN114325932B CN 114325932 B CN114325932 B CN 114325932B CN 202210102732 A CN202210102732 A CN 202210102732A CN 114325932 B CN114325932 B CN 114325932B
- Authority
- CN
- China
- Prior art keywords
- module
- soa
- optical
- output end
- amplitude
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Landscapes
- Optical Modulation, Optical Deflection, Nonlinear Optics, Optical Demodulation, Optical Logic Elements (AREA)
Abstract
The invention discloses an on-chip integrated all-optical neural network optical computing chip, which comprises: a multi-amplitude pulse light source module comprising at least two lasers for generating a multi-amplitude light calculation input pulse signal; the cross gain modulation module comprises a Mach-Zehnder interferometer SOA-MZI module based on a semiconductor optical amplifier and a semiconductor optical amplifier SOA module, and is used for performing wavelength conversion on the light calculation input pulse signals with multiple amplitudes according to a nonlinear activation Sigmoid function based on cross gain modulation of the SOA-MZI module and the SOA module so as to output pulse light signals with balanced amplitudes; the threshold adjusting module is used for adjusting the SOA-MZI module and the saturation state of the SOA in the SOA module so as to adjust the threshold of the Sigmoid function; and the photonic integrated chip substrate is used as the chip substrate of the all-optical neural network optical computing chip. The invention adopts the all-optical monolithic integration technology to realize the neural network optical computation of a chip-level all-optical structure with high speed, low power consumption, low cost and high integration.
Description
Technical Field
The invention relates to the technical field of semiconductor photoelectron, in particular to an on-chip integrated all-optical neural network optical computing chip.
Background
With the rapid development of current information science, many important calculation tasks are not realized by software similar to brain neural network calculation in the fields of image processing, voice recognition, artificial intelligence technology, deep learning application and the like. The neural network calculation is a calculation method similar to human brain neurons, is formed by connecting a plurality of neurons with adjustable connection weights, and has the characteristics of large-scale parallel processing, distributed information storage, good self-organizing and self-learning capabilities and the like.
The traditional neural network computing is an electronic neural network which takes an Integrated Circuit (IC) chip as a carrier, is an 'electronic + logic' information processing mode built based on a von Neumann architecture, consists of two mutually independent storage units and a processor unit, and physically separates the core computing functions of memory and processing, so that a large amount of tidal data load is generated between the memory and the computing unit, the computing rate is reduced, and the single computing energy consumption is increased. In addition, the electronic neural network computation has the following two inherent defects: firstly, signals are easy to interfere with each other, and certain difficulty is brought to the calculation of a high-density connected neural network; secondly, the energy demand is too high, resulting in higher calculation cost. Limited by the inherent limits of electronics, it is difficult for conventional electronic neural networks to further increase computational rate and power efficiency.
In order to break through the technical limitation of the electronic neural network computing method, a neural network computing method of a photon technology is provided, and photons are used as a basic carrier for information processing and transmission, namely, neural network optical computing. The neural network optical calculation is a neural network calculation mode combined with an optoelectronic device, and has the functions of linear weighting and nonlinear activation on the device level, the linear weighting needs large-scale matrix calculation, and the nonlinear activation unit can realize nonlinear calculation which cannot be realized by the matrix calculation, so that the complete neural network calculation is realized. The optical network optical computation adopts optical interconnection to realize three-dimensional relation between optical logic elements and image information, and optical correlation memory is applied to realize image correlation, convolution, extraction, symbol substitution, matrix operation and high-density information storage. Compared with the electronic technology, the neural network optical computation has the advantages of large bandwidth, low loss, high transmission information content and the like, and fundamentally changes the structural system of modern electronic computation, so that the optical computation of a chip-level all-optical structure with high speed, low power consumption and high integration degree becomes possible.
Disclosure of Invention
The embodiment of the invention provides an on-chip integrated all-optical neural network optical computing chip, which solves the technical problems of low computing rate, high energy consumption, high cost and difficulty in high-density integration of the traditional electronic neural network.
The embodiment of the invention provides an on-chip integrated all-optical neural network optical computing chip, which comprises:
a multi-amplitude pulsed light source module comprising at least two lasers for generating a multi-amplitude light calculation input pulse signal;
the cross gain modulation module comprises a Mach-Zehnder interferometer SOA-MZI module based on a semiconductor optical amplifier and a semiconductor optical amplifier SOA module, and is used for performing wavelength conversion on the light calculation input pulse signals with multiple amplitudes according to a nonlinear activation Sigmoid function based on the cross gain modulation of the SOA-MZI module and the SOA module so as to output the pulse light signals with balanced amplitudes;
the threshold adjusting module is used for adjusting the SOA-MZI module and the saturation state of the SOA in the SOA module so as to adjust the threshold of the Sigmoid function;
and the photonic integrated chip substrate is used as a chip substrate of the all-optical neural network optical computing chip to realize the monolithic integration of the multi-amplitude pulse light source module, the cross gain modulation module and the threshold value adjusting module.
