200 results about "Optical neural network" patented technology

A diffractive optical neural network device includes a plurality of diffractive substrate layers arranged in an optical path. The substrate layers are formed with physical features across surfaces thereof that collectively define a trained mapping function between an optical input and an optical output. A plurality of groups of optical sensors are configured to sense and detect the optical output, wherein each group of optical sensors has at least one optical sensor configured to capture a positive signal from the optical output and at least one optical sensor configured to capture a negative signal from the optical output. Circuitry and / or computer software receives signals or data from the optical sensors and identifies a group of optical sensors in which a normalized differential signal calculated from the positive and negative optical sensors within each group is the largest or the smallest of among all the groups.

The invention discloses an end-to-end optical computing chip based on multi-mode analog signal fusion processing, and belongs to the optical computing technology. Comprising a multi-mode input fusion front-end module, an optical fiber input interface and an end-to-end reasoning module. The multi-mode input fusion front-end module can convert different types of original analog signals such as images, spectrums and radio frequencies into unified broadband spectrum input signals. The end-to-end reasoning module constructs a deep optical neural network, and the deep optical neural network comprises a sensing-convolution integrated unit realized by an arrayed waveguide grating (AWG). The end-to-end reasoning module further comprises a photoelectric non-linear-pooling integrated unit, the average pooling function and the non-linear activation function are achieved at the same time, and light path loss is effectively compensated through injection of the supply light source. And finally, integrating the signals subjected to multi-layer processing by a full connection layer, and outputting a classification result by an output layer. According to the chip architecture, direct and efficient processing of multi-mode analog signals is realized.

The invention provides a multiband photoresponse nerve synapse device based on a two-dimensional and organic heterostructure, an array and application, and belongs to the technical field of organic electric solid-state devices. The device provided by the invention adopts a three-end bottom gate top contact structure and comprises a gate substrate, an insulating layer, a channel layer and a source / drain electrode, and the channel layer forms an II-type heterojunction by a two-dimensional semiconductor material with a narrow-band gap characteristic and an organic semiconductor material. The heterostructure realizes ultraviolet to far infrared (365nm-10 [mu] m) multiband photoelectric response, and shows synaptic plasticity under light pulse modulation, including short-term enhancement, long-term enhancement and learning-forgetting-re-learning characteristics. Large-area array preparation of the device can be realized through methods such as chemical vapor deposition, solution spin coating, jet printing or layer-by-layer self-assembly. The technology not only can be used for brain-like learning and memory simulation, but also can be applied to the fields of sensing and memory integrated intelligent systems, optical neural networks, artificial vision, multi-mode intelligent sensing and the like.

The invention relates to the technical field of optical computing, and discloses an optical neural network and an implementation method thereof, and an optical neural network chip, and the method comprises the steps: an input layer converts and processes an input external signal, obtains an optical signal, and inputs the optical signal to a hidden layer; according to the task requirement of an external signal, the hidden layer dynamically adjusts the weight parameter of the optical neural network under the external magnetic field control condition; an optical signal output by the hidden layer is transmitted to the nonlinear activation layer; aiming at the task requirement of an external signal, the nonlinear activation layer dynamically adjusts the nonlinear characteristic of an activation function under an external magnetic field control condition; and the output layer converts the optical signal output by the nonlinear activation layer into an electric signal, carries out corresponding processing, and outputs a final result. Therefore, the optical neural network is endowed with dynamic adjustability and rapid adjustment capability among different network layers, the weight parameter and the nonlinear function of the full-optical domain dynamic adjustment optical neural network are realized, and the overall operation efficiency is improved.

Systems and methods are provided for monolithic photonic integrated devices for implementing convolutional neural network (CNN). Examples provide for photonic devices comprising a first optical transform device configured to perform a first transform on input optical signals to output a plurality of intermediate optical signals, a plurality of optical modulation devices configured to perform a convolution on the intermediate optical signals, and second transform device configured to perform a second transform of the plurality of intermediate optical signals and output a plurality of output optical signals. Each of the modulation devices comprises an amplitude modulator and a phase modulator. The photonic device, including an optical source that generates an input optical signal, can be formed monolithically on a common substrate to provide a convolutional optical neural network (CONN) integrated onto a single chip.

