68results about How to "Improve reasoning efficiency" patented technology

A multi-hop reasoning method based on dynamic reasoning guidance and multistage self-feedback retrieval belongs to the field of natural language processing, and comprises the following steps: deconstructing a multi-hop reasoning process into a target-oriented sequence decision problem, carrying out dynamic reasoning guidance by using a large language model, generating a sub-problem sequence matched with a reasoning progress in real time, and carrying out multi-level self-feedback retrieval on the sub-problem sequence; target document retrieval is guided, and sub-questions are dynamically generated; according to the generated sub-questions, obtaining associated documents by adopting a three-level collaborative retrieval mechanism; and performing information refining on the associated document through a large language model, fusing the refined information into an inference chain, and performing inference to generate an answer. The invention further discloses a multi-hop reasoning system, a storage medium and a computer program product. The method aims at solving the complex multi-hop problem that multiple dispersed knowledge fragments need to be integrated, high-accuracy and high-efficiency reasoning is achieved, the retrieval requirement is dynamically generated through an explicit thinking chain guiding mechanism, and evidence obtaining is optimized and redundant information is filtered in combination with a three-level self-feedback retrieval mechanism.

The invention discloses a self-adaptive multi-level forged voice detection method based on a belief propagation mechanism. The method comprises the following steps: performing feature extraction on voice to be detected to obtain an initial feature vector; reasoning the initial feature vector by using a lightweight model to obtain an initial confidence coefficient representing the voice forgery possibility; based on a low-confidence threshold value and a high-confidence threshold value which are determined through performance optimization, performing multi-level discrimination on the input voice and screening difficult samples; generating expert feature vectors for the difficult samples by adopting a depth feature extraction network regulated and controlled by the initial confidence coefficient; constructing a query vector based on the initial confidence, and fusing expert features through an attention mechanism; and voice authenticity determination is completed according to the fusion features. The initial confidence is used as the cross-stage control quantity, so that the self-adaptive adjustment of the detection process is realized, and the average reasoning delay is reduced while the detection accuracy is ensured.

This invention discloses a request processing method, apparatus, device, and storage medium, comprising: when the matching result is determined to be a matching failure, inferring the user request using a synchronous direct transmission mode through the target instance to obtain a first key-value cache, and determining the storage method of the first key-value cache based on its cost; when the matching result is determined to be a match with the target key-value cache, inferring the user request using a prefetch incremental mode through the target instance to obtain a second key-value cache, and directly saving the second key-value cache to a storage center. In the synchronous direct transmission mode, the cost of the key-value cache obtained through user request inference is calculated, and key-value caches that meet the cost requirements are saved to the storage center, avoiding invalid space occupation in the storage center and improving the inference efficiency of subsequent user requests. In the prefetch incremental mode, when a key-value cache hit meets the cost requirements, it is directly used to improve inference efficiency and avoid redundant calculations and waste of computing resources.

The embodiment of the specification provides a model training method and platform, and a text reasoning method and platform, wherein the model training method comprises the following steps: obtaining an initial sample set, a student reasoning model and a reasoning ability label of the student reasoning model, wherein the initial sample set comprises a plurality of sample thinking chains, the plurality of sample thinking chains are generated by using different reasoning methods by a teacher reasoning model, each sample thinking chain is labeled with a reasoning method label of a corresponding reasoning method, a sampling weight of the plurality of sample thinking chains is determined based on a deviation between the reasoning ability label of the student reasoning model and the reasoning method labels labeled by the plurality of sample thinking chains; a target sample set is obtained by sampling from the initial sample set based on the sampling weight; and the student reasoning model is trained based on the target sample set to obtain a target reasoning model. The target reasoning model obtained by the training method can dynamically adjust the reasoning process according to the complexity of the problem, and can ensure the reasoning accuracy while improving the reasoning efficiency.

The application discloses a material mechanical property inference method based on multi-modal pre-training representation. In view of the problems that the existing method needs to rely on expensive scanning electron microscope images in the inference stage, and the prediction accuracy is low in the small sample scene, the application constructs a multi-modal network containing an image encoder, a table encoder, a shared projection branch and a modal specific projection branch, and carries out joint pre-training through cross-modal alignment loss, intra-modal decoupling loss based on contrast logarithmic ratio upper bound mutual information estimation and table mask recovery loss. Then, the pre-trained table encoder and its auxiliary projection head are migrated to the downstream multi-task regression network, and the mechanical properties can be predicted only by inputting process parameters. The method can also maintain stable performance in the small sample scene of iterative sample collection, significantly reduces the material representation cost, and improves the prediction accuracy and generalization ability.

