36 results about "Linear operators" patented technology

The invention discloses a reactor thermal hydraulic characteristic rapid prediction method based on dynamic mode decomposition and deep learning, and the method comprises the steps: obtaining thermal hydraulic characteristic data of a reactor under different operation conditions, and selecting steady-state operation data to establish a snapshot; initial dimension reduction is carried out on the snapshot through singular value decomposition, and snapshot representation after dimension reduction is obtained; constructing an order reduction model based on the snapshot representation after dimension reduction; wherein the step of constructing the reduced-order model comprises the steps of performing dynamic mode decomposition on the snapshot after dimension reduction, constructing a parameterized low-dimensional linear operator, and fitting a mapping relation between the operator and an operation parameter through deep learning; verifying the reduced-order model according to a simulation result of the nuclear reactor mechanism model, and obtaining a parameterized reduced-order model when the verification is passed; and inputting the actual operation condition data into the parameterized reduced-order model to obtain a prediction result of the thermal hydraulic characteristics of the reactor. According to the method, the simulation efficiency is remarkably improved by combining the dynamic mode decomposition and the deep learning technology.

The invention provides a virtualization-based big language model parameter encryption protection method and system, and the method comprises the steps: carrying out the corresponding preprocessing of a current to-be-protected big language model in a trusted environment, obtaining a modified linear operator parameter, and carrying out the encryption of a nonlinear operator; the non-trusted virtual machine is a main interface which can be accessed by a user, executes a reasoning task on the preprocessed large language model, and when operation of nonlinear operator encryption needs to be carried out, the non-trusted virtual machine is switched to the trusted virtual machine by calling the service of the virtual machine hypervisor, and the trusted virtual machine completes the operation of nonlinear operator encryption; and the trusted virtual machine processes the encryption operator operation unauthorized to be processed by the non-trusted virtual machine, and returns a result to the non-trusted virtual machine for next calculation of the non-trusted virtual machine. According to the method, mathematical characteristics of different operators of the model are identified, a mode of dynamically modifying the weight of the linear operator without influencing the output result of the model is provided, and safe and efficient operation of the linear operator is ensured.

A system and method are provided for training a neural network for controlling operation of a system having non-linear dynamics represented by partial differential equations (PDEs). The method includes collecting a digital representation of time series data indicative of an instance of a function space of the system and a measurement of a state of operation of the system. A configuration point corresponding to the solution of the PDE is generated. A neural network is trained using training data including the collected time series data and the configuration points to train parameters of the non-linear operator. The neural network has an autoencoder architecture, the autoencoder architecture comprising: an encoder to encode each instance of training data into a potential space; a non-linear operator for propagating the encoded instance into a potential space using a transformation determined by a parameter of the non-linear operator; and a decoder to decode the transformed encoded instance of the training data to minimize the hybrid loss function.

This invention proposes a GQA-LUT method, utilizing a genetic algorithm and LUT-based circuit to efficiently approximate non-linear operators in Transformers. It adaptively finds optimal solutions for various non-linear functions, outperforming conventional neural network methods. A novel rounding mutation (RM) algorithm enhances approximation accuracy during quantization, improving low-bit integer precision. The invention also introduces a LayerNorm folding strategy as a near-memory computing principle, reducing IO and energy overheads with a two-stage memory hierarchy. Additionally, an additive partial sum quantization method is proposed to reduce energy consumption by quantizing accumulated PSUMs in matrix multiplication, alongside a PSQ-APSQ grouping strategy and floating-point regularization.

A training system for a neural network model includes a memory and a processor. The memory is configured for storing the neural network model and several instructions. The processor is configured for executing the instructions to perform a training method including: (a) receiving an image data; (b) performing a feature calculation based on the image data to obtain a feature data; (c) performing a linear classification calculation based on the feature data by using a mathematical operator; (d) performing a non-linear classification calculation based on the feature data by using a non-linear operator and another mathematical operator; and (e) performing a combination calculation based on a first result of the linear classification calculation and a second result of the non-linear classification calculation.

