86 results about "Optical neural network" patented technology

An optical neural network is constructed based on photonic integrated circuits to perform neuromorphic computing. In the optical neural network, matrix multiplication is implemented using one or moreoptical interference units, which can apply an arbitrary weighting matrix multiplication to an array of input optical signals. Nonlinear activation is realized by an optical nonlinearity unit, which can be based on nonlinear optical effects, such as saturable absorption. These calculations are implemented optically, thereby resulting in high calculation speeds and low power consumption in the optical neural network.

The invention provides an optical neural network method for realizing the digital recognition. The optical neural network method comprises a step of acquiring digital image features and a step of constructing an optical neural network. The optical neural network is composed of an optical interference module, an optical nonlinear module and a detector array, and the optical interference module comprises a Mach-Zehnder interferometer array and a variable optical attenuator, and can realize any matrix multiplication. The optical non-linear module is composed of a saturable absorber and other devices with the non-linear effects, and can realize the function of an activation function in an artificial neural network. According to the method, the calculation time is shortened through optical calculation, and the calculation energy consumption is reduced.

The invention discloses a method and a device for achieving an all-optical nonlinear activation function of an optical neural network. The method comprises the following steps: acquiring a to-be-processed signal optical signal and a reference optical signal coherent with the to-be-processed signal optical signal; inputting the to-be-processed signal optical signal and the reference optical signalinto a first phase shift module, wherein the first phase shift module performs phase shift operation on the to-be-processed signal optical signal and the reference optical signal to obtain an opticalsignal of a first array; and inputting the optical signal of the first array into an optical interference module, performing nonlinear operation on the optical signal of the first array in the opticalinterference module to obtain an optical signal of a second array, and outputting the optical signal of the second array as a nonlinear response of the signal optical signal to be processed. According to the technical scheme, the problem that an existing optical nonlinear function calculation unit has high requirements for optical power and a transimpedance amplifier is solved, parameters are adjustable, and a provided nonlinear function is flexible and controllable.

The invention provides an optical neural network processor and a training method thereof. The processor comprises: a numerical mapping means for realizing a mapping between a numerical value and a numerical value in a positive integer domain which can be represented by an optical neuron; An optical computing device including an optical neuron for performing a corresponding calculation of a networklayer of the neural network model according to an input value and a weight value within a positive integer domain represented by the optical neuron; a photoelectric converter which is used for converting an optical signal of a calculation result of the optical calculation device into an electric signal; And a nonlinear activation device which is used for executing nonlinear activation on the electric signal of the calculation result of the corresponding network layer.

The invention discloses a scattering medium optical imaging method based on a neural network. The method includes the steps of establishing a scattering medium optical neural network model and conducting imaging. In the step of establishing the scattering medium optical neural network model, modularization optimized signals before a scattering medium serve as the output layer of the neural networkmodel, light speckle intensity information formed after the signals penetrate through the scattering medium serve as the input layer of the model, the neural network is trained through the data of the input layer and the output layer, parameters of all middle layers in the neural network model are obtained, and therefore the specific scattering medium optical neural network model is obtained. Inthe scattering medium optical imaging process, a to-be-imaged target pattern serves as the input layer of the neural network model, light wave pre-modularized signals are calculated and output by theneural network, an incident light wave face is modularized through the pre-modularized signals, and the target image is directly formed through the scattering medium.

The invention discloses an image recognition method based on an optical neural network structure, an image recognition device and an electronic device, and the optical neural network structure consists of an X-layer neural network; The image recognition method comprises the steps of obtaining a to-be-recognized image; inputting the to-be-identified image into the optical neural network structure;determining a recognition result of the to-be-recognized image based on an output result of the optical neural network structure; wherein the optical neural network structure is used for obtaining aninput vector of the ith neural network for the ith neural network, and i is a positive integer greater than 0 and less than X + 1; performing linear transformation on the input vector based on Yi inner product calculation units to obtain Yi linear transformation results; activating the Yi linear transformation results through a nonlinear crystal to obtain Yi activation results; and taking the Yi activation results as output vectors of the layer of neural network. According to the scheme, a novel optical neural network structure is applied, and the image recognition speed is further increased.

