16results about How to "Increase computational overhead" patented technology

The invention discloses a fine tuning method and system for precise parameters of a pre-training language model. The method comprises the following steps: constructing a robust adapter; a plurality of constructed robust adapters are connected in series to form a robust adapter chain, and the maximum chain node is determined; preparing a weight of a pre-training language model and a data set of a corresponding task label; firstly, each robust adapter in the robust adapter chain is subjected to preliminary training, optimization and approximate fine tuning of parameters of the robust adapters; after the parameters of the robust adapter are approximately fine-tuned, the parameters are combined with the weight of the preorder frozen pre-training language model to complete calculation of a current robust adapter chain node, residual information used for next iteration is generated, and the parameters of the current robust adapter are updated at the same time; and carrying out iterative training, updating residual information until the number of iterations is equal to the maximum node of the robust adapter chain, and obtaining the optimal weight of the pre-training language model and the optimal parameters of the robust adapters, so that more stable convergence and simpler hyper-parameter adjustment can be provided.

The invention discloses a large-scale pipeline laying sequence optimization method and system based on machine learning and a storage medium, and belongs to the field of pipeline system design. Comprising the following steps of 1, dividing extracted information data according to a space region, and constructing a quadtree; 2, performing large-scale pipeline grouping according to a Mean Shift clustering algorithm; 3, setting a pipeline laying sequence optimization rule; and 4, introducing reinforcement learning to carry out pipeline sequence optimization, and calculating a Q value. According to the method, the pipeline laying sequence can better meet the actual engineering requirements, the later algorithm complexity is reduced, the algorithm robustness is improved, and the capability of adapting to the dynamic and complex pipeline laying environment is improved.

The application discloses a method for alleviating Non-IID influence based on explainable federated learning. The application mainly introduces a validation set explainable mechanism to depict the explainable results of the validation set samples in the center server based on the influence of local client update on the learning ability of each category of the aggregated model, and evaluates the explainable results of each category by using a structural similarity index (SSIM), so as to deduce the data imbalance clients. Then, the parameters of the data imbalance clients are adjusted, the gradient distance between the model of the data imbalance client and the parameter of the global aggregated model in the last round is minimized, and the parameters of the model of the data imbalance client are corrected through the convergence of the gradient distance, so that the negative influence caused by the data imbalance is weakened.

The present invention provides a grayscale release migration method and system for an information system, which relates to the field of information system migration. The method controls the grayscale ratio of each module by initializing a grayscale ratio vector, dynamically diverts user requests; monitors the abnormality and slow response of the system with migration based on time slices, and generates a system statistical vector; performs weighted statistics and standardization on the statistical vector using a sliding window to form a system state vector; combines health status information to generate a final system state vector reflecting the real-time operating status of the system; calculates the stability index of each module through a stability discrimination model, dynamically adjusts the grayscale ratio, prioritizes the transition of stable modules, and rolls back the grayscale ratio of unstable modules. The system supports automatic adjustment of grayscale ratios, real-time risk response, and multi-architecture compatibility, significantly reduces the risk of the migration process of the information system, and improves the efficiency and stability of grayscale release.

The application provides a laser transmission method and system for atmospheric turbulence adaptive compensation, and relates to the technical field of optical communication. A distorted laser beam is received by a receiving aperture and physically split for detection, and an instantaneous sequence of light intensity, a two-dimensional tilt vector in time domain and a residual high-order phase array are synchronously extracted. The tilt vector is converted into a low-order control instruction in a first control period, the residual high-order phase array is corrected by a physical wind speed feedforward in a second control period and converted into a high-order control instruction, and the stroke saturation and the physical driving voltage saturation are output at the end of the period. The light intensity sequence is collected to calculate a light intensity flicker index, a multi-dimensional channel state matrix is constructed by combining the bit error rate, the stroke and the voltage saturation, a three-state transfer decision is made according to the relationship between the numerical value of the elements in the matrix and a preset threshold, and finally, corresponding communication modulation format switching and receiving aperture physical adjustment are performed according to the decision result, so that time sequence decoupling and cross-layer linkage are realized, and the strong turbulence link breakage rate is greatly reduced.