The invention discloses an optical neural network non-linear activation system, method and device, and relates to the technical field of optical neural networks, because a magneto-optical material is integrated in a phase adjustment waveguide arm of a non-linear activation unit, a controller can adjust the current activation function type and weight of the optical neural network according to the current activation function type and weight of the optical neural network. And controlling a magneto-optical material of the phase adjustment waveguide arm to generate a magneto-optical effect so as to perform phase adjustment on an optical signal transmitted through the phase adjustment waveguide arm, so that the nonlinear activation unit can adapt to the type and weight of a current activation function, and perform corresponding nonlinear calculation on a target input signal to obtain a target input signal. Dynamic switching of various nonlinear activation functions is realized, and the flexibility of the optical neural network is improved.

The invention discloses a system and a method for realizing positive and negative weighting in an optical domain, and the system comprises a multi-wavelength light source array which provides multi-channel optical signals, and each wavelength corresponds to an input channel; the adjustable bias photoelectric modulator arrays are in one-to-one correspondence with the light sources, a quiescent working point is biased in a positive / negative slope linear region of a transmission curve through an independent bias circuit, positive and negative weight symbol control is realized, and an input electric signal is loaded; the multi-wavelength micro-ring weighting unit is used for weighting the modulated optical signal by regulating and controlling the transmissivity of the micro-ring; the photoelectric detection unit adopts a single-path detector to receive a composite optical signal and convert the composite optical signal into an electric signal; and the capacitor blocking and signal synthesis unit filters a direct current component, synthesizes an alternating current weighted signal and outputs a real number weighted sum. The system does not need to balance detectors and redundant devices, simplifies hardware architecture, reduces cost and size, reduces power consumption, improves weighting precision and reliability, supports flexible expansion of multi-wavelength channels, adapts to requirements of a large-scale optical neural network, and provides support for optical calculation integration and low-power-consumption landing.

An all-analog optical neural network includes multiple all-analog optical neural network layers; a laser and splitter configured to distribute light signals from the laser equally across all of the multiple all-analog optical neural network layers; integrated MZI switches configured to switch the all-analog optical neural network to a hybrid backpropagation training configuration that measures the light signals in forward and backward directions, and a trains a linear portion of the all-analog optical neural network. Preferably, each of the all-analog optical neural networks comprises: an integrated silicon photonic neural network (PNN) of Mach-Zehnder interferometers (MZIs) and programmable phase shifters (η) configured to implement a programmable unitary matrix-vector multiplication (MVM) operation U; photonic meshes configured to send input forward and backward inference signals to the PNN and configured to measure using both amplitude and phase detection an output forward signal and a backward adjoint signal from the PNN.

The invention provides an optical neural network computing system and method based on time domain excitation characteristics, and the system comprises a pulse clock sequence module which is used for generating a reference time sequence; the multi-channel pulse laser array is used for generating semiconductor neuron laser pulses with subnanosecond pulse width based on a reference time sequence; the time domain and wavelength division multiplexing module is used for simultaneously performing time domain coding and wavelength division multiplexing on the laser pulse based on a reference time sequence to generate a first optical signal; the reconfigurable weight matrix module is used for performing all-optical linear matrix operation and weight mapping on the first optical signal to obtain a second optical signal; the neuron activation module is used for performing nonlinear activation processing on the second optical signal based on an LIF theory to obtain an activated optical signal; and the output module is used for carrying out decoding and photoelectric conversion on the activated optical signal to obtain a network reasoning result.

The invention provides a ferromagnetic artificial neuron film and application thereof. The ferromagnetic artificial neuron film comprises a transparent substrate and a ferromagnetic heterojunction constructed on the transparent substrate, the ferromagnetic heterojunction comprises a ferromagnetic alloy layer, a heavy metal layer and an anti-oxidation layer. Nonlinear spin oscillation of the ferromagnetic artificial neuron film is excited through femtosecond laser, terahertz radiation with intensity nonlinear response is directly generated, and the ferromagnetic artificial neuron film serves as an artificial neuron of a space optical neural network. The activation function response meeting the performance requirement of the artificial neuron can be effectively generated without introducing the regulation and control of an external magnetic field, and a better artificial neuron function facing the optical space neural network is realized.

