46results about How to "Enhance reasoning ability" patented technology

The present disclosure relates to a model deployment method and device, electronic equipment, storage medium and product. The method comprises: obtaining an image saliency segmentation model; the image saliency segmentation model is used for saliency segmentation of an input image, the image saliency segmentation model comprises a preset activation function, the output value of the preset activation function is in a preset range, the maximum value of the preset range is a first preset value; converting the model format of the image saliency segmentation model into a target model format supported by a target chip; deploying the image saliency segmentation model in the target model format on the target chip. The present disclosure can limit the size of the output value of the preset activation function, thereby improving the accuracy of the image saliency segmentation model. In addition, by converting the model format of the image saliency segmentation model into the target model format, the image saliency segmentation model can be efficiently run on the target chip, thereby improving the inference efficiency and inference performance of the image saliency segmentation model.

This application discloses a data-aware intelligent analysis method, system, storage medium, and computer device. The method includes: receiving a target problem and initializing state information, wherein the state information includes key sub-conclusions, a reasoning state logic diagram, and key data descriptions; a language model-based agent iteratively performs the following operations: obtaining input information for the current step; updating the state information of the current step based on the action and environmental feedback information of the previous step, and making decisions based on the state information of the current step to obtain the action and tools to be invoked for the current step; invoking the tools to execute the action of the current step, and obtaining environmental feedback information for the current step after the action is completed; checking whether the target problem has been analyzed; if not, proceeding to the next loop; if completed, generating analysis results. This method can improve the reasoning ability, response consistency, and interpretability of deep reasoning models.

The invention relates to a knowledge-enhanced big language model reasoning method based on experience knowledge. Comprising the following steps: constructing a knowledge base; obtaining reasoning problems and corresponding problem solving experience; extracting the problem solving experience into structured knowledge entries; a reasoning structure of the problem is extracted in the structure abstraction process, and experience is refined into a reusable solution, judgment and correction mode in the mode generalization process; retrieval of knowledge: for a new problem, related items are retrieved by using a reasoning structure of the new problem; the diversity reordering step then selects a final set of entries, providing a strategically diverse context for reasoning. The problem-based core reasoning structure adopts a diversity perception retrieval mechanism, so that the LLM reasoning accuracy is improved, and the LLM can learn and evolve from practice under the condition that re-training is not needed.

The invention relates to a test case generation model training method and a test case generation method, and the method comprises the steps: carrying out the first-stage reinforcement learning training of a preset model through employing a first training data set, and obtaining a first target model; wherein the first training data set comprises a first function specification sample of a plurality of function points and a first standard test case, and the preset model is used for generating a first prediction test case according to the first function specification sample; in the first-stage reinforcement learning training process, a target reward value is adopted to guide a prediction model to optimize a generation strategy, and the target reward value is determined according to a correctness reward rule, a format reward rule and a coverage rate reward rule according to a first function specification sample, a first standard test case and a first prediction test case. On the basis, the test case generation model obtained through training can reverse multiple test scenes such as a test scene, a test scene corresponding to a market problem and a test scene corresponding to a boundary condition.

This application proposes a data merging method, apparatus, device, and computing medium, comprising: obtaining the micro-rank matrix corresponding to the merging matrix of each model to be merged, and a first matrix and a second matrix corresponding to each micro-rank matrix, wherein the matrix parameters of each first matrix conform to a uniform distribution, and the matrix parameters of each second matrix conform to a Gaussian distribution; performing pruning processing on each first matrix and each second matrix to obtain corresponding first pruned matrices and second pruned matrices; obtaining a singular value scaling matrix for each first matrix based on each first matrix and the corresponding first pruned matrix; and obtaining the merging matrix of the merged model based on each singular value scaling function, each first pruned matrix, and each second pruned matrix. The embodiments of this application process the pruned micro-rank matrix using the singular value scaling matrix to adjust the singular values ​​of the merging matrix, thereby avoiding distribution shift during inference of the merging model and improving the inference performance of the merging model.