Optionally, in an embodiment of the present invention, the multi-value pulsed light source module further includes:
the input end of the phase modulation module is connected with the output ends of the at least two lasers, the phase modulation module comprises at least two phase modulators PM and is used for carrying out phase modulation on multi-wavelength high-speed pulse optical signals output by the at least two lasers corresponding to the at least two phase modulators to obtain pulse optical signals with at least two amplitudes;
and the input end of the MMI is connected with the output end of the phase modulation module and is used for coupling the pulse optical signals of at least two amplitudes to a single waveguide to obtain the multi-amplitude optical calculation input pulse signal.
Optionally, in an embodiment of the present invention, the multi-amplitude pulsed light source module includes at least two lasers, and is configured to generate a multi-amplitude light calculation input pulse signal, specifically:
varying the output power of the at least two lasers to adjust the weight of the at least two amplitudes of pulsed optical signals when coupled to a single waveguide in the MMI;
determining the multi-amplitude light calculation input pulse signal according to the weight.
Optionally, in an embodiment of the present invention, the SOA-MZI module includes: the first 1x2MMI, the first branch, the second branch and the first 2x1 MMI;
the input end of the first 1x2MMI is connected with the threshold value adjusting module, and the output end of the first 1x2MMI is connected with the input end of the first branch and the input end of the second branch;
the input end of the first 2x1 MMI is connected with the output end of the first branch circuit and the output end of the second branch circuit, and the output end of the 2x1 MMI is connected with the input end of the SOA module;
the first branch comprises a second 2x1 MMI, a first SOA, a first PM and a second 1x2MMI which are sequentially connected in series;
the second branch comprises a third 2x1 MMI, a second SOA, a second PM and a third 1x2MMI which are sequentially connected in series.
Optionally, in an embodiment of the present invention, the threshold adjusting module includes a first laser and a second laser;
the output end of the first laser is connected with the input end of the first 1x2 MMI;
and the output end of the second laser is connected with the input end of the second branch circuit.
Optionally, in an embodiment of the present invention, the cross-gain modulation module further includes:
the input end of the signal control module is connected with the output end of the multi-amplitude pulse light source module, the first output end of the signal control module is connected with the input end of the first branch circuit, and the second output end of the signal control module is connected with the output end of the second branch circuit; and the intensity of the input pulse signal entering the first branch and the second branch is calculated by adjusting the light with multiple amplitudes, and the curve slope and the range of the nonlinear region of the Sigmoid function are further adjusted.
Optionally, in an embodiment of the present invention, the signal control module includes: a third SOA, a fourth 1x2MMI and a fourth SOA;
the input end of the third SOA is connected with the output end of the multi-amplitude pulse light source module, and the output end of the third SOA is connected with the input end of the fourth 1x2MMI and used for adjusting the range size of the nonlinear region of the Sigmoid function;
a first output end of the fourth 1x2MMI is connected with an input end of the fourth SOA, and a second output end of the fourth 1x2MMI is connected with an output end of the second branch circuit;
and the output end of the fourth SOA is connected with the input end of the first branch circuit and is used for adjusting the curve slope of the nonlinear region of the Sigmoid function.
Optionally, in an embodiment of the present invention, the SOA module includes a filtering module, a fourth 2x1 MMI and a fifth SOA;
the input end of the filtering module is connected with the output end of the SOA-MZI module, the output end of the filtering module is connected with the input end of the fourth 2x1 MMI, and the output end of the fourth 2x1 MMI is connected with the input end of the fifth SOA.
Optionally, in an embodiment of the present invention, the threshold adjusting module further includes a third laser; and the output end of the third laser is connected with the input end of the fourth 2x1 MMI.
Optionally, in one embodiment of the invention, the photonic integrated chip substrate comprises a group III-V compound semiconductor material comprising any one of InP, gaAs, alAs, inGaAsP, inGaAlAs, and InGaAs.
Optionally, in an embodiment of the present invention, the cross-gain modulation module is configured to, when performing wavelength conversion on the multiple-amplitude optical computation input pulse signal according to a nonlinear activation Sigmoid function to output a pulsed optical signal with equalized amplitude, specifically configured to:
comparing the amplitudes of the multi-amplitude light calculation input pulse signal with a threshold value of the Sigmoid function;
if the amplitude is smaller than the threshold value, the cross gain modulation module outputs a pulse optical signal with balanced low amplitude;
and if the amplitude is not smaller than the threshold value, the cross gain modulation module outputs a pulse optical signal with balanced high amplitude.