The invention belongs to the field of silicon-based photonics, and particularly relates to an MZI optical neural network construction method based on polarization control and a set port, singular value decomposition is carried out on a weight matrix W of a traditional neural network model to obtain a unitary matrix and a diagonal matrix, a mode of combining a plurality of MZI units is utilized, a unitary matrix or diagonal matrix structure is adopted, and the MZI optical neural network is constructed. Constructing an optical neural network; the method has the advantages that a larger-scale optical matrix is constructed, a higher-dimension data processing task is supported, and then the processing capacity and accuracy of the whole network are improved.

The invention discloses an optical neural network multi-architecture adaptive computing chip and a computing method thereof. The chip comprises an optical signal input unit, an input data modulation unit, a weight and bias processing unit and a photoelectric detection unit. The optical signal input unit is used for generating and outputting an optical signal; the input data modulation unit is used for modulating and loading information carried by the source data and the bias data to an optical signal output by the optical signal input unit; the weight and bias processing unit comprises a photon weight module and a photon bias module, and the photon weight module is used for carrying out optical coding on weights and realizing optical multiplication and addition calculation of input data of optical signals and corresponding weight parameters; the photon bias module is used for bias coding, summing the bias coding and an optical multiplication and addition operation result, and outputting an operation result containing bias compensation; the photoelectric detection unit is used for converting operation results into electric signals to be output. The method has high universality, and the calculation accuracy of the optical neural network is effectively improved.

The invention discloses a photon convolutional neural network system, which belongs to the technical field of optical neural networks and comprises dynamic convolution kernel modules in one-to-one correspondence with convolution layers in a pre-trained convolutional neural network model. The dynamic convolution kernel module comprises a photon synaptic device based on a phase change material, and convolution kernel weights corresponding to convolution layers are stored in the photon synaptic device in a non-volatile manner; before the dynamic convolution kernel module performs convolution operation, according to a dynamic coefficient generated by a coefficient generation model according to the content of input data, the weight of a corresponding photon synaptic device in the dynamic convolution kernel module is dynamically adjusted in a maladjustment-prone mode, so that the dynamic convolution kernel module can flexibly adapt to the change of the input data; according to the invention, coding is carried out not only depending on a non-volatile phase state of a phase change material, and volatile modulation coding is further introduced on the basis, so that the coding expression capability of a dynamic convolution kernel module is greatly enriched; on the basis, the feature extraction capability of the convolution kernel in the photon convolutional neural network system provided by the invention is relatively high.

The invention relates to an all-optical non-linear layer, an all-optical neural network system and a calculation method, the all-optical non-linear layer comprises a transparent carrier and at least one light limiting material covering the transparent carrier, and the all-optical non-linear layer is used for limiting the light intensity of an incident first light signal under the condition that the light intensity of the incident first light signal is larger than a preset threshold value. Performing nonlinear transformation on the first optical signal to transmit a second optical signal; wherein under the condition that the light intensity of the first light signal is greater than a preset threshold value, the absorptivity of the light limiting material to the first light signal is increased along with the increase of the light intensity of the first light signal. According to the all-optical nonlinear layer disclosed by the embodiment of the invention, the nonlinear response of light intensity can be realized under the condition that electric or magnetic assistance is not needed, and more flexibility and possibility are provided for the construction of an optical neural network.

The invention discloses an extensible on-chip optical neural network chip based on a partially coherent light source, and belongs to the field of optical calculation. The optical neural network chip comprises a partially coherent light source, an input coding unit, a linear calculation unit, a nonlinear activation unit, an output detection unit and a calculation and control unit. According to the invention, a narrow linewidth laser is replaced by a partially coherent light source, so that the dependence of the system on coherent detection is reduced, and the large-scale capability of optical matrix calculation is improved. An optical-electrical-optical (O-E-O) nonlinear activation unit is adopted to realize the expandability of the network level, and the calculation efficiency and the energy utilization rate are effectively improved in combination with an optimized convolution calculation architecture. The chip provided by the invention adopts an on-chip integrated manufacturing process, can realize end-to-end optical reasoning, supports a complex mode recognition task, has the characteristics of high bandwidth, low energy consumption, expandability and the like, and is suitable for large-scale artificial intelligence calculation.