Most artificial neural networks are implemented electronically using graphical processing units to compute products of input signals and predetermined weights. The number of weights scales as the square of the number of neurons in the neural network, causing the power and bandwidth associated with retrieving and distributing the weights in an electronic architecture to scale poorly. Switching from an electronic architecture to an optical architecture for storing and distributing weights alleviates the communications bottleneck and reduces the power per transaction for much better scaling. The weights can be distributed at terabits per second at a power cost of picojoules per bit (versus gigabits per second and femtojoules per bit for electronic architectures). The bandwidth and power advantages are even better when distributing the same weights to many optical neural networks running simultaneously.

An input layer outputs light having a relatively narrow emission angle distribution to a middle layer as an output signal if the signal level of input signal is relatively high and outputs light having a relatively broad emission angle distribution to the middle layer as the output signal if the signal level of input signal is relatively low. The middle layer outputs light having a relatively narrow emission angle distribution as an output signal to an output layer if the signal level of the output signal from input layer is relatively high and outputs light having a relatively broad emission angle distribution to the output layer as an output signal if the signal level of the output signal from the input layer is relatively low.

In order to solve the technical problem that the calculation rate and loss of a photoelectric hybrid system are still limited by the rate and ohmic loss of an electric clock, and the high cost, largesize, limited expansibility and amplification of errors during cascading of a conventional all-optical neural network system implemented by using discrete optical elements, the invention provides an all-optical diffraction neural network and system implemented on an optical waveguide and / or an optical chip. All-optical connection is realized through the diffraction free transmission region on theoptical waveguide and / or the chip, and more waveguide neuronal connections can be realized under the same size, so that the problems of few neuronal connections and weak system expansion can be effectively solved; and more neuronal connection error-tolerant rates are better, so that the method is higher in recognition precision.

The invention discloses on-chip cascaded MZI reconfigurable quantum network architecture based on lithium niobate. The on-chip cascaded MZI reconfigurable quantum network architecture comprises an input waveguide array serving as a neural network input layer, a plurality of MZIs which are optically connected with the input waveguide array and serve as a neural network hidden layer, a saturable absorber array, a nonlinear optical unit and a detector array serving as a neural network output layer. The MZIs are connected with each other and linearly converted into second array optical signals according to the first array optical signals; each saturable absorber in the saturable absorber array receives a corresponding optical signal in the second array optical signal and nonlinearly converts the optical signal into a third array optical signal, and the third array optical signal is detected by the detector array. According to the invention, the technical problem of on-chip coherent opticalneuromorphic calculation based on a photon integrated circuit is solved, and the limitation of calculation efficiency and power consumption in micro-electronic and hybrid optical electronic implementation is eliminated by a universal and reconfigurable quantum optical neural network.

An optical neural network convolution layer chip is applied to the field of artificial intelligence and comprises a first coupler, a first beam splitter, a plurality of photon calculation modules anda convolution summation module which are connected in sequence, wherein the first coupler is used for coupling a received optical signal into the first beam splitter; the first beam splitter comprisesa plurality of output ports, the beam splitter is used for splitting the coupled optical signals to obtain a plurality of beams of optical signals, and the plurality of beams of optical signals are input to the photon calculation modules through the output ports one by one; the photon calculation module is used for carrying out amplitude modulation and phase modulation on each beam of optical signals so as to represent input data and a convolution kernel parameter through each beam of modulated optical signals, and converting all the modulated optical signals into electric signals; and the convolution summation module is used for carrying out convolution summation on all the electric signals and completing photon convolution operation of all the input data and convolution kernel parameters. Photons have the characteristics of high speed, high bandwidth and low power consumption, convolution calculation is realized by utilizing the photons, the calculation speed can be greatly improved, and the calculation energy consumption is reduced.

The invention provides an optical neural network processor and a computing method thereof. The processor comprises a numerical mapping device, a positive value optical calculation device, a negative value optical calculation device, a photoelectric converter, a subtracter and a nonlinear activation device. During calculation, weights and input values of a network layer in a neural network model are mapped into an integer domain which can be represented by optical neurons and are divided into positive and negative optical paths for calculation of the network layer, and two paths of calculationresults are combined into one path through a subtracter.