Optionally, in one embodiment of the invention, the photonic integrated chip substrate comprises a III-V compound semiconductor material; the group III-V compound semiconductor material includes any one of InP, gaAs, alAs, inGaAsP, inGaAlAs, and InGaAs.
To sum up, the chip provided by the embodiment of the present invention includes a multi-amplitude pulse light source module, which includes at least two lasers, for generating a multi-amplitude light calculation input pulse signal; the cross gain modulation module comprises a Mach-Zehnder interferometer SOA-MZI module based on a semiconductor optical amplifier and a semiconductor optical amplifier SOA module, and is used for performing wavelength conversion on the multi-amplitude optical calculation input pulse signal according to a nonlinear activation Sigmoid function based on cross gain modulation of the SOA-MZI module and the SOA module so as to output a pulse optical signal with balanced amplitude; the threshold adjusting module is used for adjusting the SOA-MZI module and the saturation state of the SOA in the SOA module so as to adjust the threshold of the Sigmoid function; the device comprises a photonic integrated chip substrate so as to realize the monolithic integration of the multi-amplitude pulse light source module, the cross gain modulation module and the threshold value adjusting module. The invention can realize the all-optical on-chip integration of the optical calculation activation function of the neural network, has high integration level and system integrity and low power consumption, and is beneficial to realizing large-scale neural network calculation.
Additional aspects and advantages of the invention will be set forth in part in the description which follows and, in part, will be obvious from the description, or may be learned by practice of the invention.
Drawings
The foregoing and/or additional aspects and advantages of the present invention will become apparent and readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings of which:
fig. 1 is a schematic structural diagram of an on-chip integrated all-optical neural network optical computing chip according to an embodiment of the present invention;
fig. 2 is a schematic structural diagram of an SOA-MZI module provided in an embodiment of the present invention;
fig. 3 is a schematic structural diagram of an all-optical neural network optical computing chip provided in an embodiment of the present invention.
Detailed Description
Reference will now be made in detail to embodiments of the present invention, examples of which are illustrated in the accompanying drawings, wherein like or similar reference numerals refer to the same or similar elements or elements having the same or similar function throughout. The embodiments described below with reference to the accompanying drawings are illustrative only for the purpose of explaining the present invention, and are not to be construed as limiting the present invention. On the contrary, the embodiments of the invention include all changes, modifications and equivalents coming within the spirit and terms of the claims appended hereto.
According to some embodiments, in the related art, the neural network optical computation has the following three schemes:
the first scheme is an optical-electrical combined neural network optical calculation scheme; linear and nonlinear calculation is completed through conversion of optical-electrical-optical signals by optical-electrical combined neurons;
the second scheme is a silicon-based all-optical neural network optical calculation scheme;
and the third scheme is a III-V compound semiconductor material-based all-optical neural network optical calculation scheme.
The second scheme and the third scheme both belong to all-optical neural network optical computing technologies, and specifically, the neural network optical computing process is completed through optical signals.
In some embodiments, the optical-electrical combined neural network optical calculation scheme has high loss and low integration level, and is not beneficial to realizing large-scale calculation; the silicon-based all-optical neural network optical calculation scheme cannot be monolithically integrated with active devices such as a semiconductor laser and a semiconductor optical amplifier because the silicon-based material belongs to an indirect bandgap semiconductor, and the integration level is low; the common InP-based all-optical neural network optical calculation scheme of the III-V group compound semiconductor material only performs activation function unit verification on the level of discrete devices, and has certain defects in integration level and system integrity.
The present invention will be described in detail with reference to specific examples.
Fig. 1 is a schematic structural diagram of an on-chip integrated all-optical neural network optical computing chip according to an embodiment of the present invention.
As shown in fig. 1, an on-chip integrated all-optical neural network optical computing chip provided in an embodiment of the present invention includes:
a multi-amplitude pulsed light source module 101 comprising at least two lasers for generating a multi-amplitude light calculation input pulse signal;
the cross gain modulation module 102 comprises a Mach-Zehnder interferometer SOA-MZI module based on a semiconductor optical amplifier and a semiconductor optical amplifier SOA module, and is used for performing wavelength conversion on the light calculation input pulse signals with multiple amplitudes according to a nonlinear activation Sigmoid function based on the cross gain modulation of the SOA-MZI module and the SOA module so as to output the pulse light signals with balanced amplitudes;
the threshold adjusting module 103 is used for adjusting the SOA-MZI module and the saturation state of the SOA in the SOA module so as to adjust the threshold of the Sigmoid function;
and the photonic integrated chip substrate 104 is used as a chip substrate of the all-optical neural network optical computing chip to realize monolithic integration of the multi-amplitude pulse light source module, the cross gain modulation module and the threshold adjusting module.
