512 results about "Forward propagation" patented technology

The invention is applied to the field of mechanical fault diagnosis signal processing, and particularly provides a pulse neural network mechanical fault diagnosis method based on STFT dimension transformation, and the method comprises the steps: collecting a one-dimensional mechanical vibration signal, carrying out the wavelet decomposition, carrying out the wavelet reconstruction of a low-frequency component and a denoised high-frequency component, and carrying out the wavelet reconstruction of the low-frequency component and the denoised high-frequency component; obtaining a denoised one-dimensional vibration signal; performing short-time Fourier transform, and converting the time-frequency two-dimensional matrix into a time-frequency two-dimensional matrix; inputting the time-frequency two-dimensional matrix into an improved HH threshold neuron model, carrying out Poisson sparse coding on the time-frequency two-dimensional matrix, and only carrying out pulse response on signal significant features; constructing a suprathreshold coding convolutional network with residual connection, inputting a sparse coding matrix, training by adopting an unsupervised learning rule based on STDP, and adaptively adjusting a network synaptic weight; and inputting to a trained above-threshold coding convolutional network, and obtaining pulse emission activity of neurons of an output layer through network forward propagation to determine a fault diagnosis result.

A set of features of a training document are identified in a training document for training a machine learning model. A subset of the features is selected to be omitted from a training forward propagation. As a result of omitting the subset of the set of features, a different subset of the set of features is used to train the machine learning model to classify documents and distinguish between an out-of-domain document and in-domain document.

The invention belongs to the field of image processing, and discloses a medical image segmentation method based on multi-branch distillation, which comprises the following steps: constructing and training a multi-branch collaborative distillation image segmentation model based on uncertainty perception, and establishing a network architecture comprising a public encoder, a first decoder and a second decoder, including a guiding decoder and a guided decoder; dropout disturbance is differentiated on the output of the public encoder, multiple features are generated, multi-branch forward propagation is executed, and a virtual label is generated through fusion of a CTF module; calculating supervision loss and distribution alignment loss of the original feature map through a guided branch; performing multi-level consistency constraint on the output of the disturbance characteristic graph by the guided model to realize multi-view structure consistency, automatically balancing parameter weights through an adaptive task balancing mechanism, and constructing a total loss function; and obtaining a to-be-segmented medical image, and inputting the to-be-segmented medical image into the trained image segmentation model to obtain a segmentation result. According to the invention, high-quality medical image segmentation under a low marking rate can be realized.

The embodiment of the invention provides a model reinforcement learning method, device and equipment. The scheme comprises the following steps: in a sampling stage, using an inference engine and adopting a to-be-trained target model to generate an output sequence for an input sequence under a first strategy parameter, recording a first probability value of each lexical element in the output sequence, and calculating a dominant value of each lexical element, in a training stage, after a training engine is used for forward propagation to obtain a second probability value of each lexical element generated by a target model under a first strategy parameter, a first ratio of the second probability value to the first probability value of each lexical element can be calculated, and then the lexical elements with the first ratios within a preset numerical range are screened out to participate in calculation of a target function; and optimizing the target function to update the parameters of the target model.

The invention relates to a quantitative perception training method and device of a neural network model, electronic equipment and a storage medium. The method comprises the steps of obtaining a to-be-trained first neural network model; an operator pair and a non-module operator in the first neural network model are identified, the non-module operator is an operation which is not realized based on a module class, and the operator pair comprises a convolution operator and a batch normalization operator which are connected; the non-module operators are packaged into module operators, the operator pairs are packaged into new convolution operators, a second neural network model is obtained, the new convolution operators run the fusion process of the convolution operators and batch normalization operators during forward propagation, and the parameters of the convolution operators and the batch normalization operators are updated during back propagation; and performing quantitative perception training on the second neural network model. By adopting the method, the reasoning precision of the quantitative model can be guaranteed, and the deployment efficiency of the quantitative model is improved.

The invention discloses a large model deployment method based on heterogeneous resource scheduling and multi-model collaborative reasoning, and relates to the field of artificial intelligence and server system deployment. Comprising the following steps of: 1, calculating a deployment score of a target model by utilizing a deployment scoring function according to a GPU video memory residual rate, a CPU load, network delay and a current task quantity of each node through a heterogeneous resource sensing scheduler, and selecting a node with the highest score of the target model to load or reuse the target model; a cache index key is generated, and whether key value KV pairs supporting multiplexing of the target model exist or not is inquired; step 3, if the cache is hit, skipping a forward propagation stage, and directly decoding based on a cache result, otherwise, executing complete forward reasoning and writing the result into the cache; 4, based on a plurality of user requests, performing fusion according to user priorities, model weights, reasoning costs and cue word lengths to generate execution sorting weights, and scheduling execution models in batches according to the weights; and 5, returning the model response to the user, and releasing or updating node state information.