The invention provides a GQA-LUT method. According to the GQA-LUT method, a non-linear operator in a converter is effectively approximated by utilizing a genetic algorithm and an LUT-based circuit. The GQA-LUT method can adaptively find optimal solutions of various nonlinear functions, and is superior to a conventional neural network method. A novel rounding variation (RM) algorithm enhances approximation accuracy during quantization, thereby improving low order integer precision. In the invention, a LayerNorm folding strategy is also introduced as a near memory calculation principle, so that IO and energy overhead of the hierarchical structure of the two-stage memory is reduced. Furthermore, an additive partial and quantization method is proposed to reduce energy consumption by quantizing accumulated PSUM in matrix multiplication as well as a PSQ-APSQ grouping policy and floating point regularization.

The application provides a model quantification method, device and equipment and storage medium, the method comprises: using integer algorithm to execute a plurality of operators of transformer model for image processing, the plurality of operators of the transformer model comprises linear operator and nonlinear operator, wherein the linear operator comprises matrix multiplication operator, the matrix multiplication operator adopts symmetric quantization to quantize floating point value to integer value;The nonlinear operator comprises an activation function operator and a layer normalization operator, the activation function operator adopts polynomial fitting to quantize floating point value to integer value, and the layer normalization operator adopts the mean and standard deviation of input data in the channel dimension to quantize floating point value to integer value.The transformer operator is quantified, so that the inference work of the transformer model for natural language is based on integer operation, so that it can be truly deployed on FPAG chip for practical application.

The invention discloses a nonlinear equalization method and device for digital subcarrier multiplexing signals and an optical fiber communication system. The method comprises the following steps: caching an original symbol of a transmitting end, and preprocessing a received signal to obtain a signal to be equalized; constructing a power model according to the link parameters and deducing a back propagation power model; inputting the data into an improved digital back transmission algorithm, and calculating a nonlinear accumulation amount in each equalization step length; determining an initial splitting ratio parameter by using a nonlinear symmetry principle, and embedding the initial splitting ratio parameter into the neural network as a trainable variable; inputting a received signal and an original symbol into a network, and optimizing a splitting ratio through gradient descent to obtain an optimal value; therefore, the position of a nonlinear operator in the improved digital back transmission algorithm is adjusted, linear and nonlinear combined compensation of the signal is realized, and a high-precision equalization result is output. According to the method, extra communication overhead is not needed, and the link power non-uniformity characteristic can be effectively adapted.

The invention relates to the technical field of numerical weather forecasting, in particular to a hierarchical adaptive Gaussian conversion method and device for cloud and rain control variables, which can obtain observation data and input the observation data into a numerical forecasting model to obtain three-dimensional cloud and rain control variables; constructing a variational assimilation model comprising a forward transformation operator, an inverse transformation operator, a tangent linear operator and an adjoint operator; the forward conversion operator can calculate and adjust the conversion intensity layer by layer, so that finer and more accurate Gaussian processing is realized; the tangent linear operator and the adjoint operator are used for assimilating minimization iterative calculation in the model, and then the model is restored to the original physical magnitude through the inverse transformation operator to output a three-dimensional cloud and rain control variable analysis field according with reality. According to the technical scheme, the vertical distribution difference of cloud and rain variables can be accurately adapted, and the compatibility with an existing assimilation system is ensured by providing a complete operator chain, so that the quality of an analysis field and the forecasting capability are improved.

A method of reconstructing a spatial distribution of a characteristic (F) in an object, including a) acquisition of measurements (M) by a sensor, each measurement being able to be estimated a linear operator ((HF),P*F, (F)), applied to the spatial distribution of the characteristic (F), forming a forward model; b) with a processing unit, reconstruction of the spatial distribution of the object characteristic, by iterative minimization of an error, each iteration comprising an update of the spatial distribution of the object characteristic; where in step b), the minimized error includes a data attachment component (εD(f), εD (F)) including a deviation between the acquired measurements and the measurements estimated by the forward model; a regularization component (εR (f), εR (F)), including a sum of a norm of a spatial gradient of the feature, determined at different coordinates in the object.

The embodiment of the invention provides a nonlinear operator approximate calculation device and method, a neural network processor and a medium, and belongs to the technical field of neural networks. The device comprises: a floating point number evaluation unit receiving original floating point data of an original neural network operator, and performing compensation interval evaluation on the original floating point data according to a preset floating point value domain range to obtain compensation interval evaluation information; if the compensation interval evaluation information represents that the original floating point data is not in the preset floating point value domain range, the index splitting unit splits the original floating point data into a first floating point number and a second floating point number; the operation compensation unit performs fitting compensation on the first floating-point number to obtain first output data; and the splicing unit splices the first output data and the second floating-point number passing through the original neural network operator to obtain target output data. According to the invention, the computing resources and time of a computer system can be reduced, the full-value-domain compensation of the neural network operator is completed with few hardware resources, the computing precision is improved, and the reasonability of network reasoning is ensured.