According to some embodiments, the laser array formed by at least two lasers in the multi-amplitude pulsed light source module 101 is a laser array with a wavelength difference of preset nanometers, and a multi-wavelength high-speed pulsed light signal can be generated through high-speed modulation. The laser is not specifically referred to as a fixed laser. The laser may be, for example, a semiconductor laser. Including but not limited to distributed feedback semiconductor lasers, distributed bragg grating semiconductor lasers, and the like.
In the embodiment of the present invention, the multi-amplitude pulse light source module 101 further includes:
the phase modulation module is connected with the input ends of the at least two lasers, comprises at least two phase modulators PM and is used for carrying out phase modulation on multi-wavelength high-speed pulse optical signals output by the at least two lasers corresponding to the at least two phase modulators to obtain pulse optical signals with at least two amplitudes;
the input end of the MMI is connected with the output end of the phase modulation module and is used for coupling the pulse optical signals with at least two amplitudes to the single waveguide to obtain the optical calculation input pulse signals with multiple amplitudes.
According to some embodiments, the phase modulators in the phase modulation module correspond to the lasers in the multi-amplitude pulsed light source module one to one, and at least two paths of optical pulse signals output by at least two lasers may be superimposed on a time domain, so as to obtain at least two amplitude pulsed optical signals.
In the embodiment of the present invention, the multi-amplitude pulse light source module 101 includes at least two lasers, and is configured to generate a multi-amplitude light calculation input pulse signal, specifically:
changing the output power of at least two lasers to adjust the weight of at least two amplitudes of pulsed optical signals when coupled to a single waveguide in the MMI;
determining the multiple amplitude light calculation input pulse signal according to the weight.
According to some embodiments, the resulting multi-amplitude light calculation input pulse signal is a linearly weighted multi-amplitude light calculation input pulse signal by adjusting the weights of at least two amplitudes of the pulsed light signal when coupled onto a single waveguide in the MMI.
In the embodiment of the present invention, fig. 2 is a schematic structural diagram of an SOA-MZI module provided in the embodiment of the present invention. As shown in fig. 2, the SOA-MZI module includes: the first 1x2MMI, the first branch, the second branch and the first 2x1 MMI;
the input end of the first 1x2MMI is connected with the threshold adjusting module, and the output end of the first 1x2MMI is connected with the input end of the first branch circuit and the input end of the second branch circuit;
the input end of the first 2x1 MMI is connected with the output end of the first branch circuit and the output end of the second branch circuit, and the output end of the 2x1 MMI is connected with the input end of the SOA module;
the first branch comprises a second 2x1 MMI, a first SOA, a first PM and a second 1x2MMI which are sequentially connected in series;
the second branch comprises a third 2x1 MMI, a second SOA, a second PM and a third 1x2MMI which are sequentially connected in series.
According to some embodiments, the 1x2MMI may split a single waveguide's optical signal onto both waveguides; the 2x1 MMI may couple optical signals from two waveguides to a single waveguide.
According to some embodiments, the first PM is for phase modulating the output light of the first SOA; the second PM is used for performing phase modulation on output light of the second SOA; thereby realizing the phase control of the cross gain modulation module;
in the embodiment of the present invention, the threshold adjusting module 103 includes a first laser and a second laser;
the output end of the first laser is connected with the input end of the first 1x2 MMI;
the output end of the second laser is connected with the input end of the second branch.
In this embodiment of the present invention, the cross gain modulation module 102 further includes:
the input end of the signal control module is connected with the output end of the multi-amplitude pulse light source module, the first output end of the signal control module is connected with the input end of the first branch circuit, and the second output end of the signal control module is connected with the output end of the second branch circuit; and the light for adjusting the multiple amplitudes calculates the intensity of the input pulse signal entering the first branch and the second branch, and further adjusts the curve slope and the range of the nonlinear region of the Sigmoid function.
In an embodiment of the present invention, the signal control module includes: a third SOA, a fourth 1x2MMI and a fourth SOA;
the input end of the third SOA is connected with the output end of the multi-amplitude pulse light source module, and the output end of the third SOA is connected with the input end of the fourth 1x2MMI and used for adjusting the range of the nonlinear region of the Sigmoid function;
a first output end of the fourth 1x2MMI is connected with an input end of the fourth SOA, and a second output end of the fourth 1x2MMI is connected with an output end of the second branch circuit;
and the output end of the fourth SOA is connected with the input end of the first branch circuit and is used for adjusting the curve slope of the nonlinear region of the Sigmoid function.
In the embodiment of the invention, the SOA module comprises a filtering module, a fourth 2x1 MMI and a fifth SOA;
the input end of the filtering module is connected with the output end of the SOA-MZI module, the output end of the filtering module is connected with the input end of the fourth 2x1 MMI, and the output end of the fourth 2x1 MMI is connected with the input end of the fifth SOA.