The invention discloses a private data security sharing method based on federal learning, particularly relates to the field of private data security protection and sharing, and is used for solving the problems of insufficient model credibility and lack of precision control of data exchange in the existing cross-mechanism data collaboration process. According to the method, homomorphic encryption processing is carried out on local data of all participants, a ciphertext verifiable calculation task is constructed in combination with federated learning, forward propagation and back propagation calculation of a ciphertext are carried out by outsourcing calculation nodes, credible verification is carried out on a calculation process by utilizing zero-knowledge proof, and gradient aggregation and model updating are completed in a ciphertext domain; after model training is completed, based on federal learning model output, intelligent judgment is conducted on the sharing value of local data samples, ciphertext re-encryption and data exchange control are driven, multi-party safe and controllable data circulation is achieved, and therefore on the premise that privacy safety and compliance requirements are guaranteed, cross-mechanism data collaboration efficiency and decision-making precision are improved.

The embodiment of the invention provides a dynamic multifunctional optical metasurface design method and system, and the method comprises the steps: obtaining function demands in different application scenes, and setting a target reflection spectrum according to the function demands; the target reflection spectrum is input into a reverse retrieval network, and super-unit structure prediction design parameters are obtained through encoder compression and decoder parameter prediction; the super-unit structure prediction design parameters are input into a forward prediction network, and a prediction reflection spectrum is obtained through network forward propagation calculation; based on the matching degree between the predicted reflection spectrum and the target reflection spectrum, establishing a bidirectional relation between a design domain and a physical domain by using a phase recovery algorithm, and predicting design parameters by continuously adjusting the super-unit structure, so that the predicted reflection spectrum gradually approaches the target reflection spectrum, and obtaining ideal design parameters of the super-unit structure; and according to ideal design parameters of the super-unit structure, super-units are arranged in an operation space, and the super-surface structure capable of realizing dynamic multifunctional regulation and control is formed.

A multi-modal large model illusion detection and inhibition method based on attention time sequence difference comprises the following steps: inputting text lexical elements of cue words and visual lexical elements of images into a multi-modal large model, and obtaining an internal attention graph of the decoding stage of the multi-modal large model; then calculating the attention proportion of the visual lexical units at the current generation moment, and making a difference between the attention proportion and the proportion at the previous moment; and if the difference value exceeds a set threshold value, determining that the lexical elements are visual related lexical elements. When the visual related lexical elements are recognized, performing secondary forward propagation of primary visual enhancement to obtain more accurate output; and if not, directly entering the next step of generation. According to the method, visual related lexical elements in text generation are recognized and refined through attention time sequence difference, two-time forward propagation is adopted, the visual attention of second-time forward propagation is enhanced based on a visual attention graph of first-time forward propagation, and illusion can be recognized and corrected on the premise that the language expression ability is not reduced.

The invention discloses a model illusion detection method and device based on internal state fusion and a medium, and relates to the technical field of natural language processing. The method comprises the following steps: extracting multi-modal features in a forward propagation process of a target large language model, wherein the multi-modal features comprise a hidden layer embedding feature, an attention feature, a feedforward network activation feature and a text feature; aligning the multi-modal features to the lexical element length of the generated text through an interpolation method, and calculating the weight of the position of the lexical element corresponding to the multi-modal features; performing weighted fusion on the multi-modal features according to the weights to generate a fusion feature sequence; and constructing the fusion feature sequence into a graph structure, reasoning the graph structure by using a multi-layer attention network, and outputting a lexical-level illusion probability through a classification head. According to the method, the target model parameters do not need to be modified, and high-precision detection can be completed only through single-time forward propagation.

This application discloses a model training method and apparatus based on a hybrid parallelism manner. In this method, a neural network model is divided into a plurality of pipeline stages, and each pipeline stage includes a plurality of sub-stages of the neural network model. Computing nodes corresponding to the plurality of pipeline stages are invoked in a hybrid parallelism manner according to a sequence of sub-stages in the neural network model. When iterative training is performed on a network layer in a corresponding pipeline stage, because sub-stages at same locations in adjacent pipeline stages are consecutive in the neural network model, the computing node does not need to wait for completion of forward propagation of a previous pipeline stage, and can perform forward propagation on the corresponding pipeline stage only after forward propagation of the 1st sub-stage in the previous pipeline stage is completed.