Method for reconstructing a spatial distribution of a characteristic () in an object, comprising: acquisition of measurements () by a sensor (15), each measurement being able to be estimated a linear operator ((), ), applied to the spatial distribution of the characteristic (), forming a direct model; using a processing unit (20), reconstruction of the spatial distribution of the characteristic of the object, by iterative minimization of an error, each iteration comprising an update of the spatial distribution of the characteristic of the object; the method being characterized in that during step b), the minimized error comprises: a data attachment component () comprising a difference between the acquired measurements and the measurements estimated by the direct model; a regularization component (), comprising a sum of a norm of a spatial gradient of the characteristic, determined in different coordinates in the object.

The embodiment of the invention provides an artificial neural network model conversion method, a storage medium and a program product, and relates to the technical field of artificial intelligence, and the method comprises the steps: after obtaining an artificial neural network model obtained through pre-training, converting each nonlinear operator in the artificial neural network model into a corresponding pulse module, each pulse module comprises a difference expectation compensation module, the difference expectation compensation module is used for calculating an output increment according to the accumulated membrane potential and inserting a difference pulse neuron into each pulse module, the difference pulse neuron updates a coding activation value when issuing a pulse, otherwise, the coding activation value is kept unchanged, and the difference expectation compensation module is used for outputting the difference pulse neuron. According to the method, the bias term of the linear operator located on the previous layer of each nonlinear operator is removed, the initial membrane potential of the differential pulse neuron inserted into the pulse module corresponding to the nonlinear operator is set as the bias term, and the coding activation value is updated only when the pulse is emitted, so that the loss caused by updating the coding activation value no matter whether the pulse is emitted or not is avoided, and the accuracy of the coding activation value is improved. And the energy consumption is obviously reduced.

The application relates to the technical field of network confrontation, and provides a network space confrontation knowledge graph construction method based on an operator framework, which comprises the following steps: mining an operator set in a network space, wherein the operator set comprises a first operator, a second operator and a third operator; the first operator is an identity operator, the second operator is a strong operator, and the third operator is a total bounded linear operator; performing convergence judgment on the second operator, combining the first operator and the third operator to form a target operator value framework; performing attack and defense in the network space based on the operator value framework, obtaining network attack and defense data; mapping and corresponding the network attack data and the network defense data to obtain a construction data set; constructing a network confrontation knowledge graph; and performing information confrontation control in the network space based on the network confrontation knowledge graph. The method can solve the technical problem that the key node confrontation effect is poor in the network confrontation process.

The application discloses a boundary level grid physical constraint variational assimilation method embedded in a deep neural network, comprising: establishing a momentum equation containing a boundary level grid turbulent friction term, wherein the boundary layer turbulent friction term is simulated through a deep neural network; and constructing a weak constraint term of a variational assimilation framework cost function with the momentum equation; training the deep neural network with a dataset constructed from historical numerical weather prediction model simulation results; linearizing the trained deep neural network to obtain a corresponding tangent linear operator and an adjoint operator, and embedding the tangent linear operator and the adjoint operator into the variational assimilation framework cost function; obtaining multi-source remote sensing observation data and a numerical weather prediction model background field, and solving an analysis field by taking minimization of the cost function as an objective, to complete data assimilation. The application can improve the data assimilation and numerical prediction level of disastrous weather such as typhoon, especially the observation assimilation level related to the boundary layer.

The invention discloses a boundary hierarchical grid physical constraint variation assimilation method embedded with a deep neural network, which comprises the following steps: establishing a momentum equation containing a boundary hierarchical grid turbulence friction item which is simulated through the deep neural network; constructing a weak constraint term of a variational assimilation framework cost function by using the momentum equation; training the deep neural network by using a data set constructed by a historical numerical weather forecast mode simulation result; linearizing the trained deep neural network to obtain a corresponding tangent linear operator and an adjoint operator, and embedding the tangent linear operator and the adjoint operator into the variational assimilation framework cost function; and acquiring multi-source remote sensing observation data and a numerical weather forecast mode background field, and solving an analysis field by taking the cost function minimization as a target to complete data assimilation. According to the method, the data assimilation and numerical forecasting level of disastrous weather such as typhoon can be improved, and particularly the observation assimilation level related to a boundary layer can be improved.