This application discloses a method and system for identifying 2D engineering drawing elements based on dynamic query-guided consistent projection, proposing a dynamic query-guided consistent projection decoder. This decoder achieves a single forward propagation projection from a random noise vector to an accurate element state vector. The method first initializes a set of vectors following a predetermined noise distribution as initial dynamic queries. These queries are fed into the dynamic query-guided consistent projection decoder (DQ-CPD). At each layer of the decoder, the current dynamic query vector not only interacts with other queries through a self-attention mechanism, but it is also input as a controller into a "contextualized hypernetwork." This hypernetwork generates a set of dedicated weight parameters in real time based on the current inference information carried by the query vector. This method organically combines dynamic queries, dynamic context, and one-step generation theory to construct a novel, efficient, accurate, and adaptive element identification method.

This invention discloses a short-window gamma-gamma turbulence parameter prediction method for space-to-ground laser communication, belonging to the technical field of space-to-ground laser communication and atmospheric turbulence channel parameter prediction. This method addresses the problems of traditional methods, such as strong dependence on long observation windows, significant degradation in prediction accuracy under short-window scenarios, and insufficient robustness under low signal-to-noise ratio conditions. It establishes a short-window observation model for the space-to-ground optical link and a gamma-gamma channel statistical model, constructs time-dependent short-window training data, and designs a short-window gamma-gamma network. Temporal features are extracted through a convolutional backbone, and a scintillation exponential physical regularization auxiliary head is used to achieve joint parameter prediction under physical constraints, outputting predicted gamma-gamma distributed parameters. This method can achieve high accuracy and good stability in parameter prediction under short observation windows and low signal-to-noise ratio conditions, and can be used for turbulence channel state characterization at the space-to-ground laser communication receiver.

The application discloses a large model inference scheduling method and system, a storage medium and a computer program product, and relates to the technical field of large model inference scheduling. The application changes the "blind pushing" mode of the gateway in the traditional scheme, and adjusts to be dominated by the computing node. The computing node can dynamically and comprehensively perceive state data of the node, such as processor state and inference task state borne by the node. Based on the state data, the performance index of the computing node can be determined. The computing node actively reports the performance index to the gateway. The gateway can allocate target computing nodes for target inference tasks to be allocated based on the latest performance indexes of the computing nodes, better realize load balancing, and then improve the overall throughput of the system, improve inference efficiency, and improve the stability of inference service.

The invention relates to the field of intelligent medical treatment, in particular to a children disease cause analysis method and device based on interlayer information reuse and a program product. The method includes: acquiring a child case data set and a disease tag; inputting the child case data set and the disease label to a to-be-trained neural network model for training to obtain a child pathogenesis analysis model; the neural network model carries out parameter sharing through interlayer similarity calculation to reduce the total parameter quantity of the model, the model comprises a first attention layer and a second attention layer, the similarity of the first attention layer and the second attention layer is calculated, and when the similarity of the first attention layer and the second attention layer is larger than a preset threshold value, the first attention layer and the second attention layer are selected. If the layer is judged to be a functional similar layer, endowing the parameters of the first attention layer to a second attention layer; otherwise, the second attention layer obtains parameters through vector feature learning. According to the model, the understanding ability and the generation ability of the child disease data can be met, and the time and space overhead needs to be reduced.

The application relates to the technical field of image enhancement, in particular to a mine environment super-resolution image reconstruction method based on vibration induction for mine monitoring, which comprises collecting vibration data and image data and performing pretreatment; the processed vibration data sequence is input into a double-layer LSTM network, vibration data is analyzed by using the LSTM, vibration characteristics at the next moment are predicted, and dynamic compensation parameters are generated; the processed image is input into a convolutional neural network constructed by using a blueprint separable convolution, a residual attention module and a coordinate attention mechanism, and multi-scale feature extraction is performed; time alignment of the vibration data and the image data is performed, corresponding compensation parameters are generated, and meanwhile, spatial alignment and fusion of the two are performed to obtain fusion features; based on the fusion features, super-resolution reconstruction is performed on the image to generate a high-resolution image. Through technical fusion innovation, the recognition quality of image quality in a complex vibration scene is significantly enhanced.