The invention discloses a cross-modal attention collaborative jail break attack method, and belongs to the technical field of artificial intelligence security. The method comprises the following steps of: constructing an input sequence representation containing a system prompt, an adversarial image representation, a malicious query and an adversarial text suffix according to a causal self-attention mechanism; inputting the sequence into a visual language model to execute forward propagation; based on the designed attention-oriented loss collaborative function, optimizing adversarial image representation through a joint gradient optimization algorithm and updating an adversarial text suffix to optimize an attack target; iteratively circulating until convergence, and outputting the optimized unified multi-modal knowledge; and finally, utilizing the knowledge to construct an attack sequence to realize jailbreak. According to the method, accurate control on an internal attention mechanism of the visual language model is realized for the first time, and through visual-text dual-mode cooperative attack, the attack success rate is remarkably improved while high concealment is kept, and the important driving force for promoting the progress of a safe alignment technology is achieved.

The invention provides a model training method and device, equipment and a storage medium. The method comprises the following steps: determining a first optimizer state parameter at a current time step according to a loss function value obtained by forward propagation calculation of a neural network model; the first optimizer state parameter, the second optimizer state parameter and the loss function value have a first corresponding relation; substituting the first optimizer state parameter and the loss function value into a first corresponding relation in response to the detection that the training of the current time step meets a preset training abnormal condition, so as to obtain a second optimizer state parameter; substituting the second optimizer state parameter into the loss function, and updating a loss function value; substituting the updated loss function value and the second optimizer state parameter into the first corresponding relation, and determining a third optimizer state parameter; and using the third optimizer state parameter to execute the training of the next time step on the neural network model.

The invention discloses a method and a system for realizing remote operation and maintenance of equipment through cooperation of an industrial intelligent gateway and a cloud platform, and relates to the technical field of remote operation and maintenance of industrial Internet of Things, and the method comprises the following steps: the gateway collects multi-modal data of the equipment and locally preprocesses the multi-modal data to generate a structured data packet; lightweight Transform model forward propagation is executed, and an encrypted feature vector is extracted and uploaded to a cloud platform; the cloud platform adopts federated learning to update model parameters based on the feature vectors, issues the model parameters to the gateway, and synchronously constructs equipment digital twin bodies; analyzing the natural language operation and maintenance instruction through the digital twin to generate a control command; an equipment association graph is constructed, a fault propagation path is analyzed by using a graph convolutional network (GCN), and an early warning is generated; and after executing the control instruction, the gateway feeds back state data to the cloud platform, updates the digital twin and model input, and forms closed-loop control. According to the invention, collaborative optimization of data privacy protection, cross-device fault prediction and intelligent decision is realized, and the operation and maintenance efficiency and security are improved.

The invention relates to the technical field of artificial intelligence, and discloses a Bayesian neural network method fused with uncertainty quantization and a system thereof, the method comprises five steps of probability weight reconstruction modulo, variational posterior inference, re-parameterization sampling, multi-scale uncertainty quantization and adaptive rejection decision, the network weight reconstruction modulo is probability distribution, and the probability distribution of the network weight reconstruction modulo is improved. Sparse induction is realized by adopting scale Gaussian mixture prior, calculation complexity is reduced through a local re-parameterization technique, reasoning time consumption is controlled within two times of a deterministic network, a multi-scale uncertainty quantification module quantifies cognitive uncertainty by counting output variance of multiple forward propagation, and the accuracy of reasoning is improved. According to the method, the technical problem that a deep learning model lacks reliable prediction confidence estimation is effectively solved, and the misdiagnosis sample omission ratio in a medical image classification task is reduced.