The invention discloses a non-linear operator calculation method and system for privacy protection machine learning, and provides efficient and safe basic support for private calculation of operations such as a non-linear activation function in a machine learning model on the premise that multi-party cooperative calculation is performed and a plurality of participants do not expose local data. According to the framework, additive secret sharing and mask secret sharing are combined, and an efficient sharing conversion protocol is constructed and used for supporting conversion operation between different sharing types. Furthermore, the invention provides a series of sub-protocols such as security replacement, security comparison and most significant bit (MSB) extraction, and lays a key foundation for subsequent construction of security calculation protocols of non-linear operators such as ReLU, DReLU, MaxPool and the like.

The invention relates to a stability analysis method and system for a boundary layer of a high-enthalpy ablated wall surface, and is applied to the technical field of aerospace, and the method comprises the steps: taking a Landau-Teller equation and a chemical reaction rate equation as source items, constructing a control equation, and solving the control equation to obtain a steady laminar flow field; analyzing the steady laminar flow field by using LST to obtain a transmission coefficient eigenvalue; if the first modal time scale and the second modal time scale are greater than or equal to 1, correcting the first modal time scale and the second modal time scale based on the roughness of each position of the wall surface of the high-speed aircraft, and constructing a first modal linear operator and a second modal linear operator; further constructing NPSE, and solving to obtain an NPSE result; when an NPSE result is decomposed and linearized according to Fourier subharmonics, a coupling source item of a subharmonic equation is constructed according to a second modal linear operator, secondary instability analysis is carried out, and a transition initial position and the amplitude and frequency of disturbance development are obtained. And the accuracy of stability analysis can be improved.

A system and method are provided for training a neural network to control the operation of a system having nonlinear dynamics represented by a partial differential equation (PDE). The method includes collecting digital representations of time series data representing instances of the system's function space and measurements of the system's state of operation. Co-location points corresponding to solutions to the PDE are generated. The neural network is trained using training data including the collected time series data and the co-location points to train parameters of a nonlinear operator. The neural network has an autoencoder architecture including an encoder and a decoder. The encoder encodes each instance of the training data into a latent space, a linear operator propagates the encoded instance into the latent space using a transformation determined by the parameters of the nonlinear operator, and the decoder decodes the transformed encoded instances of the training data to minimize a hybrid loss function.

The present application belongs to the field of integrated circuit manufacturing online measurement, and discloses a snapshot overlay error measurement method and system, which comprises the following steps: measuring a measured object, obtaining two measurement spectra with positive and negative preset deviations respectively, coherently demodulating the measurement spectra, and shifting the specific frequency channel to the zero frequency channel; processing the spectrum data after frequency shifting by using a linear operator to obtain the coefficients corresponding to the positive and negative preset deviations; constructing a characteristic quantity according to the linear combination of the real part and the imaginary part of the coefficients, then calculating the corresponding characteristic quantity according to the coefficients corresponding to the positive and negative preset deviations; and calculating the overlay error according to the characteristic quantity based on the linear relationship between the characteristic quantity and the overlay error. The present application can solve the overlay error with multi-wavelength coupling without using traditional Fourier analysis and truncation operation, has high precision and robustness to noise, and can be used for data processing of snapshot overlay error measurement in multiple scenes.

The application belongs to the technical field of artificial intelligence, and particularly relates to a pure integer quantization method of a visual Transformer model and a related device; the pure integer quantization method of the visual Transformer model comprises the following steps: based on a selected visual Transformer model, linear components and nonlinear components are respectively quantized and integrated to obtain a pure integer quantization model; when the nonlinear components are quantized, integer approximation algorithms are used to respectively reconstruct integer Softmax, GELU and LayerNorm nonlinear operators to obtain reconstructed integer Softmax, GELU and LayerNorm operators; a tensor virtual machine compiler framework is used to perform operator packaging and computation graph optimization on the pure integer quantization model to obtain an optimized pure integer computation graph. The technical scheme disclosed by the application realizes full integer calculation and efficient deployment of the ViT model on the FPGA end side by reconstructing integer operators of the three types of nonlinear operators, namely Softmax, GELU and LayerNorm, and combining computation graph scheduling optimization.