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.

This invention discloses a cross-modal data processing system for the safe operation of hydrogen refueling stations, including a raw data acquisition and processing module, a full-variable safety scanning module, a physical relationship coupling diagnosis module, an adaptive operating condition clustering module, a question-answer pair construction module, a hydrogen refueling station time-series command data acquisition module, and a fault type identification module. By constructing a three-layer semantic enhancement logic and model adaptation strategy, this invention effectively solves the technical problems in the prior art, such as the lack of supervision signals in the raw data of hydrogen refueling stations, the difficulty in identifying hidden faults under complex operating conditions, and the difficulty in adapting heterogeneous feature space models.

The application discloses a zero sample video classification method based on frame-level cache and video-level cache updating. First, visual features and text features of a test video are extracted. Then, cosine similarity calculation is performed on the test video and text prototypes of each category to obtain cross-modal zero sample matching results, which contain predicted pseudo labels and prediction result entropy. According to the prediction result entropy and time sequence probability difference criterion, the credibility of each frame is measured, and a two-stage Top-K strategy is adopted. First, high-confidence frames satisfying a preset value are selected according to the prediction entropy, and then frames most discriminative in time sequence are further selected according to the time sequence probability difference for caching. A video-level cache updating module is used to maintain the cache. Visual features of the test video are matched with visual prototypes in the cache of each category to obtain same-modal matching results. The cross-modal zero sample matching results and the same-modal matching results are fused to obtain final inference results.

The application discloses a large language model end-edge collaborative adaptive reasoning method, a terminal device and a storage medium, which are suitable for terminal devices and belong to the technical field of edge computing. The method comprises the following steps: acquiring working state information of a terminal device at a current time, channel state information between the terminal device and an edge server, a plurality of selected exit points of a preset large language model, exit point information of each selected exit point, and a current reasoning task; calculating the total reasoning time delay of the current reasoning task at each selected exit point and under different node modes, eliminating timeout exit points, and obtaining a plurality of initial exit points; then, the optimal exit point and the optimal node mode are selected from the initial exit points; and finally, the current reasoning task is reasoned according to the preset large language model, the optimal exit point and the optimal node mode. Through the implementation of the application, the problem that the existing technology uses a pre-set model division point for edge-end collaborative reasoning and easily leads to a decrease in reasoning efficiency can be solved.

The invention discloses a drop-out fuse state monitoring method based on multi-size features and deep learning, and the method comprises the steps: S1, forming a data set for obtained drop-out fuse pictures, and dividing the data set into a training set and a test set; s2, adding an improved receptive field block and a coordinate attention module to a YOLOx backbone network, adding an adaptive spatial feature fusion module to PANet, carrying out secondary fusion on features of different scales, then introducing a loss function of weighting loss and positioning loss fusion, and finally carrying out lightweight improvement, constructing a state monitoring model, and carrying out state monitoring. Performing training verification on the constructed state monitoring model through the training set and the test set; s3, identifying a drop-out fuse picture acquired in real time by adopting the trained and verified state monitoring model; according to the image data collected by the application, the construction of a special database is realized, the YOLOx is improved and lightweight operation is carried out, and the complexity of the model is reduced, so that the method is suitable for an embedded platform of an electric unmanned aerial vehicle, and the detection effect of the drop-out fuse is improved.

The application provides a SAR target scattering center parameter fast extraction method, and mainly solves the problems of slow speed and low precision of a traditional SAR target scattering center parameter extraction method; firstly, the application models a scattering center parameter solving problem based on a sparse representation theory; then, a SAR target scattering center parameter extraction problem model based on a semi-quadratic splitting method is constructed, a deep unfolding network is constructed according to an optimization process based on the semi-quadratic splitting method, and the SAR target scattering center parameter is extracted by using the deep unfolding network; compared with a traditional method, the method constructs a deep unfolding network based on the semi-quadratic splitting method and a scattering center model, and realizes efficient and interpretable SAR target scattering center parameter extraction.

The application provides an image reasoning method, device and equipment and a storage medium. The method comprises the following steps: inputting a user-inputted static image to be reasoned and a natural language prompt into a trained task routing module, identifying at least one task category corresponding to the natural language prompt, and activating a trained expert agent corresponding to each task category; inputting the static image to be reasoned and the natural language prompt into each trained expert agent for first reasoning, obtaining a shared semantic representation of each task category, and inputting the shared semantic representation into a collaborative memory pool for graph fusion to generate a consensus representation of the static image to be reasoned; and inputting the consensus representation of the static image to be reasoned and the natural language prompt into a trained target agent for second reasoning to obtain a final answer corresponding to each task category of the static image to be reasoned. The application avoids interference between tasks and improves the generality, reasoning efficiency, reliability and stability.