The invention discloses a data processing method and device based on an optical neural network, a computer readable storage medium and the optical neural network. An optical interference unit of the optical neural network comprises a first interference light path structure, a phase shifter and a second interference light path structure, and each of the two interference light path structures comprises an internal phase shifter and an optical splitter. The method comprises the following steps: if the splitting ratios of the optical splitters of the two interference optical path structures both meet the splitting compensation condition, obtaining initial optical information and final output optical information of an input optical signal, and inputting and outputting optical information in the middle of an input / output port of the phase shifter; when the initial optical information and the intermediate input optical information as well as the intermediate output optical information and the final output optical information both meet preset light splitting conditions of the optical neural network, calculating parameters of internal phase shifters of the two interference optical path structures, and performing data processing by using the optical neural network based on the parameters to obtain a phase shifter. The optical neural network performance and the data processing accuracy can be effectively improved.

An all-optical neural network that utilizes light beams and optical components to implement layers of the neural network is disclosed herein. The all-optical neural network includes an input layer, zero or more hidden layers, and an output layer. Each layer of the neural network is configured to simulate linear and nonlinear operations of a conventional artificial neural network neuron on an optical signal. In an embodiment, the optical linear operation is performed by a spatial light modulator and an optical lens. The optical lens performs a Fourier transformation on the set of light beams and sums light beams with similar propagation orientations. The optical nonlinear operation is implemented utilizing a nonlinear optical medium having an electromagnetically induced transparency characteristic whose transmission of a probe beam of light is controlled by the intermediate output of a coupling beam of light from the optical linear operation.

The invention discloses an ultra-precise displacement measuring system based on an optical neural network. The system comprises a light source, an optical displacement measuring device, an optical neural network, a detector array and a signal processing device. When a target object moves, the system takes a measurement optical signal output by the optical displacement measuring device as a signalinput, the signal is received by the detector array after being processed by the optical neural network, and finally the signal is converted into displacement information of the target object throughthe signal processing device. The invention further discloses an ultra-precise displacement measuring method based on the optical neural network. The optical neural network is used for processing themeasurement optical signal, so that the displacement of the target object can be directly measured; the phase discrimination process of the electronic signal is not needed; the response speed is extremely high; the size can be zoomed; the energy utilization rate is high; and the system is suitable for ultra-precise measurement occasions with high speed and high dynamic performance requirements. According to the system, the multi-degree-of-freedom pose measurement of the target object also can be realized by increasing the number of input measurement optical signals and the number of detector arrays.

The invention discloses a coherent light QPSK judgment method based on an optical neural network. The method comprises the steps that firstly, according to a QPSK judgment problem, an optical neural network ONN model structure is designed, model parameters are trained, a judgment photon circuit is set up and set up according to the parameters, and the judgment photon circuit is a 2 * 4 optical network and is provided with two input ports and four output ports; and then, the QPSK receiving signal and the local oscillator signal are input into two input ports of a judgment photon circuit respectively at the same time, then processed and output through four output ports, and the phase difference between the receiving signal and the local oscillator signal is judged according to the port with the highest output power so as to complete QPSK judgment. The judgment method can meet the requirements of an all-optical communication network and an optical mobile communication system, received signals and local oscillation signals are directly processed in an optical band, and judgment or branching of the optical signals is directly completed according to phase difference information of two paths of input signals. The invention also discloses a coherent light QPSK decision system based on the optical neural network.

The invention discloses an optical chip and a manufacturing method thereof. Machine learning is performed by designing an all-optical diffraction deep neural network architecture. According to the architecture, various functions can be realized on the basis of joint work of passive diffraction layers based on deep learning; an all-optical neural network architecture is made into an integrated chipby a grating based on a 3D printing technology, a specific task of training can be executed at a light speed, powerful functions can be realized by only using optical diffraction and passive opticalcomponents or layers, and the working efficiency is greatly improved. The invention also discloses an optical identification device.

The invention provides a parallel optical neural network system based on a micro-cavity optical frequency comb and an identification method, and solves the problems that an existing optical neural network system is high in cost, large in size, unfavorable for monolithic integration and the like. The parallel optical neural network system based on the micro-cavity optical frequency comb comprises anarrow linewidth laser, a micro-cavity optical soliton optical frequency comb unit, an optical Bragg grating, an optical amplifier, a wavelength division multiplexing module, a first modulator and anoptical neural network unit which are arranged in sequence. According to the system, the micro-cavity optical soliton optical frequency comb is combined with the optical neural network, the micro-cavity optical soliton optical frequency comb has hundreds of frequency components, hundreds of different characteristics can be identified in parallel, and future application can be met.