According to some embodiments, the first SOA, the second SOA, the third SOA, the fourth SOA and the fifth SOA are all used for amplification of the optical signal; the first SOA, the second SOA and the fifth SOA are specifically used for differential gain modulation; the third SOA is specifically configured to adjust a range size of a nonlinear region of the Sigmoid function; the fourth SOA is specifically used to adjust the slope of the curve of the nonlinear region of the Sigmoid function.
In the embodiment of the present invention, the threshold adjusting module 103 further includes a third laser; the output end of the third laser is connected with the input end of the fourth 2x1 MMI.
According to some embodiments, the first laser, the second laser, and the third laser are all configured to generate continuous light injection into the corresponding semiconductor optical amplifier, and changing the optical power of the first laser, the second laser, and the third laser can adjust the saturation state of the corresponding semiconductor optical amplifier, thereby adjusting the threshold of the Sigmoid function.
According to some embodiments, the filtering module is configured to filter and extract a pulse light signal with a first laser wavelength in output light signals of the SOA-MZI module, and is configured to filter and extract a pulse light signal with a third laser wavelength in output light signals of a fifth SOA in the SOA module, so that the all-optical neural network optical computing chip completes a function of a Sigmoid function. The filter module is not specifically referred to a fixed filter module, and for example, the filter module may be an on-chip filter, such as an Arrayed Waveguide Grating (AWG).
In the embodiment of the invention, III-V group semiconductor materials are used as the chip substrate of the all-optical neural network optical computing chip to realize the monolithic integration 103 of the multi-amplitude pulse light source module 101, the cross gain modulation module 102 and the threshold value adjusting module, so that the integration level is improved, and the chip cost is reduced.
According to some embodiments, the III-V semiconductor material may be any one of InP, gaAs, alAs, inGaAsP, inGaAlAs, and InGaAs, that is, the present invention may use any one of III-V compound semiconductor materials as a chip substrate, and the specific choice of which may be determined according to practical situations, which is not limited herein.
In an embodiment of the present invention, the cross gain modulation module is configured to perform wavelength conversion on a light calculation input pulse signal with multiple amplitudes according to a nonlinear activation Sigmoid function, so as to output a pulse light signal with an equalized amplitude, and is specifically configured to:
comparing the light with multiple amplitudes to calculate the amplitude of the input pulse signal and the threshold value of the Sigmoid function;
if the amplitude is smaller than the threshold value, the cross gain modulation module outputs a pulse optical signal with balanced low amplitude;
if the amplitude is not smaller than the threshold value, the cross gain modulation module outputs a pulse optical signal with balanced high amplitude.
According to some embodiments, according to a cross-gain modulation principle, by using the deep saturation characteristic of the SOA, when the amplitude of the light calculation input pulse signal input by the SOA-MZI module is not less than the threshold value of the Sigmoid function, the first SOA and the second SOA in the SOA-MZI module are in a deep saturation state, so that a pulse light signal with balanced low amplitude is output; and when the amplitude of the input light calculation input pulse signal is smaller than the threshold value of the Sigmoid function, outputting an unbalanced high-amplitude pulse light signal. The initial saturation states of the first SOA and the second SOA can be adjusted by changing the light injection sizes of the first SOA and the second SOA in the SOA-MZI module through the first laser and the second laser, and therefore the threshold value of the Sigmoid function is adjusted. And the third SOA and the fourth SOA are used for changing the light with multiple amplitudes to calculate the power of the input pulse signal entering the first branch and the second branch of the SOA-MZI, so that the bias state of the SOA-MZI can be changed, and the curve slope and the range of the nonlinear region of the Sigmoid function can be adjusted.
In some embodiments, the initial state of the fifth SOA is set to be close to a deep saturation state. And inputting the pulse optical signal modulated by the SOA-MZI module into the SOA module for cross gain modulation again. When the pulse optical signal modulated by the SOA-MZI module is a balanced low-amplitude signal, the SOA outputs a balanced high-amplitude signal; when the pulse optical signal modulated by the SOA-MZI module is an unbalanced high-amplitude signal, the SOA is in a deep saturation state, and therefore a balanced low-amplitude pulse optical signal is output.
It is easy to understand that by combining the SOA-MZI module and the SOA module, the function of the Sigmoid function can be realized. That is, by comparing the amplitude of the input pulse signal calculated by the light with the threshold value of the Sigmoid function, a pulse light signal of 0 or 1 of the equalized amplitude can be output. If the amplitude is smaller than the threshold value, the cross gain modulation module outputs a pulse optical signal '0' with balanced high amplitude; if the amplitude is not smaller than the threshold value, the cross gain modulation module outputs a pulse optical signal '1' with balanced low amplitude.
