1791 results about "Hidden layer" patented technology

Provided in the present invention is a method for analyzing the vibration of a fluid conveying pipe on the basis of a Fourier featured PINN. The method comprises: constructing a physics-informed neural network (PINN), determining initial conditions and boundary conditions on the basis of a fluid conveying pipe model, and obtaining loss functions corresponding to the initial conditions and the boundary conditions; using a "soft" constraint to express a second-type boundary condition and initial excitation of a pipe into PINN loss functions, and using a "hard" constraint to code a first-type boundary condition and an initial amplitude into a DNN system structure; adding an observation anchor point to the network for correction when a peak value is insufficient during a network model training process; constructing a Fourier feature mapping function, adjusting a neural tangent kernel of the network, training two mapping results by means of a set network, and finally, in the network, connecting spatial and temporal hidden layers by means of point multiplication, so as to establish an FFNN; and by means of training and learning, finally outputting a prediction result by means of a linear layer connection.

Systems and methods for adaptive edge computing using artificial intelligence (AI) include monitoring real-time accuracy of a neural network by using a feedback loop configured to detect changes in inference accuracy and dynamically adjusting the structure of the neural network by adding or removing hidden layers based on monitored error rates and predetermined computational constraints. A Kalman gain computation determines neural network weight adjustments based on monitored error rates. Weight matrices undergo incremental updates derived from these adjustments. Incremental weight adjustments remain stored in memory to enable low-bandwidth model updates. The neural network stores inference results and refined weights in an inference result database. Pre-trained models periodically receive incremental updates based on stored adjustments. Predictive holistic inference logic (PHIL) applied to stored inference results improves the accuracy of the inference results.

The invention provides an intelligent incubation bin anti-interference control method and system based on error self-learning, and relates to the technical field of intelligent control, and the method comprises the steps: respectively calculating a temperature deviation value and a humidity deviation value according to an optimization parameter sequence of a temperature prediction error and an optimization parameter sequence of a humidity prediction error; performing normalized weighted summation on the temperature deviation value and the humidity deviation value, and performing dynamic scaling through a Gaussian kernel function to obtain a correction factor; dynamically adjusting the activation function slope of a neural network hidden layer according to the correction factor and a BP neural network, reconstructing the weight of the fuzzy rule base based on the numerical distribution characteristics of the correction factor, and generating the corrected weight of the fuzzy rule base; and based on the corrected fuzzy rule base weight, constructing a three-dimensional parameter adjustment curved surface, and dynamically adjusting proportion, integral and differential parameters of a PID controller through a curved surface gradient search algorithm to generate a control signal. According to the invention, the control accuracy is improved.

The invention discloses a hybrid expert network training method, device and equipment and a computer program product, experts in a hybrid expert layer are equally divided into a plurality of groups and equally distributed to a plurality of pieces of equipment, and different pieces of equipment share the same routing module. In the training process, for an input hidden layer sequence of the mixed expert layer, the score of each path of experts is calculated through a routing module, for each expert group, activation experts in the group are selected according to the scores of the experts in the group, the number of the activation experts in different groups is the same, and each expert group processes the complete hidden layer sequence. And processing the input hidden layer sequence by the activation experts in each expert group to obtain the output of the activation experts, and obtaining the output of the hybrid expert layer based on the output of each activation expert. According to the method and the device, absolute balance of calculation among the devices is ensured, and the effect of the model is not influenced by drop-pad (namely data discarding and 0 filling).

The present disclosure relates to the field of biological information, and in particular to a gene regulatory network inference method based on spatiotemporal transcriptomic data, and the use thereof. In the present disclosure, on the basis of a graph neural network, a graph variational auto-encoder is used, an adjacency matrix A is used as node connection information, cells and genes are used as node feature information, and position information of the cells is fully used, so as to reconstruct a feature matrix of nodes by means of three hidden layers, thereby performing more accurate cell clustering and dimensionality reduction analysis while inferring a gene regulatory network.

The invention relates to the technical field of model construction, and discloses a neural network-based frozen soil unfrozen water content detection model construction method and system. The method comprises the following steps: carrying out segmented sampling test on a clay sample in a temperature interval of 0 DEG C to-20 DEG C to obtain a multi-dimensional original data set, and carrying out standardization and principal component analysis on the data set to obtain a standardized feature vector; and performing three-stage classification marking on the feature vectors to obtain a training data matrix. A deep neural network is adopted to carry out five-fold cross validation training, and a prediction model of three hidden layers is constructed. And performing correction calculation on model output according to soil mineral composition and porosity parameters, and finally identifying an abnormal value through root-mean-square error analysis and a three-time standard deviation principle to obtain a standardized prediction result. According to the method, the mapping relation between the thermal parameters and the unfrozen water content is established, correction is carried out in combination with the soil body characteristic parameters, and the technical problem that the prediction precision of the unfrozen water content is not high in the prior art can be solved.