The invention provides a robot system based on the optical neural network. The robot system comprises a communication system for receiving instructions sent to a robot from the outside, an optical neural network center processing system used for resolving and outputting optical signals carrying decision information, a control system, an execution system and a sensing system, wherein the control system is used for resolving and outputting the optical signals carrying control information and encoding the optical signals into electric signals carrying control information for output, the executionsystem is used for receiving the electric signals carrying the control information and executing corresponding instructions, the sensing system is used for monitoring the execution system and the outer environment and transmitting the sensed sensing information to the optical neural network center processing system in the form of optical signals, and the optical neural network center processing system decides the current robot behavior according to the sensing information. Most of calculating work of the robot system is conducted with the speed of light, the calculating speed is high, and real-time performance is high.

The embodiment of the invention provides an optical neural network device, a chip and an optical implementation method for neural network calculation. The device comprises: a light generation sub-device which is used for generating N paths of optical signals with different wavelengths, wherein the N is an integer greater than 1; a first modulation sub-device which is used for modulating the intensity of the N paths of optical signals according to the N first voltages to obtain N paths of first optical signals; a first conversion sub-device which is used for carrying out parallel-serial conversion on the N paths of first optical signals to obtain a second optical signal; an optical splitter which is used for splitting the second optical signal into N paths of third optical signals; a secondmodulation sub-device which is used for separately modulating the intensity of the N paths of third optical signals according to the N first voltage sets to obtain N paths of fourth optical signals;a second conversion sub-device which is used for performing serial-parallel conversion on each of the N paths of fourth optical signals to obtain N paths of fifth optical signals; and a processing sub-device which is used for adjusting the values of the N first voltages and the N first voltage sets based on the N paths of fifth optical signals.

The invention discloses an optical neural network all-optical nonlinear activation layer and an implementation method thereof. The MZI waveguide configuration is combined with the graphene heterogeneous enhanced Bi2Te3 nonlinear material, and the nonlinear response of the graphene heterogeneous enhanced Bi2Te3 nonlinear material is further amplified by utilizing the on-chip waveguide structure design, so that the design of the on-chip integrated nonlinear activation layer is completed, the optical nonlinear calculation in the on-chip integrated waveguide is realized, the problem that the nonlinear degree of an optical nonlinear material is weak is solved, the function of an optical neural network is expanded, and possibility is provided for use of a multi-layer pure optical neural network; and the optical nonlinear activation layer provided by the invention not only can be used for an on-chip integrated optical neural network, but also can be used for scenes needing nonlinear calculation in other integrated optical signal processing platforms, is extremely high in response speed, and can meet the requirements of low-energy-consumption and high-speed calculation.

The invention relates to a self-adaptive optical system based on an all-optical neural network, which belongs to the technical field of self-adaptive optical systems and comprises an all-optical neural network solver, a photovoltaic conversion array and a high-voltage amplifier. According to the invention, the all-optical neural network solver formed by optical diffraction plates is used for solution and modulation of a target light beam, and the target light beam is converted to an optical signals. The self-adaptive optical system is used for replacing a wavefront sensing device, a signal resolving device, a digital-to-analog conversion device and the like in a traditional self-adaptive optical system, the response bandwidth can reach the KHz magnitude, the high-bandwidth wavefront control requirement under a non-cooperative target scene can be met, meanwhile, a traditional circuit is replaced by the optical path, and the cost is reduced. Operation from a target light beam to a deformable mirror driving electric signal is achieved with extremely low power consumption and extremely high response speed, the effect of real-time wavefront correction is achieved, and the self-adaptiveoptical system can be applied to military and other scenes with high response bandwidth requirements for light beam wavefront correction.