A method for generating a training sample for an image detection model comprises the steps that tampering detection is carried out on a target image by detecting a large model, and an initial detection report is obtained and comprises abnormal visual information and abnormal logic information; the abnormal visual information is determined based on pixel analysis on a target image, and the abnormal logic information is determined based on semantic analysis on the target image; inputting the initial detection report and the first cue word into a large reasoning model to generate a thinking chain text; the thinking chain text is used for generating a training sample corresponding to the target image.

This invention discloses a portable edge-side large-scale computing device based on a solid-state battery, comprising a casing, a solid-state battery core power supply module, a heterogeneous computing and storage motherboard, an intelligent composite three-dimensional heat dissipation system, and a temperature management model. Through the collaborative optimization design of the solid-state battery core power supply module, the heterogeneous computing and storage motherboard, the intelligent composite three-dimensional heat dissipation system, and the temperature management model, this invention solves the technical challenges of insufficient battery life, heat dissipation bottlenecks, poor power management adaptability, and low system integration in edge-side large-scale model deployment, achieving efficient and stable operation of large-scale mobile models.

This application provides an inference optimization method, apparatus, and storage medium for a generative diffusion model. The optimization method includes one or more of the following: first optimization of the inference parts of the U-Net and VAE in the generative diffusion model through code recompilation; second optimization of the performance bottleneck region in the VAE through data layout transformation; and multi-level operator optimization of key computational operations in the generative diffusion model based on performance bottleneck analysis. Through the above technical solutions, this invention can improve the generative diffusion model on a heterogeneous acceleration platform, aiming to locate the performance bottleneck part in the model through performance profiling, improve the inference efficiency on the heterogeneous acceleration platform through data layout transformation, and systematically improve the model's inference performance by utilizing multi-level optimization schemes from the graph level, data level, to the operator level.

The invention discloses a monitoring task processing method based on a natural language, and the method comprises the steps: carrying out the semantic analysis of a to-be-processed target monitoring task, and obtaining a target entity and a core execution action corresponding to the target monitoring task; the target monitoring task is described in a natural language; determining a target tool from the candidate tools based on the core execution action and the function types of the candidate tools, and determining candidate entities which can be identified by the target tool and the total number of the entities; respectively determining a target semantic pole corresponding to the target entity and a candidate semantic pole corresponding to the candidate entity on the basis of a mutual exclusion attribute model, the inference computing power of the edge computing equipment and the total number of the entities; based on the target semantic pole and the candidate semantic pole, selecting an associated entity corresponding to the target entity from the candidate entities; generating a task processing rule based on the target monitoring task, the target entity and the associated entity which can be identified by the target tool; according to the method and the device, the obvious reduction of task processing resource consumption and the synchronous guarantee of execution accuracy are realized on the edge computing equipment.

This invention provides a method for constructing and reasoning a ship navigation knowledge graph based on multi-source heterogeneous data, involving the intersection of artificial intelligence and maritime technology. The method includes: Step S1, establishing a ship navigation knowledge model; Step S2, acquiring and preprocessing multi-source heterogeneous data; Step S3, performing named entity recognition on the preprocessed data based on a BiLSTM-CRF hybrid model incorporating domain dictionary features; Step S4, extracting entity relationships from the entities identified in Step S3 using a pre-trained BiLSTM hybrid model incorporating interactive attention mechanisms; Step S5, fusing knowledge from the entities and entity relationships extracted in Steps S3 and S4 to construct a preliminary knowledge graph; Step S6, performing link prediction on the preliminary knowledge graph based on an RGCN model incorporating temporal constraints and rule logic to achieve knowledge completion and reasoning. This invention improves the ability to respond to risks in complex navigation scenarios.