Disclosed in embodiments of the present description are a transfer learning model training method and device, a storage medium, and a terminal. The method comprises: at a target transfer moment, initializing a source domain model and a target domain model which satisfy a preset similarity condition; on the basis of parameter transfer networks corresponding to target domain hidden layers in the target domain model, obtaining target source domain hidden parameters in the source domain model; and transferring the target source domain hidden parameters to obtain a target domain transfer model, and training the target domain transfer model on the basis of target sample data corresponding to a target domain. Target source domain hidden parameters reflecting deep network features are selected from source domain hidden layers of the source domain model by means of the parameter transfer networks and transferred to corresponding target domain hidden layers of the target domain model, so that the target domain hidden layers directly obtain effective deep feature parameters, and the target domain transfer model is still trained by using the target sample data of the target domain.

The invention discloses a quasi-brittle material damage field inversion method based on a physical information neural network. The method comprises the steps of data acquisition and processing; presetting initial damage field data, and taking the modulus of each unit in the initial damage field data as an independent to-be-inverted parameter; taking the position information of each node in the space as the input of a neural network, processing through a hidden layer of the neural network, and taking displacement data corresponding to the position information as network output data; constructing a loss function of a neural network according to the preprocessed displacement data, network output data and a mechanical law; and performing training optimization on the initial damage field data by using a loss function of a minimized neural network, optimizing hyper-parameters of the neural network by adjusting weights of control item loss, boundary item loss and data item loss until a preset convergence condition is met, and determining target damage field data. According to the method, the dependence of a neural network model on a data set is remarkably reduced, and the inversion result is ensured to have relatively high physical interpretability.

The invention relates to the technical field of air conditioning, in particular to a laboratory heating and ventilation load prediction and self-adaptive regulation and control method. According to the method, the infrared frame and the power sampling time mark are synchronized, the sensing flow is aligned and packaged, and the thermal diffusion evolution rate is generated. Lagging characteristics are determined in combination with power jump and temperature rise moments, and a heterogeneous dynamic coupling model is established. Extracting a physical evolution parameter as a mechanism operator, injecting the mechanism operator into a hidden layer of the prediction model, reconstructing a phase space, and calculating an air load increment in a lag window. And reverse mapping is executed based on the heat exchange characteristics to generate a feedforward instruction, and when the rate exceeds a threshold value, the weight is issued and dynamically corrected, so that closed-loop correction is completed. According to the method, deep coupling of feedforward prediction compensation and feedback residual adjustment is executed, and accurate regulation and control of the air volume and cooling and heating loads of the laboratory are achieved.

The invention relates to the technical field of piston ring group wear monitoring, in particular to a piston ring group wear life prediction method and system. The method comprises the following steps: collecting working state data of a piston ring group in real time, de-noising by using improved wavelet transform, processing missing data by using a KNN interpolation method, standardizing and normalizing the processed data, adopting a filtering label generation and auditing mechanism to ensure the ordered homogeneity of the data, and constructing an over-limit learning machine model to obtain a piston ring group model; a hidden layer weight and an offset vector are randomly generated to calculate a hidden layer output matrix, a whale optimization algorithm is adopted to optimize model parameters, the prediction precision is improved, then the residual life is predicted, a final prediction result is used for calculating a fault risk coefficient and comparing the fault risk coefficient with a preset threshold value, maintenance suggestions are provided, and real-time data and the prediction result are displayed. And a user is assisted to make a maintenance decision. The advanced signal processing technology and the machine learning algorithm are combined, and the maintenance efficiency of the piston ring set and the reliability of equipment operation are improved.

The invention relates to the field of machine learning, and discloses a charging pile fault diagnosis method, which comprises the following steps: S1, preparing an environment, building a deep learning development environment, and providing a necessary technical environment for the training and deployment of a subsequent model; and S2, data acquisition and processing: installing a current sensor, a voltage sensor and a temperature sensor on a new energy automobile charging detection vehicle, and acquiring operation data of the charging pile in real time. And S3, model definition: a fault diagnosis model based on deep learning is constructed, a fully connected neural network (DNN) is adopted, and the model comprises an input layer, a hidden layer and an output layer. The key operation data of the charging pile are collected in real time through the sensor, the fault state is intelligently judged in combination with the deep learning model, rapid and accurate fault positioning and diagnosis are achieved, and the maintenance efficiency of the charging pile is remarkably improved. The system can automatically extract data features and analyze fault types, and manual intervention is reduced.