The invention discloses an infrared and visible light image fusion method based on semantic prior. The method comprises the following five steps: defining a source input image and network structure parameters, constructing a three-level optimization objective function, constructing a grid coding network by adopting a residual representation module with a two-dimensional scanning characteristic, constructing a double-branch multi-level visual priori self-prompting mechanism, and performing staged processing on the characteristics by prompting an interactive fusion network. According to the method, the problem that visual quality and downstream task adaptability are difficult to consider in an existing method is effectively solved, semantic consistency, space structure integrity and visual effect of the fused image are remarkably improved, dependence on manual annotation is reduced, generalization ability and robustness of a model in a complex scene are enhanced, and the method is suitable for popularization and application. The method is suitable for multiple fields of robots, remote sensing, automatic driving and the like.

The application discloses a point cloud classification neural network optimization method and device based on dimension transformation, electronic equipment and medium, comprising: obtaining a point cloud classification neural network to be optimized; embedding a pre-constructed dimension change module between each submodule in the feature extraction module of the point cloud classification neural network; wherein the dimension change module is used for nonlinearly transforming the classified point cloud features into a high-dimensional space through a high-dimensional projection mode to improve the linear separability of the classified point cloud features; and using a data set to optimize and train the network parameters of the point cloud classification neural network embedded with the dimension change module to obtain an optimized point cloud classification neural network. The application fully considers the spatial structure features of the point cloud, improves the linear separability of the point cloud through the high-dimensional projection mode, solves the problem of spatial position information loss in the operation process, and further improves the classification accuracy of the network.

The invention provides an incremental SfM reconstruction method based on relative orientation uncertainty, and relates to the technical field of computer vision, and the method comprises the steps: obtaining a plurality of to-be-reconstructed images, calculating the relative orientation uncertainty index between each pair of images in the plurality of images, and calculating the relative orientation uncertainty index between each pair of images in the plurality of images; the relative orientation uncertainty index is calculated based on the minimum singular value of the relative orientation coefficient matrix; images are used as nodes, connection relations between the images are used as edges, each edge is endowed with a weight to construct an undirected graph model, and the weights are in negative correlation with relative orientation uncertainty indexes; selecting a node with the highest weight and an adjacent node with the lowest relative orientation uncertainty index of the node to form an initial image pair, and resolving an initial relative pose; and selecting a node with the lowest relative orientation uncertainty index from the alternative nodes connected with the current model, adding the selected node into the model, resolving the camera pose of a newly added image, gradually adding the image until all the images are added, obtaining a reconstruction path with the lowest uncertainty, and further completing model reconstruction.

The invention relates to a probability uncertainty closed-loop robust trajectory planning method considering execution mechanism saturation, belongs to the technical field of aircraft control, and aims to reduce the sensitivity degree of a guided ammunition scheme trajectory to uncertain factors and give full play to the limited control ability while avoiding execution mechanism saturation as much as possible. Firstly, a control quantity of a deterministic energy optimal trajectory is acquired as an initial value of a subsequent algorithm; secondly, modeling is carried out on main uncertain factors, and an uncertainty quantification module is built based on a PCE method; then, designing a corresponding constraint condition and a target function, establishing an open-loop robust programming model, weighing and considering algorithm precision and calculation time consumption, designing a closed-loop tracking PID controller, and coupling with the open-loop programming model; and finally, carrying out closed-loop robust planning by adopting a direct targeting method. The method fully considers the matching of the trajectory and the guide control system, and overcomes the problems of insufficient description precision of a nonlinear model and sensitivity to an initial value in a traditional method.

This application discloses a photovoltaic panel vibration monitoring method based on heterogeneous fusion network communication. The method includes: after each sensor terminal completes vibration data acquisition and processing within the same acquisition cycle, it sends the local relative timestamp generated by the free-running timer to the gateway base station along with the uplink data packet; the gateway base station calculates the relative transmission delay difference between any pair of nodes based on the packet reception time and the local relative timestamp, and constructs a frequency domain phase correction factor at each discrete frequency point accordingly to correct the uncorrected cross-spectral density, obtaining the corrected cross-spectral density; further, it calculates the coherence function value by combining the self-spectral density of each node, and locates the vibration anomaly area based on the spatial distribution of the coherence function value on the photovoltaic array topology. This invention can improve the accuracy of multi-node vibration joint analysis and the reliability of anomaly location, and is suitable for online monitoring of photovoltaic power plants.