By way of example in a scenario, fig. 3 is a schematic structural diagram of an all-optical neural network optical computing chip according to an embodiment of the present invention. As shown in fig. 3, the all-optical neural network optical computing chip adopts an indium phosphide InP material as a photonic integrated chip substrate, wherein LD-1, LD-2, LD-3, and LD-4 are semiconductor laser arrays with a wavelength difference of a preset nanometer;
PM-1, PM-2, PM-3 and PM-4 are phase modulators, respectively perform phase modulation on LD-1, LD-2, LD-3 and LD-4, superpose four paths of optical pulse signals on a time domain, and generate pulse optical signals with four amplitudes;
the 4x1 MMI is a 4x1 multi-mode interference coupler and couples four paths of optical signals to a single waveguide;
SOA-1, SOA-2, SOA-3, SOA-4 and SOA-5 are semiconductor optical amplifiers;
LD-5, LD-6, LD-7 are semiconductor lasers, produce the continuous light injection to the semiconductor light amplifier;
PM-5 and PM-6 are phase modulators, carry on the phase modulation to the output light of SOA-3, SOA-4 separately;
the Filter is an on-chip Filter and is used for filtering and extracting the pulse optical signal with the LD-5 wavelength in the output optical signal of the SOA-MZI structure and filtering and extracting the pulse optical signal with the LD-7 wavelength in the output optical signal of the SOA-5.
To sum up, the chip provided by the embodiment of the present invention includes a multi-amplitude pulse light source module, which includes at least two lasers, for generating a multi-amplitude light calculation input pulse signal; the cross gain modulation module comprises a Mach-Zehnder interferometer SOA-MZI module based on a semiconductor optical amplifier and a semiconductor optical amplifier SOA module, is used for performing cross gain modulation based on the SOA-MZI module and the SOA module, and performs wavelength conversion on the light calculation input pulse signals with multiple amplitudes according to a nonlinear activation Sigmoid function so as to output the pulse light signals with balanced amplitudes; the threshold adjusting module is used for adjusting the SOA-MZI module and the saturation state of the SOA in the SOA module so as to adjust the threshold of the Sigmoid function; and the photonic integrated chip substrate is used as the chip substrate of the all-optical neural network optical computing chip to realize the monolithic integration of the multi-amplitude pulse light source module, the cross gain modulation module and the threshold value adjusting module. According to the invention, through an all-optical scheme, the power loss caused by optical-electrical-optical conversion is reduced, and large-scale neural network calculation is facilitated; by realizing the all-optical on-chip integration of the neural network optical computation Sigmoid activation function, the defects of the all-optical neural network optical computation scheme in the related technology on the integration level and the system integrity can be overcome.
It should be noted that the terms "first," "second," and the like in the description of the present invention are used for descriptive purposes only and are not to be construed as indicating or implying relative importance. In addition, in the description of the present invention, "a plurality" means two or more unless otherwise specified.
Any process or method descriptions in flow charts or otherwise described herein may be understood as representing modules, segments, or portions of code which include one or more executable instructions for implementing specific logical functions or steps in the process, and alternate implementations are included within the scope of the preferred embodiment of the present invention in which functions may be executed out of order from that shown or discussed, including substantially concurrently or in reverse order, depending on the functionality involved, as would be understood by those reasonably skilled in the art of the present invention.
It should be understood that portions of the present invention may be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, various steps or methods may be implemented in software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, any one or combination of the following technologies, which are well known in the art, may be used: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application specific integrated circuit having an appropriate combinational logic gate circuit, a Programmable Gate Array (PGA), a Field Programmable Gate Array (FPGA), or the like.
It will be understood by those skilled in the art that all or part of the steps carried out in the method of implementing the above embodiments may be implemented by hardware related to instructions of a program, which may be stored in a computer readable storage medium, and the program, when executed, includes one or a combination of the steps of the method embodiments.
In addition, functional units in the embodiments of the present invention may be integrated into one processing module, or each unit may exist alone physically, or two or more units are integrated into one module. The integrated module can be realized in a hardware mode, and can also be realized in a software functional module mode. The integrated module, if implemented in the form of a software functional module and sold or used as a stand-alone product, may also be stored in a computer readable storage medium.
The storage medium mentioned above may be a read-only memory, a magnetic or optical disk, etc.
In the description herein, references to the description of the term "one embodiment," "some embodiments," "an example," "a specific example," or "some examples," etc., mean that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the invention. In this specification, the schematic representations of the terms used above do not necessarily refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.
Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention, and that variations, modifications, substitutions and alterations can be made to the above embodiments by those of ordinary skill in the art within the scope of the present invention.