The invention discloses a recommendation system-oriented privacy sensitive parameter identification and accurate deletion method, and relates to the technical field of recommendation system model optimization and privacy protection. In order to solve the problems that privacy sensitive parameters in an existing recommendation system model are difficult to accurately position and the model performance loss is too large after deletion, according to the evaluation thought of parameter importance and data feature relevance, importance scores of model parameters on two types of data sets are calculated respectively by dividing a forgetting set and a reservation set, and the privacy sensitive parameters are obtained. The privacy sensitive parameters with the forgetting set importance higher than the reserved set importance are screened out, and mask zero setting processing is carried out. According to the method, the model does not need to be retrained or subjected to weight updating, sensitive parameter recognition and deletion can be completed only through single-time forward propagation, and the original recommendation performance of the model is reserved to the maximum extent while the privacy security of a recommendation system is guaranteed. According to the method, accurate identification of privacy sensitive parameters is realized on a mainstream recommendation model, the reduction range of model recommendation accuracy after deletion is controlled within an acceptable range, and the privacy leakage risk is remarkably reduced.

The invention relates to a freight train future speed prediction method and device based on brain-like acceleration and a medium, the method inputs obtained freight train related data into a speed prediction model, a freight train future speed prediction result is output, the speed prediction model adopts a three-level network structure of an MLP encoder-Transformer encoder-MLP decoder, and the speed prediction model is used for predicting the freight train future speed. After the MLP encoder encodes the two paths of input respectively, the two paths of input are input into the Transform encoder for spatio-temporal feature extraction, decoding is carried out through the MLP decoder to obtain a prediction result, the forward propagation process of the Transform is converted into a brain-like instruction sequence, and the brain-like chip is controlled based on the brain-like instruction sequence to carry out operation so as to realize accelerated prediction. Compared with the prior art, the method has the advantages that the static characteristics can be dynamically processed in a time sequence mode, and the prediction efficiency can be improved by fully utilizing the brain-like chip.

The invention discloses a network security uncertainty quantitative detection method based on generative AI, and belongs to the technical field of network security detection, and the method comprises the steps: collecting security event data from a plurality of heterogeneous data sources, and forming a standardized to-be-detected event sequence; constructing a dynamic baseline model based on the historical event sequence of the normal behavior by using a probability generation model, and generating at least one anomaly indication feature for the to-be-detected event sequence; inputting the anomaly indication features into a Bayesian deep learning model, and obtaining a mean value representing an anomaly probability and a variance representing prediction confidence in the event sequence to be detected through multiple forward propagation sampling; and carrying out risk grading on the security events, and carrying out retraining on the probability generation model and / or the Bayesian deep learning model based on feedback annotation of a grading result to realize closed-loop optimization. According to the invention, the decision-making efficiency is improved in a leap-over manner, and the gray attack behavior is effectively detected and the precision is improved.

The invention discloses a dynamic multimedia data hash retrieval method and system based on extensible increment, and relates to the technical field of multimedia data retrieval. The method comprises the following steps: acquiring dynamic multimedia data to be retrieved; a pre-trained Hash retrieval large model is constructed, the Hash retrieval large model is trained by taking a bit extensible Hash center as global supervision information and taking tag cosine similarity as local supervision information, and specifically, generalization feature representation of new multimedia data maintaining new and old class discrimination is obtained through forward propagation; constructing a linear mapping relationship between the generalization feature representation and the hash code, and introducing an auxiliary variable to continuously update the hash function without playback; and utilizing the trained Hash retrieval large model to generate a query Hash code for the dynamic multimedia data to be retrieved, and utilizing the query Hash code to retrieve. According to the method, low memory occupation, high updating efficiency and non-forgetting retrieval of the dynamic multimedia data stream in the open environment are realized.

The invention discloses an uncertainty guided few-sample harmful speech detection method (U-GIFT). According to the method, a pre-training language model is finely adjusted based on a small number of labeled samples, and a semi-supervised self-training and uncertainty guiding strategy is combined. Monte Carlo Dropout is started in the reasoning stage, multiple times of random forward propagation are carried out to obtain sample posterior distribution, prediction entropy and information gain are calculated, pseudo-label samples are sorted and screened, and only high-confidence samples are selected to be added into a training set. And in order to reduce the influence of a pseudo labeling error, designing a stability weighting mechanism, giving a sample weight according to a prediction variance, and constructing a joint loss function, so that the model preferentially learns a stable sample to improve the detection performance. According to the method, the semantic and attention mechanism of the pre-training model is utilized, the detection effect is remarkably improved under the conditions of few samples, imbalance, multiple languages and cross domains, models such as BERT, RoBERTa, XLM-R, LLaMA2 and DeepSeek-R1 are compatible, and the method is suitable for content auditing and risk prevention and control.