The invention discloses a wavefront restoration method and system based on a diffractive optical neural network. The method comprises the following steps: 1) selecting or constructing a data set composed of wavefront-coefficient data pairs containing first N orders of Zernike; 2) constructing an optical neural network model, and fitting the data set to obtain two-dimensional phase distribution ofeach phase modulation plate in the model; determining the thickness of the corresponding phase modulation plate according to the two-dimensional phase distribution of each phase modulation plate, thewavelength of the light wave to be measured, and the refractive index and transmittance of the required phase modulation plate; and (3) manufacturing corresponding phase modulation plates according tothe thicknesses of the phase modulation plates determined in the step (2), respectively placing the phase modulation plates behind the wavefront to be measured according to the positions in the optical neural network model, modulating the complex amplitude of the optical wave, then detecting the light intensity distribution after passing through the phase modulation plates, and carrying out wavefront restoration according to the light intensity distribution. Photoelectric conversion and dependence on an electronic computer are avoided, and the method has the advantages of being low in energyconsumption, high in speed and the like.

The invention discloses an optical signal processing method, a photonic neural network chip, and a design method of the chip. The optical signal processing method comprises the following steps: determining to-be-processed data and a target modulation range of an optical signal; modulating the initial input optical signal according to the to-be-processed data and the target modulation range to obtain a target input optical signal; performing parallel convolution calculation based on the target input optical signal to obtain a target output optical signal; and performing signal conversion on the target output optical signal to obtain an output electric signal, and executing corresponding processing operation according to the output electric signal. An operation module in the photonic neural network chip provided by the embodiment of the invention adopts an optical matrix multiplier, an accumulator and a nonlinear optical element to realize convolution parallel calculation of the same layer, so that the network operation efficiency is greatly improved, and the advantage of high operation speed of an optical simulation accelerator is fully embodied. And meanwhile, in cooperation with electrical storage and control, a multilayer optical neural network is constructed, and mature popularization of the optical neural network is promoted.

The invention discloses a non-linear activation function RELU implementation method and equipment and a medium. The method comprises the steps of obtaining an original input quantity; mapping the original input quantity to light intensity to obtain a light signal; inputting the optical signal to the micro-ring resonator; applying corresponding current to the micro-ring resonator to adjust the phase of the micro-ring resonator to achieve activation function RELU operation. Therefore, on the basis of the micro-ring resonator, the characteristics of the RELU of the activation function can be simulated without a large amount of photoelectric conversion, the method has the advantages of rapidness and low power consumption, and the advanced layout of the optical neural network chip is realized.

The invention discloses a Hadamard product implementation method and device, and a storage medium. The method comprises the steps of obtaining to-be-processed optical signals of various different wavelengths; inputting a to-be-processed optical signal to a wavelength division multiplexer; feeding a to-be-processed optical signal to the micro-ring resonator structure by using a wavelength division multiplexer, wherein the micro-ring resonator structure comprises a plurality of micro-ring resonator groups consisting of two micro-ring resonators with the same radius; applying corresponding current to the micro-ring resonator structure, and obtaining a Hadamard product result according to the output light intensity. Therefore, the micro-ring resonator is used as the basis for realizing the artificial neural network scheme, the wavelength division multiplexer is used for feeding the optical signal to be processed to the micro-ring resonator structure, the effective refractive index and the phase of the micro-ring resonator can be changed through current heating, and the result of the Hadamard product can be obtained according to the light intensity of the output optical signal; therefore, a simulation solution suitable for the Hadamard product in the optical neural network is realized.

Provided are an optical device which is capable of optically implementing an activation function of an artificial neural network and an optical neural network apparatus which includes the optical device. The optical device may include: a beam splitter splitting incident light into first light and second light; an image sensor disposed to sense the first light; an optical shutter configured to transmit or block the second light; and a controller controlling operations of the optical shutter, based on an intensity of the first light measured by the image sensor.

The invention relates to the technical field of electrical digital data processing, in particular to a photoelectric integrated circuit for message compression in a message hash algorithm. The photoelectric integrated circuit is provided with first to Nth levels of optical neural networks and a photoelectric detector array in an integrated manner. The first to Nth levels of optical neural networks are used for carrying out message compression operation in a message hash algorithm on multiple paths of initial optical signals which are input in parallel through the waveguide step by step from the first to Nth levels, carrying out message compression operation on messages loaded to the multiple paths of initial optical signals through each level of optical neural network, and carrying out step-by-step compression to obtain final optical signals meeting compression conditions; and the photoelectric detector array is used for converting the final optical signal into an electric signal for carrying a message compression result. Therefore, the problems of high power consumption during message compression, reduction of operation performance and the like caused by high power consumption of dynamic flipping during implementation of each round of operation of a message hash algorithm based on a hardware circuit in related technologies are solved.