Claims (7)
1. An on-chip integrated all-optical neural network optical computing chip is characterized by comprising:
the multi-amplitude pulse light source module comprises a phase modulation module, a multi-mode interference coupler MMI and at least two lasers, wherein the input end of the phase modulation module is connected with the output ends of the at least two lasers, and the output end of the phase modulation module is connected with the input end of the MMI;
the cross gain modulation module comprises a Mach-Zehnder interferometer SOA-MZI module based on a semiconductor optical amplifier and a semiconductor optical amplifier SOA module, and is used for performing wavelength conversion on the light calculation input pulse signals with multiple amplitudes according to a nonlinear activation Sigmoid function based on the cross gain modulation of the SOA-MZI module and the SOA module so as to output the pulse light signals with balanced amplitudes, wherein the SOA-MZI module comprises a first 1x2MMI, a first branch, a second branch and a first 2x1 MMI, the input end of the first 1x2MMI is connected with a threshold value adjusting module, the output end of the first 1x2MMI is connected with the input end of the first branch and the input end of the second branch, the input end of the first 2x1 MMI is connected with the output end of the first branch and the output end of the second branch, the output end of the 2x1 MMI is connected with the input end of the SOA module, the first branch comprises a second 2xMMI, a first SOA, a first PM and a second 1x2MMI which are sequentially connected in series, and the second branch comprises a third 2x1 MMI, a second SOA, a second PM and a third 1x2MMI which are sequentially connected in series;
the threshold adjusting module is configured to adjust saturation states of SOAs in the SOA-MZI module and the SOA module to adjust a threshold of the Sigmoid function, where the threshold adjusting module includes a first laser and a second laser, an output end of the first laser is connected to an input end of the first 1x2MMI, and an output end of the second laser is connected to an input end of the second branch;
and the photonic integrated chip substrate is used as a chip substrate of the all-optical neural network optical computing chip to realize the monolithic integration of the multi-amplitude pulse light source module, the cross gain modulation module and the threshold value adjusting module.
2. The on-chip integrated all-optical neural network optical computing chip of claim 1, wherein the multi-amplitude pulse light source module is further configured to:
varying the output power of the at least two lasers to adjust the weight of the at least two amplitudes of pulsed optical signals when coupled to a single waveguide in the MMI;
determining the multiple amplitude light calculation input pulse signal according to the weight.
3. The on-chip integrated all-optical neural network optical computing chip of claim 1, wherein the cross-gain modulation module further comprises:
the input end of the signal control module is connected with the output end of the multi-amplitude pulse light source module, the first output end of the signal control module is connected with the input end of the first branch circuit, and the second output end of the signal control module is connected with the output end of the second branch circuit; and the intensity of the input pulse signal entering the first branch and the second branch is calculated by adjusting the light with multiple amplitudes, and the curve slope and the range of the nonlinear region of the Sigmoid function are further adjusted.
4. The on-chip integrated all-optical neural network optical computing chip of claim 3, wherein the signal control module comprises: a third SOA, a fourth 1x2MMI and a fourth SOA;
the input end of the third SOA is connected with the output end of the multi-amplitude pulse light source module, and the output end of the third SOA is connected with the input end of the fourth 1x2MMI and used for adjusting the range size of the nonlinear region of the Sigmoid function;
a first output end of the fourth 1x2MMI is connected with an input end of the fourth SOA, and a second output end of the fourth 1x2MMI is connected with an output end of the second branch circuit;
and the output end of the fourth SOA is connected with the input end of the first branch circuit and is used for adjusting the curve slope of the nonlinear region of the Sigmoid function.
5. The on-chip integrated all-optical neural network optical computing chip of claim 1, wherein the SOA module comprises a filtering module, a fourth 2x1 MMI and a fifth SOA;
the input end of the filtering module is connected with the output end of the SOA-MZI module, the output end of the filtering module is connected with the input end of the fourth 2x1 MMI, and the output end of the fourth 2x1 MMI is connected with the input end of the fifth SOA;
the threshold adjustment module further comprises a third laser; and the output end of the third laser is connected with the input end of the fourth 2x1 MMI.
6. The on-chip integrated all-optical neural network optical computing chip of claim 1, wherein said photonic integrated chip substrate comprises a III-V compound semiconductor material comprising any of InP, gaAs, alAs, inGaAsP, inGaAlAs, and InGaAs.