This specification provides a method and apparatus for compiling a plaintext processing program into a ciphertext processing program. The method includes: obtaining a plaintext processing program, which includes multiple operators; converting each operator in the multiple operators into a corresponding homomorphic operation to obtain a sequence of homomorphic operations; wherein, when any operator is a linear operator, it is directly converted into a corresponding homomorphic operation; when the operator is a nonlinear operator, a target polynomial is determined by polynomial fitting, and the operator is converted into a homomorphic operation corresponding to the target polynomial; traversing each homomorphic operation in the sequence of homomorphic operations, and calling the fully homomorphic encryption function corresponding to the homomorphic operation to obtain the ciphertext processing program. This method can improve the development efficiency of ciphertext processing programs.

The disclosure provides a computing system, a data processing method, equipment and a medium for high-bandwidth inference, and relates to the technical field of integrated circuits. The computing system is used for performing a decoding stage of a Transformer-based model inference, and comprises a host processor for performing a decoding stage nonlinear operator, an offload subsystem for performing at least a part of a decoding stage linear operator, and a standard high-speed interface module for transmitting an input activation vector and an output result vector between the host processor and the offload subsystem. The offload subsystem comprises a weight lock storage array for storing a weight matrix in a static residence manner, an input vector streaming interface for streamingly receiving the input activation vector, a matrix vector multiplication calculation unit for performing a matrix vector multiplication operation on the input activation vector and the weight matrix, and a result processing module for reducing or arranging the operation result to obtain the output result vector.

Aspects of the techniques may include generating (910), by a processor, a non-linear output of a layer of a first model, the first model including a first neural network (310), the layer of the first model including non-linear operators and corresponding to a substance distribution, generating (920), by the processor, a linear output based on a layer of a second model comprising a second neural network (320) and a non-linear output, the layer of the second model comprising linear operators and corresponding to a plurality of beams respectively configured to generate radiation, outputting (930), by the processor and based on a linear response, an indication of an energy distribution of the plurality of beam outputs, the processor is configured to output energy to correspond to the substance distribution, and cause (940), by the processor, one or more of the plurality of beams to output radiation in accordance with the output energy distribution.

The application discloses a neural network operator design method and device for aircraft flow field modeling, equipment and medium, relates to the field of operator design, and comprises the following steps: extracting a linear operator to be replaced from a target solution model suitable for an aircraft flow field modeling scene; based on the obtained operator list to be replaced and operator constraint information, a black box operator network structure is constructed by using a tensor primitive library; the obtained operator generation result and quantum annealing algorithm are used to iteratively optimize the black box operator generation parameters to obtain a candidate black box operator; the candidate black box operator is evaluated by using a preset precision evaluation index and a preset performance evaluation index, and a target deep neural network operator suitable for the aircraft flow field modeling scene is determined; and the target deep neural network operator is used to update the target solution model, so that an aircraft flow field modeling operation is performed based on the updated target solution model. The application can automatically generate an operator with controllable precision, meet the design rules and requirements, and improve the aircraft flow field modeling effect.

The application relates to the technical field of numerical weather prediction, in particular to a layered self-adaptive Gaussian conversion method and equipment for cloud and rain control variables, which can acquire observation data and input the observation data into a numerical prediction model to obtain three-dimensional cloud and rain control variables; a variational assimilation model including a forward transformation operator, an inverse transformation operator, a tangent linear operator and an adjoint operator is constructed; the forward transformation operator can calculate and adjust conversion strength layer by layer, realizing more fine and accurate Gaussian processing; the tangent linear operator and the adjoint operator are used for internal minimum iteration calculation of the assimilation model, and then the inverse transformation operator is used to restore the three-dimensional cloud and rain control variable analysis field to its original physical order of magnitude output, so that the actual three-dimensional cloud and rain control variable analysis field is obtained. According to the technical scheme, the vertical distribution difference of the cloud and rain variables can be accurately adapted, and through the provision of a complete operator chain, the compatibility with an existing assimilation system is ensured, so that the analysis field quality and the prediction ability are improved.

In one aspect, the disclosure relates to a method of clock recovery. The method includes sampling one or more signals at a symbol rate to generate a set of samples; zero-padding one or more samples in the set of samples; filtering one or more zero-padded signals with a filter to generate a set of filtered samples; and modifying the set of filtered samples by applying a non-linear operator to generate an output comprising a spectral correlation.