The application discloses a dynamic route hybrid expert model reasoning method, system, device and medium, which are corresponding solutions, and in the solutions: through automatic parallel strategy search, the application can divide the model into pipeline stages with balanced running time, reduce the idle rate of the calculation unit, and effectively improve the execution efficiency of the calculation unit; and the automatic search process of the application can reduce the maximum pipeline stage running time as much as possible under the condition of meeting the memory limit, so that the reasoning performance of the model is improved.

The application belongs to the technical field of artificial intelligence, and provides a reasoning method based on a multi-modal diffusion transformer storage image memory drive, comprising the following steps: an image memory reasoning model is constructed, a visual language model is taken as a core reasoning controller, and a multi-modal diffusion transformer is taken as a peripheral image memory storage of the reasoning controller; a model training method is constructed; a model reasoning method is constructed, the visual language model firstly outputs a thinking text and a specific recall instruction based on text and image input, then maps the image input and the recall instruction into a visual and text feature fusion tensor, takes the visual and text feature fusion tensor as a guide condition to make the multi-modal diffusion transformer output a corresponding recall image, and finally inputs the recall image into the visual language model for further reasoning. After the above model function architecture design and training steps, the visual language model has an image memory function, and uses the recall image for multi-round thinking, thereby improving the accuracy and reasoning ability of the model question and answer.

This specification provides a data processing model training method and system, and a question-answering method. The data processing model training method includes: inputting a sample question into a pre-trained data processing model to obtain answer reasoning steps; selecting a target answer reasoning step and its preceding answer reasoning steps from the answer reasoning steps, and expanding the target answer reasoning step based on the preceding answer reasoning steps to obtain an expanded answer reasoning step; and fine-tuning the pre-trained data processing model based on the preceding answer reasoning steps, the expanded answer reasoning steps, the sample question, and the sample answer to the sample question to obtain the target data processing model. By expanding the target answer reasoning step and then performing targeted fine-tuning training on the pre-trained large language model, the target data processing model's ability to understand and answer questions is improved.

The application provides a large model-based vertical field data construction method, and belongs to the technical fields of data processing and artificial intelligence, and comprises the following steps: converting a vertical field source document into an intermediate format text and dividing the intermediate format text into multiple text blocks; inputting the text blocks into a pre-trained generative language model, guiding the pre-trained generative language model to generate a plurality of candidate questions according to the contents of the text blocks according to a pre-designed prompt word, and performing preliminary screening and fine screening on each candidate question to obtain a question set; predefining a mode of a knowledge graph according to the field characteristics of the vertical field, processing all the text blocks based on an information extraction model, and constructing a field knowledge graph; and performing local context retrieval based on the text blocks of the source of each final question in the question set, performing global knowledge retrieval based on the field knowledge graph, and generating a final answer and a final thinking chain. The application is suitable for different vertical fields, and can effectively improve the quality of data and ensure the accuracy of question generation.

The application provides a data processing method and system based on network security service, relates to the technical field of network data security processing, realizes the structured expression and behavior extraction of potential threats in encrypted communication by establishing a new data structure and behavior association model, associates the originally isolated security events into a complete attack path by introducing a context causal chain modeling mechanism, establishes an event grading mechanism based on the comprehensive scoring of business criticality, response cost and attack propagation path, so that the threat processing is no longer blindly and equally allocated resources, but intelligently scheduled according to the priority, and the resource utilization efficiency is improved, can generate structured and executable defense instructions or blocking strategies according to the analysis results, supports human-computer cooperation or automatic response, and significantly improves the response efficiency and strategy adaptability. An intelligent data processing solution with high availability and high scalability is constructed for the actual application scene of network security service.

Embodiments of the present application disclose a training method of a retrieval model, a retrieval method and an electronic device. The training method comprises: fusing prior knowledge of historical wafer defect cases in a database into a multi-modal feature vector of a current training sample through a cross-attention module to generate a multi-modal enhanced query vector of the current training sample; splicing enhanced query vectors of different modes of the current training sample into a multi-modal sequence of the current training sample according to a time sequence of a process flow through a vector splicing module; inputting the multi-modal sequence of the current training sample into a sequence processing module to extract fusion features of different modes of the current training sample; calculating a distribution distance of the fusion features of different modes of the current training sample in a same feature space to construct a global alignment loss function, and optimizing parameters of the retrieval model through a back propagation algorithm. Embodiments of the present application improve the retrieval accuracy of historical wafer defect cases.