The invention relates to the technical field of data processing, and discloses a CNN-BiLSTM-MATT time sequence prediction method, which comprises a plurality of feature value inputs and a feature value output. Carrying out data normalization processing; performing normalization processing on the input data, mapping a normalized data set between [-1, 1], and performing reverse normalization processing on the output data after prediction; the hyper-parameters in CNN-BiLSTM-MATT are optimized, the hyper-parameters in CNN-BiLSTM are initialized according to an IHO algorithm, the hyper-parameters comprise the learning rate, the L2 regularization coefficient and the hidden layer number, and the optimal hyper-parameter value is obtained in the iteration process; performing feature extraction: performing convolutional feature extraction by using a convolutional neural network CNN, and taking obtained data features as input of subsequent BiLSTM; and carrying out time sequence analysis. The method overcomes the defects of a current time sequence prediction method. The method can be widely applied to the parameter optimization problem, the convergence degree is improved, and the optimal solution is efficiently found.

The invention relates to a mechanism data collaboratively driven dam safety state intelligent diagnosis method and system. The method comprises the steps of collecting environmental quantity data, dam deformation data and corresponding monitoring point coordinates; a PINN model is constructed, the PINN model comprises an input layer, a hidden layer and an output layer, the input layer receives the collected data, and the output layer outputs a predicted value of a deformation component; the PINN model is trained, the weight and bias of the PINN model are adjusted by minimizing a comprehensive loss function, and the comprehensive loss function comprises a data driving loss function, a physical constraint loss function and a boundary condition loss function; using the trained PINN model to predict a deformation field; and comparing the prediction result with the monitoring index to obtain an abnormal region, and performing abnormality diagnosis on the abnormal region. Through the mechanism data synergistic effect, it is guaranteed that model prediction follows the physical law, actual monitoring data can be better fitted, and high-precision dam safety state prediction and diagnosis are achieved.

The invention discloses an intelligent blasting sequence control system for mixed loading explosives, which is provided with a data acquisition module, a data preprocessing module, a model construction module, a parameter optimization module, a multi-target control module and a data storage module to intelligently control the blasting sequence of the mixed loading explosives. Distance attenuation features, delay features, topographic features and blasting energy features of blasting data are constructed, the blasting prediction model is updated and iterated based on a neural network algorithm and historical blasting data information, and the performance of the blasting prediction model is evaluated. The number of hidden layers and the number of nodes in the blasting prediction model are updated and optimized through a control variable method, the blasting prediction model is optimized, when mixed explosive is adjusted and the blasting sequence is optimized, the optimal mixed explosive composition is searched through a search optimization algorithm, and the optimal mixed explosive composition is obtained through control signals in combination with real-time blasting data information. And the blasting sequence and delay are dynamically adjusted, so that the blasting of the mixed explosive is intelligently and accurately controlled in real time.

The invention discloses a marine environment forecasting method based on combination of machine learning and numerical forecasting, and particularly relates to the field of marine environment forecasting, which comprises the following steps: extracting marginal region features based on real-time observation data, constructing a forecasting field through a double-loss function, extracting extreme event features through space-time decoupling, and constructing a forecasting field through a double-loss function; and outputting a forecasting result of the target sea area. According to the marine environment forecasting method based on the combination of machine learning and numerical forecasting, the problem of initial field optimization failure caused by covariance matrix estimation deviation when marginal sea area data are sparse is relieved by embedding vorticity conservation equation residual calculation in a hidden layer; the dependence on the number of extreme event samples is reduced by realizing the targeted extraction of the extreme ocean event features; through a physical hard constraint layer technology of a residual error correction network, an initial field error of a data sparse region is corrected in combination with sea surface height and a flow velocity field constraint residual error feature map, and the problem of a large prediction error of an edge sea area is relieved.