The invention relates to an image feature extraction device based on a multi-scale convolutional network model and application thereof, and belongs to the technical field of artificial intelligence and computer vision, the device comprises the multi-scale convolutional network model based on dynamic weight self-adaption, the model comprises a self-adaption heterogeneous convolution kernel dynamic weighting deep convolution module, an image feature extraction module and a feature extraction module, the multi-scale convolution kernel fusion module is used for adaptively generating and fusing weights of multi-scale convolution kernels according to input image features; the dynamic Inception mixer is used for processing multiple groups of features after channel segmentation in parallel; the dynamic mixing block is used for extracting and fusing multi-scale features and channel interaction features through a dual-path residual structure; and the integrated network module based on the C2f architecture comprises a plurality of dynamic mixing blocks which are connected in series and is used for aggregating the multi-scale features and outputting final image feature representation. According to the method, multi-scale features and long-distance dependence in the image can be efficiently captured, the model performance is improved while the parameter quantity is reduced, and the method is suitable for various computer vision tasks such as target detection and image segmentation.

The present application provides a model training method and device, equipment and medium. The model training method comprises: training a teacher model and a student model based on a training sample set, obtaining a first sequence output by the teacher model and a second sequence output by the student model; determining that a vocabulary of the teacher model and a vocabulary of the student model are heterogeneous vocabularies, determining a distribution difference loss based on a probability ranking result of the first sequence and a probability ranking result of the second sequence; updating parameters of the student model based on the distribution difference loss to achieve effective knowledge distillation in a heterogeneous vocabulary scenario, alleviate the problem of mismatch between training and inference distribution, and improve the flexibility and applicability of knowledge distillation.

The invention relates to a data verification method and system suitable for an embedded software platform, and belongs to the field of relay protection operation and maintenance intellectualization. The method comprises the following steps that data blocks needing to be protected are uniformly protected through a dynamic data linked list, and the data block protection efficiency is improved. And inserting data blocks needing to be verified into the dynamic data linked list, generating a data verification code corresponding to the data block for each data block inserted into the dynamic data linked list, and storing the generated data verification codes in the verification data block. When data verification is started for a data block, a data verification code is generated, the data verification code is compared with a data verification code stored in a verification data block, when the data verification codes are the same, the data block is not tampered, and when the data verification codes are different, the data block is tampered, and the reliability of tampering verification is improved through the data verification codes. And the security of the data of the data block is ensured, so that the security of a corresponding relay protection device system is improved.

The present disclosure provides a simulation test method and device of a large model, electronic equipment and storage medium, relates to the technical field of computers, and particularly relates to the technical field of artificial intelligence such as deep learning and large models. The specific implementation scheme is as follows: obtaining inference performance data of a large model to be tested; determining a first mapping relationship of a pre-filling stage and a second mapping relationship of a decoding stage based on the inference performance data, the first mapping relationship at least including a corresponding relationship between pre-filling time consumption and input token quantity, and the second mapping relationship at least including a corresponding relationship between decoding time consumption and concurrency quantity; and performing simulated inference on the large model to be tested by using the first mapping relationship and the second mapping relationship to obtain a simulation test result. The present disclosure can improve the accuracy of the simulation test result.

The application discloses a phase diversity aberration estimation and image reconstruction method based on implicit neural representation, adopts an implicit neural representation network to model a continuous function of a target object, adopts a phase diversity observation image to provide multi-observation constraints, adopts a Zernike polynomial and a microscopic imaging physical model to represent system aberration, and jointly solves the target object and the wavefront aberration under a unified unsupervised differentiable optimization framework, so that the target object recovery and the aberration estimation are realized without external labeled training data and additional wavefront measurement hardware. The application reduces the redundancy of the target object representation, improves the reconstruction stability, does not need external labeled training data, enhances the method adaptability, improves the aberration estimation accuracy, alleviates the multi-solution problem, enhances the continuous structure and high-frequency detail expression capability, enhances the physical consistency and result reliability, and improves the stability of the joint optimization process.

The application belongs to the field of brain-computer interface, and particularly relates to a convolutional neural network construction method for enhancing time domain information in convolutional output, which is used for classification tasks or regression tasks in brain-computer interface, and comprises the following steps: improving and training the structure of a convolutional neural network to be enhanced, wherein the improvement mode is as follows: according to the structure of a convolutional layer to be improved, the size of a feature representation output by the convolutional layer is calculated, an average pooling layer with the same output size and feature representation size is constructed, the average pooling layer is used for accompanying convolution kernel sliding point multiplication process of the convolutional layer to be improved, input data is average-pooled to generate an average pooling representation; a trigonometric function encoder is constructed to perform time sequence coding on each element in the average pooling representation to obtain a new average pooling representation; a summation unit is constructed to sum corresponding elements of the new average pooling representation and the feature representation to obtain a feature representation with enhanced time domain information, and the feature representation is input into a next layer structure. The application can enhance time domain information without expanding the size of model parameters.