The invention discloses a positioning data transmission method and device, a terminal, network side equipment and a storage medium, and belongs to the technical field of communication, and the positioning data transmission method comprises the steps that the terminal sends a first message to first network side equipment, and the first message is used for requesting the first network side equipment to provide first auxiliary data; the first message carries first information, and the first information is used for indicating the tendency of first auxiliary data requested by the terminal; the terminal receives a second message from the first network side equipment, and the second message is used for providing second auxiliary data for the terminal; and the terminal measures the downlink positioning reference signal according to the second message and acquires a data set.

The application discloses a material attribute intelligent dialogue method and system based on a knowledge graph embedding and diffusion generation mechanism, and belongs to the technical field of the cross of artificial intelligence and material science. The application solves the problems that existing methods are faced with data scarcity, insufficient prediction ability for complex crystal structures and inability to perform data privacy protection. The application firstly constructs material knowledge graph data through entity extraction; secondly, a graph neural network conditional diffusion model is combined to construct a crystal structure containing alloy components; then, a self-supervised independent pre-training model is used for general feature extraction; finally, semantic features, crystal structure features, entity features and knowledge graph features are fused, and the fused features are processed to serve as input of a generative language model to generate professional dialogue responses. The method can be applied to material attribute intelligent dialogue.

The embodiment of the invention provides a training method and platform for a diffusion language model, equipment, a medium and a product. According to the scheme, the method comprises the steps of obtaining a plurality of intermediate prediction results generated by an initial diffusion language model for sample cue words; for each of the plurality of intermediate prediction results, determining a reward value of the prediction result; and adjusting parameters of the initial diffusion language model based on the reward value to obtain a trained diffusion language model. According to the scheme, the disadvantage that the contribution degrees corresponding to the rewards and the intermediate prediction results are not matched can be effectively avoided, refined hierarchical regulation and control of the model optimization process can be realized, the overall training effect of the model can be improved, and the reasoning performance of the diffusion language model after training can be improved.

The invention discloses a medical image tool enhancement processing method based on auto-reflection reinforcement learning. The method comprises the following steps: constructing a multi-modal reasoning model for vision-language joint reasoning; and performing supervision and fine tuning on the model by utilizing a cold start data set with tool calling track marks, so that the model has basic structured reasoning capability and normative tool use behaviors. An early-stage error reasoning track and a later-stage correct reasoning track which are generated by the same input at different training check points are collected, an auto-reflection training sample is constructed, and auto-reflection fine tuning is performed on the model. And based on the reinforcement learning data set containing the real tool calling scene, constructing a reward function and executing tool enhanced reinforcement learning training. After a medical image to be processed and a corresponding natural language problem are received, the trained model can generate gradual natural language thinking, a tool calling instruction and an external tool observation result, and outputs a medical image analysis conclusion, so that refined and explainable medical visual reasoning is realized.

The application discloses a low-resource multilingual supervised data set quality improvement and construction method, belonging to the technical field of natural language processing and data construction, comprising the following steps: defining a task system and collecting original data; automatically obtaining the original data and parsing them into standardized texts; a semantic cleaning and privacy protection preprocessing stage for deeply cleaning the texts, segmenting semantic blocks and desensitizing private information; a structured encapsulation and man-machine collaborative labeling stage for converting the cleaned texts into high-quality instruction fine-tuning data; a quality verification and data set solidification stage for quality verification and solidification of the instruction fine-tuning data set; and a data asset management and continuous evolution stage for safe management, version control and continuous iteration of the data assets. The method significantly improves the alignment quality and robustness of the model in a low-resource multilingual environment, and provides a solid technical foundation for constructing high-performance multilingual LLMs.