The invention relates to the technical field of artificial intelligence, and discloses a geological mineral resource prediction model training method based on artificial intelligence, and the method comprises the steps: carrying out the expansion of geological mineral resource data through a data expansion model, and obtaining a training data set; building a prediction model through an extreme learning machine based on conjugate gradient, and inputting the geological mineral resource data in the training data set into the prediction model for training; randomly initializing hidden layer parameters of the extreme learning machine, and calculating a hidden layer output matrix; calculating an output layer weight of the extreme learning machine through a conjugate gradient method; and updating the hidden layer output matrix. According to the method, a belief propagation mechanism is adopted, the internal relevance of geological mineral resource data can be simulated, and samples are adjusted and generated through a covariance matrix, so that the method is more in line with the distribution characteristics of actual geological mineral resources; a dynamic constraint game strategy is adopted, and strategies of a generator and a discriminator are dynamically adjusted, so that adaptability to complex data is enhanced.

The invention discloses a backfill soil foundation compaction quality on-site detection method and system. The method comprises the steps that instantaneous acceleration signals of a vibrating wheel in the soil compaction construction process are collected in real time through a three-axis acceleration sensor; the acceleration signal comprises a three-axis acceleration component; recording the number of construction times through construction equipment, collecting position information of a construction area by using a GPS (Global Positioning System), and synchronously binding the acceleration signal with the number of construction times and the position information through timestamps to form a data set corresponding to the position of the construction area; performing spectral analysis on the acceleration signal in the data set, and extracting a dominant frequency amplitude and a first harmonic amplitude to generate a filtered acceleration signal; the filtered acceleration signals are input into a training single-hidden-layer BP neural network, and the soil dry density is calculated in combination with the number of construction times; based on the ratio of the calculated soil dry density to the standard maximum dry density, the soil compactness in the construction area is determined in combination with the position information, and a construction area compactness distribution diagram is generated through the upper computer to optimize the construction strategy.

The invention belongs to the field of image super-resolution reconstruction, and provides an image super-resolution reconstruction method and system based on mixed experts and stable diffusion in order to solve the problem of poor super-resolution reconstruction effect at present. The image super-resolution reconstruction method based on mixed experts and stable diffusion comprises the following steps: obtaining a multi-scale control condition and a space control condition; generating noise adding feature hidden layer space representation; after a plurality of sampling time steps, continuously subtracting noise predicted by the denoising backbone network of the current time step from the input of the current sampling time step of the denoising backbone network, then taking a denoising result as the input of the denoising backbone network of the next time step, and after the plurality of sampling time steps are completed, obtaining the hidden layer space expression of the reconstructed image; and the hidden layer space expression of the reconstructed image is decoded to the image pixel space, so that an image super-resolution reconstruction result is obtained, and a real and clear super-resolution reconstruction effect is improved.

The invention discloses a big data prediction system based on machine learning, and relates to the technical field of machine learning and big data, and the system comprises a data collection module, a data preprocessing module, a feature engineering module, a machine learning prediction model module, a model training module and a prediction output module. According to the method, a specific algorithm is applied through the feature engineering module, correlation is considered, a penalty factor is introduced, related feature subsets are focused to improve precision, a hidden layer and parameters of the model module can be adaptively adjusted, for example, high precision is obtained in power load prediction, the data acquisition module updates and processes a format in real time, and data timeliness and availability are ensured. The preprocessing module processes abnormal values by adopting a clustering algorithm, corrects and normalizes parameters after determining the parameters, meets the requirements of the algorithm and improves the stability and accuracy, and the characteristics act together, so that the system is excellent in performance in a prediction task, and a solid foundation is laid for accurate prediction.

The invention relates to the technical field of structural design, and discloses a topology and size joint optimization design method and system for a hierarchical composite structure, and the method comprises the following steps: dividing a structural domain into a plurality of hierarchical sub-domains, defining the size constraint range of each sub-domain, and distributing a multi-phase material for each sub-domain; grid discretization processing is carried out on the structural domain, a topological design variable is given to each discrete unit, acoustic unit modeling is carried out on the sound field domain, sound pressure boundary conditions are defined, and coupling boundary conditions are applied to the interface of the sound field domain and the structural domain; solving an acoustic-structure coupling equation based on a finite element method, calculating an objective function and constraint conditions, analyzing the sensitivity of the objective function to topological variables by adopting an adjoint method, and iteratively updating the topological design variables through an optimization algorithm until a convergence criterion is met; and on the basis of a topological optimization result, constructing a neural network agent model of an input layer, a hidden layer and an output layer, and carrying out global optimization on sub-domain size parameters by adopting a genetic algorithm so as to minimize a target function.