The invention relates to the technical field of optical computing, in particular to a deep optical neural network model, architecture and system. Wherein the model comprises an optical calculation module set, the connection mode between optical calculation modules in the optical calculation module set is determined by a target calculation task, each optical calculation module comprises an optical calculation matrix, and the number of rows and the number of columns of the optical calculation matrix are determined by the target calculation task. By adopting the scheme, a high-parameter-quantity complex intelligent model can be realized through the optical calculation module.

The invention provides a data processing method and device, a storage medium and computer equipment, and the method comprises the steps: carrying out the traversal of to-be-processed data according to a convolution calculation rule, and encoding an initial digital domain signal corresponding to the subdata into a first optical domain signal for the current traversal subdata; inputting the first optical domain signal into an optical neural network for convolution processing and nonlinear processing to obtain a second optical domain signal of the sub-data; the convolution kernel weight used in the convolution processing is represented by using the optical constant value of the phase change material in each storage and calculation integrated unit set in the optical neural network; and converting the second optical domain signal corresponding to each subdata obtained by traversing into a target digital domain signal, and inputting the target digital domain signal into a classification neural network to obtain a classification result. The method breaks through the limitation of an existing optical neural network in the aspects of power consumption and depth, has the advantages of high efficiency, low delay, expandability and the like, and is suitable for edge intelligent calculation scenes of tasks such as image classification.

The exponential growth in deep learning models is challenging existing computing hardware. Optical neural networks (ONNs) accelerate machine learning tasks with potentially ultrahigh bandwidth and nearly no loss in data movement. Scaling up ONNs involves improving scalability, energy efficiency, compute density, and inline nonlinearity. However, realizing all these criteria remains an unsolved challenge. Here, we demonstrate a three-dimensional spatial time-multiplexed ONN architecture based on dense arrays of microscale vertical cavity surface emitting lasers (VCSELs). The VCSELs, coherently injection-locked to a leader laser, operate at gigahertz data rates with a 7T-phase-shift voltage on the 10-millivolt level. Optical nonlinearity is incorporated into the ONN with no added energy cost using coherent detection of optical interference between VCSELs.

The invention discloses a rapid biological cell classification imaging method and system based on a quantum correlation diffraction optical neural network, and the method comprises the steps: regulating and controlling a signal photon to an orbital angular momentum state optimized by a neural network based on a time-energy correlation two-photon pair generated in a spontaneous parametric down-conversion process, and irradiating a biological cell sample; a diffractive optical neural network is utilized to construct a quantum state illumination diffractive optical processor, photons carrying cell complex amplitude information are mapped to different detection areas of a predefined image plane, and rapid and high-precision cell classification and recognition are realized through correlation imaging results of an enhanced charge coupled camera. The method has the signal-to-noise ratio and anti-interference capability exceeding the classic limit, can keep high-precision cell classification capability under high background noise, and is suitable for long-time monitoring of living cells under the condition of extremely low light intensity. The method has the characteristics of low energy consumption and high efficiency, and meets the requirements of green calculation and sustainable development. And the method can be used for rapidly classifying and identifying the living biological cells without marks and phototoxicity.

The invention provides an all-optical neural network system and an optical information processing method. The system comprises a pulse laser light source used for emitting a pulse laser beam; the digital micro-mirror device is used for loading the input image information and coding the input image information to the laser beam irradiated on the surface of the digital micro-mirror device; the phase modulation device is used for carrying out phase modulation on the laser beam which is emitted from the surface of the digital micromirror device and carries the input image information so as to form an optical signal with preset light intensity distribution on an output plane; the optical fiber delay line array device comprises an optical fiber delay line array and a beam combiner and is used for applying different time delays to the optical signals with preset light intensity distribution and combining and outputting the optical signals; the single-pixel high-speed detector is used for receiving the optical signal output from the optical fiber delay line array device and converting the optical signal into an electric signal at a high speed; and the signal acquisition and processing module is used for acquiring the electric signal to obtain one-dimensional pulse time domain waveform signal data so as to obtain input image information.