The invention discloses an electric power file archiving method, device and equipment, a storage medium and a product thereof, and relates to the technical field of data processing, the method comprises the following steps: responding to a received to-be-archived electric power file, retrieving a target knowledge sub-graph related to semantics of the electric power file based on a pre-constructed electric power archiving knowledge graph, generating a corresponding semantic prefix; querying whether a cache item of a semantic prefix exists in a prefix cache region, wherein the prefix cache region comprises a historical mapping relationship between the semantic prefix of the archived power file and a corresponding compliant archiving path; if the cache item exists, taking a compliance archiving path corresponding to the semantic prefix in the prefix cache region as a target archiving path; if the cache item does not exist, based on the target knowledge sub-graph, generating a target archiving path through a large language model and graph constraint decoding; and on the basis of the target filing path, the power file is filed, so that the overall reasoning performance of the power file filing robot is improved.

The application discloses a natural language and machine vision alignment method, which proposes a three-level alignment architecture, adopts a double-flow feature extraction network to extract local-global features of a visual scene and syntax-semantic features of natural language respectively, adopts a cross-modal alignment module with time-space attention enhancement to complete adaptive projection in a feature space by using a dynamic gating mechanism, constructs a joint optimization strategy based on contrast learning, optimizes the consistency of visual and language embedding spaces by using a multi-granularity contrast loss function, simultaneously introduces semantic topology constraints and visual causal reasoning to reduce the computational complexity and improve the task robustness, and achieves a Top-1 accuracy of 92.3% in image-text retrieval and a visual question answering F1 value of 83.4% through experimental test data, and the model parameter quantity is reduced by 40%, so that the method can be widely applied to intelligent interaction systems, automatic driving scene understanding, industrial quality inspection knowledge base construction and other fields, and significantly improves the semantic perception and reasoning capability of a multi-modal system.

The application provides an automatic data governance method and system based on a multi-modal large model, comprising: collecting multi-source heterogeneous industrial data and performing standardization processing to form standardized multivariate time series data; constructing a process knowledge base, performing semantic embedding coding on process knowledge text, and storing; constructing and fine-tuning a KTSF multi-modal large model, fusing process knowledge semantics and multivariate time series data through a cross-modal attention mechanism to generate joint semantic representation; based on the prediction of the KTSF multi-modal large model, outputting the residual error between the actual data, dynamically identifying abnormal data; and performing attribution analysis; based on the attribution result, calling the KTSF multi-modal large model to generate a repair value, and intelligently correcting the abnormal data; designing a quality evaluation and feedback learning module for calculating data quality scores and driving model incremental updating; designing a rule self-learning module for automatically refining governance rules through cluster analysis and updating the knowledge base.

The invention discloses a power grid wiring diagram verification method based on a multi-modal large model, and the method comprises the steps: recognizing a to-be-verified power grid wiring diagram image through a computer vision model, and obtaining initial primitive information which comprises a primitive type and a primitive position; based on the initial primitive information, adding a non-occlusion visual mark to a corresponding primitive in the power grid wiring diagram image, and generating a visual enhancement image; based on the initial primitive information, retrieving a related target naming rule from a power grid naming specification knowledge base through an improved Monte Carlo tree search algorithm; and on the basis of the vision enhancement image and a target naming rule, identifying error information in the initial primitive information through a multi-modal large model, and completing the verification of the power grid wiring diagram. According to the invention, high-precision graph-to-model conversion and logic verification of the wiring diagram are realized.