The present application discloses a method, device and related equipment for interpretability of a large model reasoning process based on Circuit and Embedding. The method of the present application first runs the large model and records the intermediate states of each hidden layer during the reasoning process. An information flow graph is constructed based on the intermediate states, in which nodes represent the Token embedding representations of each hidden layer and edges represent the computational operations between hidden layers. An attribution algorithm is used to calculate the contribution of each edge in the information flow graph to the output result, and key edges are screened according to the contribution, and the Token embedding representations of key nodes are screened in the nodes associated with the key edges. Finally, prompt instructions are used to guide the large model to interpret the Token embedding representations of key nodes to generate interpretation results. By analyzing the intermediate states, constructing the information flow graph, identifying the key paths and generating natural language explanations, the reasoning process of the large model is effectively explained, and the interpretability of the large model is enhanced.

The invention discloses a VSP sound wave full-waveform inversion method of a double-branch physical driving recurrent neural network, which is applied to the field of seismic data processing and aims to solve the problems of low interpretability, low calculation cost and the like due to the adoption of a complete data driving mode in the existing seismic velocity inversion technology. As the wave equation describes the change of the wave field in time and space, a time stepping method is adopted during numerical simulation, namely, the state of the wave field at the next moment is calculated according to the state at the previous moment. Taking a wave field state at a previous moment as a hidden layer of the RNN, performing finite difference solution on an acoustic wave equation, taking a speed parameter in the equation as a trainable parameter in a physically-driven RNN forward modeling network, and taking a shot gather record obtained by outputting through the physically-driven RNN forward modeling network each time as a predicted value; taking shot gather records obtained through physical driving RNN forward modeling output by setting real speed parameters as observation values; the process of solving the loss between the predicted value and the observed value to carry out back propagation to correct the speed parameter is called as a speed inversion process.

The invention discloses a cross-domain privacy protection method and system for advertisement recommendation, equipment and a medium, and the method specifically comprises the steps: carrying out the encryption matching of user behavior data and anonymized equipment data, and generating an initial cross-domain joint feature vector; fusing the initial cross-domain joint feature vector and the disturbance feature vector to form a target cross-domain joint feature vector; splitting a pre-trained advertisement recommendation model into a feature coding sub-module and a reasoning sub-module, deploying the feature coding sub-module to a user adjacent edge node, and retaining the reasoning sub-module in a user local device; and based on the target cross-domain joint feature vector, performing calculation of the feature coding sub-module and calculation of the reasoning sub-module, and uploading the encrypted hidden layer feature vector to a federated learning aggregation server for global model updating. According to the method, cross-domain data utilization and user privacy protection in an advertisement recommendation process are realized, and effective feature vectors are generated for personalized advertisement recommendation while data are guaranteed not to be out of a domain.

The invention discloses an unmanned platform electromechanical system fault early warning method based on a Transform-BiGRU cross attention mechanism, and belongs to the technical field of fault early warning. The method comprises the following steps: acquiring parameters, and performing time sequence alignment and key feature screening; secondly, designing a parallelization spatio-temporal feature extraction framework: on one hand, capturing a long time sequence dependency relationship through a multi-layer Transform encoder, on the other hand, realizing bidirectional context modeling of a sequence by adopting a bidirectional gating circulation unit, and introducing an interpretable global attention mechanism to perform feature weight dynamic allocation on a BiGRU hidden layer state; and a cross attention feature fusion module is constructed for a Transform and BIGRU fusion problem, and cooperative characterization of a time sequence dynamic feature and a space correlation feature is realized through interactive attention calculation of a spatial-temporal feature tensor. By means of the method, advanced early warning from diesel engine lubricating oil pressure abnormity to whole ship electric power system cascade faults is achieved.

The invention discloses an electromechanical composite transmission IGBT junction temperature prediction method based on a NODE reduced-order model. A thermal simulation model is constructed according to a real physical model mechanism, a material and a thermal transfer principle of an IGBT; in the constructed thermal simulation model, performing simulation calculation according to the monitored time sequence data to obtain the junction temperature of the IGBT corresponding to the time change, jointly establishing a data set for training and testing with the monitored data, establishing a NODE neural network order reduction model comprising an input layer, a hidden layer and an output layer, and training the NODE neural network order reduction model by using the data set; and inputting the IGBT related time sequence data into the constructed NODE neural network reduced-order model for prediction to obtain the time sequence data of the junction temperature. According to the method, the IGBT junction temperature can be predicted in real time, the calculation efficiency is high, the interpretability is high, the long-term dependence problem can be better solved, and the method is suitable for complex working conditions.