The invention relates to the cross technical field of industrial intelligent simulation and computational material science, in particular to a phase field simulation method and device based on physical constraint deep learning and electronic equipment, and the method comprises the steps: obtaining phase field parameters of a to-be-predicted material, including an initial phase field value, an initial material mobility tensor and an initial free energy density coefficient, the phase field parameters are input into a pre-trained prediction model, so that a phase field simulation result of the to-be-predicted material is obtained, and the pre-trained prediction model is used for predicting the phase field of the to-be-predicted material based on the improved LSTM network, a preset hybrid forward propagation strategy and a preset parameter coupling mechanism. And training a preset neural network by using a mixed loss function and a sub-path gradient back propagation strategy. Therefore, the problems of low phase field simulation calculation efficiency, physics non-conservation of a data driving model and the like in the related technology are solved, physical equation constraints are embedded into the neural network through a hybrid model architecture, and efficient and accurate prediction of material microstructure evolution is realized.

The invention discloses a fault tree Boolean function equivalent mapping method based on an untrained neural network, and relates to the field of fault tree analysis. In order to solve the problems that in the prior art, a Boolean function mapping structure is not beneficial to parallel expansion, the calculation efficiency is limited, and the Boolean function mapping structure is difficult to efficiently realize on high-parallel platforms such as a GPU, the invention provides a method for generating topological structure data by analyzing a fault tree model; the basic events, the intermediate events and the top events are mapped into neurons of an input layer, a hidden layer and an output layer respectively, a feedforward network with fixed weight and bias is constructed, and a logic activation function is defined in nodes to realize Boolean logic propagation. The input layer receives a basic event state vector, outputs a top event result through forward propagation, and can realize large-scale Boolean function mapping on a parallel platform through batch input matrixes. The method is suitable for reliability analysis, minimum cut set simplification, top event probability calculation, parallelization fault tree solving and the like of a large-scale complex system.

Systems and methods are provided to improve the memory throughput for storing and reading intermediate data computed by layers of a neural network during a training process. A compression operation can be performed by removing the zeros from the intermediate data and storing locations of the zeros before storing the intermediate data in the memory for a forward pass of the training process. The compressed data can be read from the memory for a backward pass of the training process and de-compressed by inserting zeros based on the stored locations. Additionally, a transpose operation can be performed before compression as a first atomic operation, or after de-compression as a second atomic operation.

The invention relates to the technical field of artificial intelligence, and provides a large language model reasoning method, device and equipment and a computer storage medium. The method comprises the following steps: acquiring an input sequence, a trained forward prompt embedding vector sequence and a trained backward completion embedding vector sequence; splicing an input embedding vector obtained by processing an input sequence through an embedding layer of a large language model with a forward prompt embedding vector sequence and a backward completion embedding vector sequence to obtain a combined input sequence; inputting the combined input sequence into a large language model, and forwards propagating the combined input sequence to at least one draft generation layer to obtain candidate draft output information corresponding to a forward prompt embedding vector sequence position; the draft generation layer is a Transform layer which is screened out from all Transform layers of a large language model and meets a reasoning precision condition and a reasoning speed condition based on a data set of a target reasoning task; and determining a reasoning result based on the confidence of the candidate draft output information, the reasoning result being the candidate draft output information, or submitting a token sequence obtained by sampling based on the candidate draft output information to all layers of the large language model for verification reasoning.

The invention discloses a space-time decoupling efficient electroencephalogram basic model construction method and device, and relates to the technical field of brain signal processing. The method comprises the following steps: constructing an electroencephalogram basic model based on linear attention and spatio-temporal data decoupling; segmenting the acquired electroencephalogram signal data, inputting the segmented electroencephalogram signal data into an electroencephalogram embedding layer, and outputting an embedding vector; inputting the embedded vector into a space-time mixing module for separation and decoupling to obtain a spatial feature and a time feature; performing coding calculation on the spatial features and the time features based on an enhanced bidirectional weighted key value algorithm of a linear attention mechanism to obtain new spatial hidden layer features and time hidden layer features; mixing the two hidden layer features to obtain spatial-temporal features; inputting the spatial-temporal characteristics into a forward propagation network, and pre-training the model to obtain a trained model; and inputting a downstream task to be executed into the trained model for processing, and outputting a task result. According to the invention, the brain signal processing efficiency can be improved.

