36 results about "Product quantization" patented technology

The invention discloses an unsupervised semi-pairing cross-modal retrieval method and system based on deep learning, relates to the field of artificial intelligence, and is used for solving the problems of annotation data dependence, asymmetric semantic association and high-dimensional storage efficiency. According to the method, a double-branch visual encoder and a dynamic prompt text encoder are combined, dynamic weighting of visual-text features is achieved through gating cross attention, and modal redundancy interference is restrained. An enhancement strategy is generated through low-frequency semantic guidance, and the long-tail word coverage rate is increased; a dual-stage quantitative hierarchical index is constructed, coarse-grained clustering and fine-grained product quantitative compression feature storage is adopted, and million-level data real-time retrieval is supported. A degradation aware increment maintenance mechanism monitors data distribution offset through a KL divergence threshold, and triggers index reconstruction to maintain long-term update precision. According to the method, limitation of a traditional strong pairing model is broken through, cross-modal sensitive content second-level positioning is achieved, asymmetric semantic alignment is effectively solved, and retrieval efficiency is improved.

The invention relates to an index and representation model joint optimization method and system and a readable storage medium, and the method comprises the steps: obtaining a data pair set formed by a plurality of query statements and corresponding query documents, labeling labels with different correlation levels, and storing the labels in a structured database; randomly selecting a plurality of data pairs, and encoding the data pairs by using the representation model to obtain corresponding query statement vectors and query document vectors; performing product quantization processing on the vectors and storing the vectors in a vector database; randomly selecting query compression vectors from the vector database, and selecting a plurality of corresponding document compression vectors as negative example document vectors by adopting a multi-stage negative example sampling strategy according to the labels; calculating a similarity score between the query compression vector and each corresponding negative example document vector, and constructing a loss function; and updating the parameters of the representation model and the clustering center through back propagation. According to the invention, calculation redundancy during query can be reduced, and query efficiency and response speed can be improved.

The invention discloses a single-stage generative sequence recommendation method, which belongs to the field of personalized recommendation systems, and comprises the following steps: S1, item code representation: performing formalized representation on items according to attribute granularity, and generating code representation of the items through a large language model and a product quantization technology; s2, constructing a user interaction sequence: splitting the user interaction sequence into a plurality of attributes to form an interaction sequence consisting of project attributes; s3, generative sequence recommendation: improving an encoder architecture of Transform, including a sequence projection layer, a spatial aggregation layer and a conversion layer, so as to adapt to a sequence recommendation task; s4, recommendation item prediction: based on the user interaction sequence, learning probability distribution, and generating prediction of a next interaction item; according to the method, prediction of the next item possibly interacted by the user can be directly generated, similarity query does not need to be carried out in the candidate set, and the sequence recommendation effect is effectively improved.

The invention relates to the technical field of intelligent information retrieval, in particular to an image-text retrieval generation method and device based on reference semantics and electronic equipment. The method comprises the steps of performing semantic segmentation on a document in a document knowledge base to obtain a plurality of paragraphs, performing semantic similarity judgment, and merging similar paragraphs; extracting semantic features of the merged paragraphs to form semantic features of lower-layer paragraphs; abstract description is carried out on the merged paragraphs, semantic features are extracted from abstract description, upper layer description semantic features are formed, and text level semantic features are constructed; image semantic features in the image knowledge base are extracted, normalization processing is carried out on the text level semantic features and the image semantic features, an inverted product quantitative index is constructed, and construction of an image-text reference semantic feature index is achieved; and processing the input text based on the image-text reference semantic feature index to obtain a retrieval result. According to the method, the problems of lack of previous knowledge, insufficient detail information and poor image-text relevance in image-text retrieval are solved.

The invention discloses a power data cross-domain dynamic trust evaluation and access control method, and belongs to the field of power data security management and control. The method comprises the following steps: S1, dynamic trust evaluation: acquiring context information of a user and equipment in real time, generating a trust value and a risk value by using a random forest algorithm, and dynamically calculating a trust degree; s2, adapting to a sequence recommendation algorithm of a large language model based on a product quantization technology: generating an access authorization strategy through behavior sequence embedding generation and vector quantization in combination with credibility; s3, fine-grained dynamic access control: dynamically adjusting permission distribution according to the sensitivity of the power data and a real-time trust evaluation result; and S4, user abnormal behavior identification: monitoring user behaviors in real time based on a TextCNN model, identifying abnormity and triggering a safety protection mechanism. According to the method, the safety and reliability of cross-domain access of the power data can be effectively improved, data abuse is effectively prevented, and the method has wide industrial application value.

The invention provides an air traffic control dialogue type speech recognition method and device based on multi-task learning. Cooperative optimization of voice segmentation and content recognition is realized by constructing a joint learning model of a voice feature representation learning module and a voice activity detection module. The method comprises the following steps: firstly, designing a segmented shielding comparison pre-training strategy, and constructing a supervision signal through a product quantization codebook, so that a speech feature representation learning module learns fine-grained acoustic features in unlabeled data; secondly, a double-branch joint training mode fine tuning model is designed, a main branch optimizes a voice recognition task through connection time sequence classification loss, an auxiliary branch improves voice segment boundary detection precision through binary cross entropy loss, and meanwhile, a dynamic task weight adjustment strategy is combined to balance a multi-target optimization direction; and finally, innovatively introducing a dynamic local window attention mechanism, and focusing context modeling of an effective voice segment according to a real-time segmentation index.