The invention discloses a coarse-to-fine image matching method based on an aggregation attention mechanism, and belongs to the technical field of computer vision. The method comprises the following steps: firstly, extracting multi-scale features of an input image through a lightweight heavy parameterized convolutional neural network; then, an aggregation attention module is used for efficiently converting coarse-grained features, and the module is used for aggregating tokens through deep convolution and maximum pooling and enhancing feature discrimination ability in combination with rotation position coding; then calculating a similarity matrix based on the converted features, and obtaining rough matching point pairs through double softmax operation; and finally, by taking rough matching as guidance, realizing sub-pixel-level accurate positioning on fine-grained features through a two-stage process of mutual nearest neighbor screening and space expectation calculation. According to the method, the calculation efficiency is remarkably improved while the high matching precision is guaranteed, and the method is suitable for scenes such as unmanned aerial vehicle visual positioning and navigation which have strict real-time requirements.

This invention discloses a high-efficiency video analysis method based on cloud-edge collaboration and adaptive MoE. The method includes: when an edge device performs a video analysis task, it first constructs a meta-request containing current task data and device resource constraints, and sends it to the cloud; the cloud uses a pre-trained basic model to extract features from the video data and clusters the data based on the feature space, dividing complex video scenes into multiple semantic subdomains; a lightweight expert model is trained in the cloud for each semantic subdomain; a routing decision model is further trained in the cloud; after deploying an expert model pool and routers on the edge device, for input data, the router first calculates the matching degree of each expert model and performs a comprehensive evaluation based on its computational cost. Through knowledge distillation, the large-scale basic model is decomposed into lightweight expert models, achieving a significant compression of the model size, enabling efficient deployment on resource-constrained edge devices.

This invention discloses a generative recommendation system and method based on quantized vector retrieval and LLM, belonging to the technical field of computer information processing and artificial intelligence recommendation systems. The method includes: Step S1, extracting continuous collaborative features of users and items; Step S2, constructing and pre-training an AHPQ segmenter quantization module, mapping continuous collaborative features to discrete ID token sequences and optimizing the quantized codebook; Step S3, using the pre-trained codebook vector to collaboratively initialize the LLM's token embedding layer; Step S4, fine-tuning the LLM using a freeze-adapt strategy, and then generating user preference vectors through a collaborative semantic fusion module; Step S5, calculating the similarity between the user preference vector and the item database vector, and retrieving Top-K items based on cosine similarity as the recommendation result. This invention solves the problems of ID discretization difficulties and collaborative signal loss in LLM recommendation, preserving the high-order collaborative structure while leveraging the semantic reasoning capabilities of LLM, and simultaneously achieving efficient reasoning and cold-start generalization.

The application discloses a kind of based on knowledge enhancement's power grid new energy operation fault discrimination method and system, it is related to wind farm fault diagnosis technical field.The method includes steps: obtaining the time series data of power grid new energy equipment operation, and according to the time series data of power grid new energy equipment operation graph structure is constructed, obtains graph data representation;According to the characteristics of power grid new energy equipment operation data expert knowledge is encoded as hierarchical rule set, and according to hierarchical rule set constructs knowledge rule base;With the knowledge enhancement mechanism of graph neural network constructs fault diagnosis model, the hierarchical rule set in knowledge rule base is as teacher, and graph neural network is as student, and the fault diagnosis model is trained;Abnormal detection is carried out to graph data representation using the trained fault diagnosis model, and wind farm fault discrimination result is obtained.The present application can realize the unsupervised accurate identification of complex fault scenarios such as gearbox fault, variable pitch system anomaly.