A method and system for enhancing large language model auxiliary network operation and maintenance capability, the method comprising: collecting a training sample, the sample collection process comprising: inputting analog network data into a network simulator, generating a text attribute graph of an analog network and a corresponding natural language query, and collecting network behavior data according to the natural language query; a large language model is constructed and trained, and the working process is as follows: natural language query and equipment configuration texts are generated through a configuration encoder, text attribute graphs are obtained through a topology encoder, then the text attribute graphs are spliced into network representation, and then the network representation is mapped into a Q layer and a V layer of an attention mechanism of the large language model in a low-rank fine tuning mode; and the large language model receives natural language query input by the user and the text attribute graph of the current network for reasoning output. The invention relates to the field of information communication networks, can efficiently model multi-modal network data, enhances the understanding ability of a large model for the current network state, and effectively assists network operation and maintenance.

The invention belongs to the technical field of typed reasoning, provides a large model typed reasoning method based on hierarchical distillation and dynamic weight merging, and realizes structured migration and cross-task fusion of SLMs typed reasoning knowledge through hierarchical thinking chain distillation and dynamic weight model fusion. The hierarchical distillation is used for extracting three types of key information including semantic representation, inference intermediate state and final thinking chain expression from the LLMs, and establishing semantic basis, logic arrangement and step generation capability required by typed inference in the SLMs in a structured mode. Meanwhile, a dynamic weight merging strategy with task types as guidance is adopted, adaptive fusion is carried out on parameter subspaces of the trained SLMs, the complementary advantages of different reasoning types are effectively reserved, and the cross-type reasoning capability is enhanced.

The invention discloses a self-adaptive computing power distribution control method and device and a storage medium, and relates to the technical field of data processing, and the method comprises the steps: extracting the computing complexity, the video memory demand and the communication overhead corresponding to a computing unit; setting a threshold parameter according to the calculation complexity, the video memory demand and the communication overhead, and clustering and integrating the calculation units to generate a plurality of sub-model modules; performing matching deployment on the sub-model modules and heterogeneous hardware nodes based on a deployment strategy, and recording state information of each node after deployment is completed; state indexes in the state information and the historical performance data are predicted and processed through exponential smoothing, and the dynamic weight value of each node is calculated in combination with the bottleneck sensitivity coefficient; and generating a node priority list according to the dynamic weight value, and allocating the hardware type and the task proportion of the sub-model module according to the node priority list and a task allocation optimization mechanism so as to complete the self-adaptive scheduling of the sub-model module. The problem of low efficiency is solved, and the resource utilization rate is improved.

This application discloses a reasoning method, system, device, and medium based on heterogeneous model reinforcement learning, belonging to the field of artificial intelligence technology. This method generates structured tree-like reasoning knowledge at low cost using a Monte Carlo tree search mechanism, introduces computational value scoring to balance accuracy and computational overhead, and applies the high-order reasoning patterns obtained from the tree-like reasoning knowledge to the reinforcement learning training process. This effectively internalizes the high-order knowledge of the heterogeneous model group into the first policy model, thereby solving the dual technical challenges of high cost in constructing high-quality distilled data and the lack of multi-source external knowledge guidance in traditional reinforcement learning algorithms. Furthermore, it enables the first policy model to reason about answers to natural language tasks through a tree-like exploratory reasoning process during the reasoning phase, significantly improving the generalization ability, exploration ability, and decision confidence of the first policy model, thus effectively enhancing its reasoning ability.

The application relates to a progressive fine-tuning method and device for power scene model migration. The method comprises the following steps: obtaining a plurality of model parameters of a target model to be trained, determining the attention weights of each model parameter to each power knowledge entity in a pre-constructed power knowledge graph; mapping the attention weights corresponding to each model parameter into the activation scores of each model parameter; classifying each model parameter into a corresponding parameter layer according to the activation scores; obtaining a plurality of power corpora, and determining the first information entropy of each power corpus; classifying each power corpus into a corresponding corpus gradient according to the first information entropy; activating the parameter layer matched with each corpus gradient in the ascending order of the corpus gradient, training the activated parameter layer by using the power corpus of the corresponding corpus gradient, and obtaining the trained target model until the training stopping condition is reached. The method can improve the training efficiency and performance of model migration learning in the power scene.