The application discloses an optical neural network nonlinear activation system, method and device, relates to the technical field of optical neural networks, and because a magneto-optical material is integrated in a phase adjustment waveguide arm of a nonlinear activation unit, a controller can control the magneto-optical material of the phase adjustment waveguide arm to generate a magneto-optical effect according to a current activation function type and weights of the optical neural network, so as to perform phase adjustment on an optical signal transmitted through the phase adjustment waveguide arm, and make the nonlinear activation unit adapt to the current activation function type and the weights, perform corresponding nonlinear calculation on a target input signal, and realize dynamic switching of multiple nonlinear activation functions, thereby improving the flexibility of the optical neural network.

The application discloses a kind of diffractive light neural network computing systems and methods for realizing pure optical nonlinearity, belong to photonic computing and artificial intelligence hardware technical field.Its system includes coherent light source, linear light computing unit, nonlinear light activation unit and optical detection unit arranged in order along optical path, wherein linear light computing unit is programmable phase modulation using spatial light modulator, nonlinear light activation unit is programmable binary amplitude modulation using digital micromirror device, and both are directly coupled to form pure optical computing path without photoelectric conversion on optical path.Its method trains neural network through customized loss function, drives nonlinear activation function output to binary convergence, and maps the parameters obtained by training into phase map and binary mask respectively and loads to hardware.The application realizes true all-optical nonlinear computation, with the advantages of high energy efficiency, extremely low delay, compact structure and strong parallel processing capability.

The application discloses a design method and device of a metasurface optical neural network, and the design method comprises the following steps: obtaining an artificial neural network model and a metasurface optical neural network model, wherein the artificial neural network model is a trained model; selecting n intermediate layers in the artificial neural network model in the order from shallow to deep, one-to-one corresponding the n intermediate layers and n metasurface layers in sequence, and calculating an intermediate feature loss according to the similarity between the features output by each layer in the n intermediate layers and the features output by the corresponding metasurface layer; training the metasurface optical neural network model; and determining the process parameters of the metasurface according to the trained metasurface optical neural network model. The design method completely transplants the calculation capacity of the artificial neural network into the metasurface optical neural network, solves the problems of low design efficiency and low precision of the metasurface optical neural network, and is beneficial to the development of the metasurface optical neural network and the deployment in an environment with limited computing power and storage.

The invention relates to a logical operation-oriented optical neural network training method and device, and the method comprises the steps: determining the light field total energy and background light intensity of each logical operation output port, so as to construct a logical contrast function; establishing a logic contrast loss function according to the logic contrast function, and constructing a mixed loss function in combination with a mean square error loss function; and based on the adaptive training strategy, training the optical neural network by using the hybrid loss function, judging whether the hybrid loss function meets a preset convergence requirement in the training process, and if the hybrid loss function does not meet the preset convergence requirement, continuing to train the optical neural network by using the hybrid loss function. And otherwise, training the optical neural network by using the mean square error loss function to generate a target optical neural network meeting a preset training requirement. Therefore, the problems that when an existing mean square error method faces logic operation, logic correctness cannot be guaranteed, and error bits occur are solved.

The invention belongs to the technical field of optical signal processing, and discloses an image classification system and method based on asymmetric pumping, and the system comprises an image obtaining module which is used for obtaining a handwritten digital image and transmitting the handwritten digital image to a nonlinear XOR module; the nonlinear XOR module is used for generating a plurality of serial signals according to the received handwritten digital image; delaying the serial signal to obtain a first signal and a second signal; sending the first signal and the second signal to a BIC resonant cavity to obtain a coupled XOR signal, and sending the coupled XOR signal to a linear classification module; the first signal is delayed by a preset time length than the second signal, and the difference between the power of the first signal and the power of the second signal is equal to a preset threshold; and the linear classification module is used for classifying the handwritten digital images according to the coupled XOR signals and outputting digital classification labels. According to the invention, all-optical nonlinear transformation is realized in the optical neural network, and the data separability is remarkably improved.