The invention relates to a power distribution network fault positioning method combined with multi-terminal traveling wave sampling and a terminal. The method comprises the following steps: synchronously acquiring a transient interference spectrum excited by a network disturbance singular point; normalizing the transient interference spectrum, and extracting a high-dimensional spectrum feature vector from the transient interference spectrum; for each candidate singular point, predicting a theoretical spectrum feature vector of the candidate singular point through a forward propagation operator, and calculating a holographic coherence index between the theoretical spectrum feature vector and an actually observed spectrum feature vector, thereby iteratively reconstructing a singular point holographic data volume representing global fault probability distribution; and carrying out image decoding analysis on the singular point holographic data body, and determining the accurate position of the network disturbance singular point by searching a global maximum value point. According to the method, the fault positioning problem is converted into an image reconstruction and understanding problem, and the positioning precision and robustness are remarkably improved by using the complete information of the traveling wave signal instead of single time information.

The invention discloses a TEE-GPU collaborative model credible training method and device based on parameter confusion, and the method comprises the steps: enabling a client to encrypt training data and a model architecture in a preprocessing stage, and uploading the encrypted training data and model architecture to a server; and the trusted execution environment of the server side decrypts the model architecture, carries out confusion processing on a linear layer and then deploys the linear layer on an external GPU (Graphics Processing Unit). In the training stage, initial forward propagation calculation of training data is firstly completed in the TEE, and then intermediate results are confused and then transmitted to the GPU so as to execute subsequent forward propagation calculation. In the back propagation process, extra confusion is applied to the gradient by the TEE, and the gradient is issued to the GPU to calculate a new gradient; and after receiving the confusion gradient, the GPU updates the parameters in a confusion form. And randomly sampling calculation data in the TEE, and carrying out integrity verification to ensure the integrity of a calculation result. Compared with an existing credible training scheme, the method has the advantages that the time overhead can be remarkably reduced while the calculation accuracy, integrity and privacy of the model are ensured.

The invention discloses a thermodynamic system analysis method and system based on a physical information neural network. The method comprises the following steps: constructing a thermodynamic system mathematical model; converting a solving problem of the mathematical model of the thermodynamic system into an optimization problem taking a minimum control equation set residual error as a target function; constructing a physical information neural network; constructing a composite loss function based on the control equation set residual sum of squares, and performing end-to-end training on the physical information neural network to obtain a trained physical information neural network; taking the trained physical information neural network as an approximate solver, and for any given thermodynamic system input working condition, predicting a group of approximate solutions as initial predicted values through one-time forward propagation; and inputting the initial predicted value into a traditional numerical solver for accurate solving to obtain a to-be-solved state variable. According to the method, a feasible technical path is provided for applications with strict requirements on the analysis and calculation speed of the thermodynamic system, such as real-time simulation, online optimization and digital twinning.

The invention discloses a semiconductor heterojunction interface thermal resistance prediction method based on machine learning, and the method comprises the steps: collecting semiconductor material data, obtaining semiconductor intrinsic attribute data through a semiconductor public database, and carrying out the preprocessing of the data; calculating statistics of element attributes through a Magpie algorithm to obtain feature descriptors, eliminating redundant features by adopting a variance filtering method and recursive feature elimination, and normalizing the redundant features; model building and training are carried out by designing CNN and XGBoost algorithms; performing parameter optimization on the trained model through forward propagation, back propagation and a DBO algorithm; and predicting the thermal resistance of the semiconductor heterojunction interface based on the optimized model. The method solves the problems that in an existing semiconductor heterojunction interface thermal resistance prediction method, the experimental measurement period is long, the cost is high, environmental parameters are difficult to control accurately, the theoretical calculation complexity is high, and the deviation between a prediction result and an actual working condition is large due to the dependence on ideal interface conditions.

The present application relates to a kind of neural network global one-time structured pruning method, system, equipment and medium, wherein, method includes: by single forward propagation, in the neural network to be pruned using calibration dataset and parallelly capturing the input activation tensor of all target layers;According to input activation tensor, the intermediate neuron weight of each target layer is synchronously calculated differential entropy index and amplitude response intensity, and after normalization and fusion, static global importance atlas is formed;According to the importance threshold value determined according to preset pruning rate, based on global importance atlas, one-time generates the index set of all global pruning to be pruned whose mixed importance score is below importance threshold value;Based on index set, the weight matrix of each target layer is executed one-time physical structured pruning.By doing so, the present application generates static global importance atlas by single forward propagation and parallelly capturing activation tensor, and completes no-mask one-time pruning by physical structured pruning.