The invention discloses a roadway anchor net cable support parameter optimization design method based on a neural network, and belongs to the technical field of crossing of mining engineering and artificial intelligence, and the method comprises the following steps: collecting related data of a roadway, and carrying out the preprocessing of the collected data; building a dynamic self-adaptive BP neural network architecture, wherein the dynamic self-adaptive BP neural network architecture comprises an input layer introducing three cross features, a hidden layer adopting a dynamic adjustment strategy and an output layer; an anchor cable model and an anchor rod model are constructed based on a dynamic adaptive BP neural network architecture, and adaptive adjustment is performed by adopting a dynamic learning rate combined with surrounding rock complexity during model training; and inputting the five key influence indexes acquired in real time on site and the three calculated cross characteristics into the trained anchor cable model and anchor rod model to generate predicted anchor cable parameters and anchor rod parameters for on-site support design. The design efficiency of the roadway support parameter scheme is improved, the consumption of resources such as manpower is reduced, and the cost is reduced.

The invention discloses a runoff prediction method for optimizing a neural network based on a hybrid optimization algorithm, and belongs to the technical field of hydrological forecasting. Neural network parameters are optimized through an SAPSO algorithm (simulated annealing particle swarm optimization), and daily runoff prediction is further carried out. The method comprises the following steps: setting parameters of a neural network and an optimization algorithm, including the number of input layer nodes, the number of hidden layer nodes, the number of training times, the population scale and the like; the method comprises the following steps: dividing ten-year daily runoff actual measurement data of a Taojiang station into a training set and a test set, and carrying out normalization processing to prepare for training and testing of a neural network; the weight and the threshold of the neural network are randomly initialized, parameters are transmitted to a simulated annealing particle swarm optimization algorithm for iterative optimization, optimal network parameters are obtained, and a prediction network model is established; and training the network model by using the grouped training set data, and predicting the runoff volume of the Taojiang station by using the training model. And the prediction precision of the runoff volume is obviously improved.

The invention discloses a quantification method and reasoning method of a large language model and electronic equipment, and belongs to the technical field of large language models.The quantification method of the large language model comprises the steps that each linear layer to be quantized in the large language model is quantized; dividing a channel of the linear layer in the hidden layer dimension into a normal channel and an outlier channel; performing INT8 quantization on the first activation matrix corresponding to the normal channel in a word segmentation token dimension to obtain a second activation matrix, and performing INT4 quantization on the first weight matrix corresponding to the normal channel according to an output channel to obtain a second weight matrix; and determining an output result of the linear layer according to the second activation matrix, the second weight matrix, a third activation matrix corresponding to the outlier channel and a third weight matrix corresponding to the outlier channel.

The invention discloses a pulse neural network image classification method and system based on a space-time adaptive threshold, mainly solving the problems of poor nonlinear expression and limited time sequence and space feature processing ability in the prior art, and the scheme comprises the following steps: obtaining an image data set, and dividing the image data set into a training set and a test set; a spiking neural network main body structure comprising an input layer, a hidden layer and an output layer is selected, an existing neuron model is improved by introducing a space-time joint threshold adjustment mechanism, and improved neurons are placed in each neuron layer in the hidden layer to form a spiking neural network based on a space-time adaptive threshold. The training set is used to carry out iterative training; and inputting the test set into the trained pulse neural network to obtain an image classification result. According to the method, a space-time adaptive threshold mechanism is introduced, the threshold can be dynamically adjusted to adapt to time and space features, the processing capacity of the network on time sequence data and complex features and the classification accuracy of images are remarkably improved, and the method can be widely applied to dynamic visual tasks and event-driven scenes.

The invention discloses a large model training method, a question and answer method, related equipment and a computer program product. The method comprises the steps of obtaining question and answer training data; the problem samples are sent to a to-be-trained large model for reasoning, and prediction output of the large model is obtained; calculating an exploration reward based on hidden layer state characteristics generated in a big model reasoning process, wherein the exploration reward is used for encouraging the big model to process the problem sample by adopting an unknown reasoning path; calculating a result reward based on the predicted output and the answer tag; and updating parameters of the large model by adopting a reinforcement learning mode according to the exploration rewards and the result rewards. Exploration rewards are additionally added in the reinforcement learning process, a large model can be encouraged to explore an unknown reasoning path, local optimum is avoided, the probability that the large model discovers a better reasoning path when facing a complex problem is improved, the performance of the large model on a complex reasoning task is improved, and the reasoning efficiency of the large model is improved. And the accuracy of the answering result of the complex reasoning question is improved.