The invention relates to the technical field of optical computing, in particular to an optical neural network reconstruction-inhibition neuron model and architecture. The model comprises an optical reconstruction module used for receiving an optical input signal and carrying out reconfigurable modulation on the optical input signal to obtain a modulated optical signal, and a modulation parameter corresponding to the reconfigurable modulation is determined by an optical calculation task corresponding to the optical input signal; and the inhibition neuron module is used for constructing a filtering channel corresponding to the optical calculation task, filtering the modulated optical signal according to the filtering channel, and obtaining and outputting an optical output signal. By the adoption of the scheme, high-performance optical calculation can be achieved.

The invention discloses an optical nonlinear activator chip based on a graphene photoelectric device and a preparation method. Belongs to the technical field of photoelectric devices and optical signal processing. The chip comprises a detector optical path and a modulator optical path, a substrate is made of a standard SOI material, and the chip structurally comprises an optical coupling module, a waveguide, a graphene detector part and a graphene modulator part. The graphene detector and the modulator are integrated on the same chip, the detector is used for converting an input optical signal into a voltage signal, then the voltage signal is input into the modulator to modulate continuous light, and an optical signal after nonlinear conversion is output and used for a nonlinear activation unit of an optical neural network. Compared with a traditional nonlinear activator of a silicon-based integrated photon chip, the method gives full play to the advantages of excellent photoelectric property and easy integration of graphene, has the advantages of high integration level, high speed, low power consumption and the like, and has better compatibility with a silicon-based optoelectronic process.

The application provides a matrix operation accelerator combining wavelength division multiplexing and MZI cascade network, relates to the field of optical neural networks, and comprises an input signal layer, a weight signal layer, a summation layer and a nonlinear layer.The input signal layer is used for realizing matrix operation of optical signals through a Mach-Zehnder interferometer array; the weight signal layer is used for applying an electrical signal to a micro-ring modulator array to adjust a weight signal; the summation layer is used for separating the results of the action of the optical signals of different wavelengths through the weight signal; and the nonlinear layer is used for converting the optical signals into electrical signals through a photodetector array to realize a nonlinear activation function in the electrical domain.The application introduces N different wavelengths in the network formed by MZI cascade, so that the number of times of executing matrix operation is increased by N times each time, high-speed convolution operation is facilitated, and the size of the micro-ring modulator is relatively small, so that the energy efficiency and area ratio of the MZI cascade network calculation can be effectively increased.

The invention discloses a method for optically realizing a restricted Boltzmann machine. According to the invention, analog calculation of the restricted Boltzmann machine and optical Gibbs sampling can be realized, and the system has the advantages of wide application range, fast information transmission, simple structure, low cost, fast calculation and the like. According to the invention, a coherent wide-spectrum light source is used as signal input, a light field is subjected to light splitting by using a light splitting device such as a grating and then enters a modulator, then spinning, interaction and magnetic field parameters are coded on light wavefront at different positions through a time or space light modulator, and optical Fourier transform is carried out by using a time or space lens system, so that the optical field is obtained. The light intensity after Fourier transform is measured through the detector, and finally the difference between the light intensity measured twice is calculated, so that optical Gibbs sampling is realized, the calculation complexity is reduced, and the calculation efficiency is improved. The method provided by the invention has important application prospects in the fields of optical neural networks and the like, can realize applications of content generation, classification and identification and the like, and is convenient to integrate in optical chips and the like.

The present application provides a port for an optical neural network and a three-dimensional port of a three-dimensional optical neural network. The port of the optical neural network provided in the present application includes: an input port and an output port; wherein the input port is used to receive input light and transmit the input light to the computing layer of the optical neural network; the output port is used to receive output light from the computing layer and emit the output light to a detector array for receiving output light; the input port includes an input optical waveguide device and an input lens array, the input lens array is arranged upstream of the optical path of the input light; the input lens array includes at least two input lenses; the output port includes an output optical waveguide device. The port for an optical neural network provided in the present application can realize independent input and output of each optical waveguide in the optical neural network, reduce the loss of input light and output light, and improve the energy efficiency of the entire optical neural network.