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 relates to the technical field of industrial manufacturing, and discloses a method and system for improving the completion parameters of understanding user input, which comprises the following steps: inputting an industrial standardized file, generating a semantic label, reasoning through the industrial standardized file and the semantic label, obtaining an optimized industrial standardized file, and obtaining an industrial rule set and a text parameter set input by a user. The application realizes accurate semantic label generation based on a WordNet synonym set, solves the polysemy ambiguity problem, calculates the similarity between a candidate parameter and a user conversation through an improved Levenshtein distance algorithm and a compensation mechanism, accurately captures the potential demand of the user, improves the consistency with the actual input intention of the user, multiplies the initial weight value, the click rate and the time decay factor to sort the candidate parameters, balances the instant production demand and the historical experience in the industrial scene, and significantly improves the accuracy, adaptability, efficiency and interpretability in four dimensions.

This invention provides a task execution method, apparatus, electronic device, and storage medium, relating to the field of embodied intelligence technology. The method includes: determining redundant feature components from the first task reference information based on the feature sources of feature components and the importance values ​​of each feature source, wherein the importance value of each feature source represents the magnitude of the influence of the feature component of that feature source on the successful output of the robot's expected action by the first VLA model; performing redundancy removal processing on the redundant feature components in the first task reference information to obtain processed first task reference information; inputting the processed first task reference information into the first VLA model to obtain the predicted action of the robot output by the first VLA model; and controlling the robot to execute the task based on the predicted action. Applying the solution provided by this invention can improve the robot's motion control frequency, thereby improving task execution performance.

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.

This invention relates to the field of artificial intelligence technology and discloses an AI application development platform that integrates convolutional networks. The platform includes: a model parsing module for parsing the model and calculating the spatial importance weights and stability coefficients of the channels; a policy generation module for generating a policy table containing load levels and binary masks based on the weights and coefficients; and an application building module for encapsulating the computing power control unit to generate the target application. During program execution, the computing power control unit determines the policy level based on the hardware state and uses masks to control the data flow to either enter the neural network processor for convolution or enter the graphics processing unit for motion vector multiplexing. This invention achieves heterogeneous hardware collaboration and dynamic policy fusion through the development platform, solving the high power consumption problem of frame-by-frame full computation and improving the inference efficiency of the application and the device's battery life.

The invention discloses an electric vehicle intelligent charging method and system based on reasoning, and belongs to the technical field of charging pile power utilization regulation and control, through real-time collection of related parameters during charging of an electric vehicle and construction of a lightweight multi-task learning network, a charging scene is predicted, the state of a power grid can be dynamically sensed, and the charging efficiency is improved. The charging strategy is dynamically adjusted according to the state of the power grid, and the multi-task parallel processing capability is improved through lightweight processing without deploying a plurality of independent models, so that the reasoning efficiency is improved, the charging efficiency is maximized on the premise of ensuring the safety of the power grid, and the balance between the charging speed and the power grid protection is realized; the problem that the operation efficiency of the charging pile is low due to the fact that static control logic is mostly adopted for electricity regulation and control of an existing charging pile, the power grid state is difficult to dynamically perceive and the output power is difficult to actively adjust according to the power grid fluctuation trend is solved.

This application relates to a method and apparatus for parallel reasoning of long texts based on linear attention. Applied to any parallel computing device in a distributed architecture, the method includes: converting a long text instruction to be processed into at least one embedded feature vector; constructing a local attention decay matrix and a global attention decay factor based on the at least one embedded feature vector and task attribute information of the long text instruction; mapping the at least one embedded feature vector to at least one core matrix based on a weight matrix; calculating the local attention matrix and the global attention matrix; constructing a total attention matrix for at least one embedded feature vector based on the local attention matrix, the global attention matrix, and linear attention; and calculating the unprocessed tokens or output text data corresponding to the long text based on task attribute information and the total attention matrix. This method can reduce the complexity of attention calculation in the long text reasoning process and shorten the computation time.

This invention belongs to the field of medical image processing technology, specifically relating to a surgical instrument segmentation method based on a visual foundation model and structural prior constraints. The specific process is as follows: The DINOv2 visual foundation model extracts visual features at different levels; a low-rank adaptation module is injected into the attention-related linear mapping layer of the DINOv2 visual foundation model encoder. The low-rank adaptation module freezes the original weight matrix of the linear mapping layer, sets two low-rank matrices to model the weight increment, and performs a linear projection mapping operation on the input features using the frozen original weight matrix; output features are obtained using a lightweight convolutional feature extraction unit constructed using GhostModule; after performing multiple dilated convolutions and average pooling operations with different dilation rates on the features using the Ghost-ASPP module, feature compression and fusion are performed to output multi-scale semantic features; based on the multi-scale semantic features, the surgical instrument segmentation result is predicted.