7. The on-chip integrated all-optical neural network optical computing chip according to claim 1, wherein the cross gain modulation module is configured to perform wavelength conversion on the multiple-amplitude optical computing input pulse signal according to a nonlinear activation Sigmoid function, so as to output a pulse optical signal with equalized amplitude, and is specifically configured to:
comparing the amplitudes of the multi-amplitude light calculation input pulse signal with a threshold value of the Sigmoid function;
if the amplitude is smaller than the threshold value, the cross gain modulation module outputs a pulse optical signal with balanced low amplitude;
and if the amplitude is not smaller than the threshold value, the cross gain modulation module outputs a pulse optical signal with balanced high amplitude.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210102732.6A CN114325932B (en) | 2022-01-27 | 2022-01-27 | On-chip integrated all-optical neural network optical computing chip |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210102732.6A CN114325932B (en) | 2022-01-27 | 2022-01-27 | On-chip integrated all-optical neural network optical computing chip |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114325932A CN114325932A (en) | 2022-04-12 |
| CN114325932B true CN114325932B (en) | 2022-12-30 |
Family
ID=81030997
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210102732.6A Active CN114325932B (en) | 2022-01-27 | 2022-01-27 | On-chip integrated all-optical neural network optical computing chip |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114325932B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115472603B (en) * | 2022-08-30 | 2023-08-08 | 浙江大学 | Monolithic photon integrated chip for realizing convolutional neural network based on optical frequency comb |
| CN116957031B (en) * | 2023-07-24 | 2024-05-24 | 浙江大学 | Photoelectric computer based on optical multi-neuron activation function module |
Family Cites Families (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101026892B (en) * | 2006-12-30 | 2011-03-02 | 电子科技大学 | Label and pay load separating method for optical packet exchange |
| CN101055402A (en) * | 2007-05-16 | 2007-10-17 | 天津大学 | High speed all optical adder based on semiconductor optical amplifier |
| CN101094177B (en) * | 2007-07-12 | 2010-09-08 | 北京交通大学 | System for extracting envelope signal of net load in optical packet signal, and implementation method |
| CN102681287B (en) * | 2012-05-18 | 2014-11-12 | 上海交通大学 | Full optical code type conversion system based on stimulated Brillouin scattering effect |
| CN103545703B (en) * | 2013-10-23 | 2016-06-22 | 华南理工大学 | Light-operated trigger laser |
| EP3202056B1 (en) * | 2014-10-02 | 2019-06-19 | B.G. Negev Technologies & Applications Ltd., at Ben-Gurion University | All-optical silicon-photonic constellation conversion of amplitude-phase modulation formats |
| CN113253539B (en) * | 2021-05-31 | 2022-04-19 | 南京邮电大学 | All-optical packet switch based on integrated SOA series tunable laser |
-
2022
- 2022-01-27 CN CN202210102732.6A patent/CN114325932B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN114325932A (en) | 2022-04-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Peng et al. | Neuromorphic photonic integrated circuits | |
| Shastri et al. | Photonics for artificial intelligence and neuromorphic computing | |
| CN114325932B (en) | On-chip integrated all-optical neural network optical computing chip | |
| Nahmias et al. | Photonic multiply-accumulate operations for neural networks | |
| CN111860822B (en) | Method and device for realizing all-optical nonlinear activation function of optical neural network | |
| De Marinis et al. | Photonic neural networks: A survey | |
| US9477136B2 (en) | Reservoir computing using passive optical systems | |
| WO2022016894A1 (en) | Photonic neural network | |
| CN111683304B (en) | All-optical diffraction neural network and system realized on optical waveguide and/or optical chip | |
| WO2023201970A1 (en) | Computing chip, system, and data processing method | |
| US20220012582A1 (en) | Neuromorphic photonics with coherent linear neurons | |
| CN115169542B (en) | Two-dimensional photon convolution acceleration system and device for convolution neural network | |
| CN112101540B (en) | Optical neural network chip and calculation method thereof | |
| CN115222035B (en) | Photon neural network convolution acceleration chip | |
| CN115472603B (en) | Monolithic photon integrated chip for realizing convolutional neural network based on optical frequency comb | |
| US20230117456A1 (en) | Optical logic element for photoelectric digital logic operation and logic operation method thereof | |
| KR20210005273A (en) | System and method for activation function for photon neural network | |
| CN113805641B (en) | Photonic neural network | |
| CN112560288A (en) | Reserve pool computing device based on optical pump spin VCSEL | |
| Kaur et al. | SOA-MZI based 4× 4 interconnected crossbar photonic wavelength switching for datacenter load balancing | |
| CN112232487B (en) | Optical neural network chip and calculation method thereof | |
| CN115496195A (en) | MZI array and FP-SA-based optical pulse neural network implementation device | |
| CN117709423B (en) | Deep neural network photon acceleration chip and operation system thereof | |
| Guo et al. | Hardware implementation of multi-layer photonic spiking neural network with three cascaded photonic spiking neurons | |
| WO2023189231A1 (en) | Nonlinear conversion processing device |
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 |