The invention relates to the technical field of artificial intelligence, can be applied to the fields of financial science and technology and medical science and technology, and discloses a voice coding and decoding method, device, equipment and medium. A continuous vector is obtained by encoding the data through an encoder trained by a first-stage mirror image architecture; segmenting a continuous vector into sub-vectors, matching the sub-vectors with corresponding sub-coding dictionaries through a quantizer to obtain sub-indexes, and combining the sub-indexes to generate an overall index; analyzing the overall index to obtain sub-indexes, calling sub-discrete vectors and splicing the sub-discrete vectors into a discrete vector; and reconstructing the discrete vector through a decoder trained by a second-stage non-mirror-image architecture to obtain a target speech spectrum feature, and converting the target speech spectrum feature into a target speech signal. According to the method, double-stage training is adopted, the first-stage mirror image architecture guarantees the coding stability, the second-stage non-mirror image architecture improves the decoding flexibility, and the product quantization technology is combined, so that the calculation efficiency and the storage overhead are balanced, and meanwhile, the voice reconstruction quality is improved.

The present invention relates to the field of private domain e-commerce technology, and specifically discloses a private domain e-commerce data search method and system based on big data. The present invention first collects user behavior characteristics (search, browsing time, product click volume) and purchase characteristics (repurchase cycle, time preference), constructs style preference models and purchase preference models, quantifies users' attention to product styles based on browsing data, and performs semantic expansion based on the explicit semantics of search keywords to capture users' potential needs. At the same time, segmented product quantization technology is used to compress the extended semantics and purchase preferences. A double-layer compact index is used to achieve parallel retrieval of explicit and associated needs. Finally, the structured search results are visualized, integrating basic product information and style extension features to meet users' potential extension intentions and avoid the need for users to conduct multiple repeated searches.

A computational storage device for executing DiskANN search is provided by an aspect of the present disclosure. The computational storage device comprises a NAND memory array configured to storage node chunks corresponding to nodes in the DiskANN. The computational storage device also comprises a non-volatile memory array configured to store multiple product quantization (PQ) vectors corresponding to the nodes of the node chunks. The computational store device also comprises a processing unit coupled to the NAND memory array and the non-volatile memory array. The processing unit configured to execute DiskANN search (DiskANNs) among the node chunks in the NAND memory array and the multiple PQ vectors in the non-volatile memory array according to a received search instruction, and output a search result.

This invention discloses an unsupervised semi-paired cross-modal retrieval method and system based on deep learning, belonging to the field of artificial intelligence, to solve problems of labeled data dependency, asymmetric semantic association, and high-dimensional storage efficiency. This invention combines a dual-branch visual encoder and a dynamic cue text encoder, utilizing gated cross-attention to achieve dynamic weighting of visual-text features and suppress modal redundancy interference. A low-frequency semantic-guided generation enhancement strategy is used to improve long-tail keyword coverage; a two-stage quantized hierarchical index is constructed, employing coarse-grained clustering and fine-grained product quantization to compress feature storage, supporting real-time retrieval of millions of data points. A degradation-aware incremental maintenance mechanism monitors data distribution shifts through a KL divergence threshold, triggering index reconstruction to maintain long-term update accuracy. This method overcomes the limitations of traditional strong pairing models, achieving second-level localization of cross-modal sensitive content, effectively solving asymmetric semantic alignment, and improving retrieval efficiency.

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for performing a machine learning task on a network input to generate a network output. In one aspect, one of the systems includes a neural network configured to perform the machine learning task, the neural network comprising an N-grammer layer and an output neural network, the N-grammer layer configured to: at each of one or more heads: receive a sequence of input embeddings; generate a discrete latent representation of the sequence of input embeddings by using a learned product quantization codebook; generate a plurality of n-gram indices from the discrete latent representation; and generate a latent n-gram representation of the sequence of input embeddings; and generate a sequence of output embeddings, and the output neural network configured to: receive the sequence of output embeddings; and process the sequence of output embeddings to generate the network output.

This invention is applicable to the fields of artificial intelligence and intelligent transportation, providing a modified vehicle identification method based on attribute fusion and fuzzy product quantization. In the comprehensive vector feature extraction, this invention employs vehicle and license plate detection, parallel feature extraction, and attribute fusion concatenation to obtain comprehensive features containing high-level license plate semantic information, as well as low-level vehicle model and color attributes. This eliminates background noise interference, enhances the response strength of three easily modified vehicle features, and helps compensate for the limitations of single features. Furthermore, it applies fuzzy clustering and cosine similarity distance calculation methods, which can accommodate the uncertainty of input information during modified vehicle retrieval. Compared to traditional Euclidean distance, cosine similarity is not affected by vector length, and can better handle differences in dimensions when retrieving modified vehicles. Compared to the original product vector method, this invention's method has higher retrieval accuracy and robustness in modified vehicle scenarios.

The present invention discloses a method for compressing the model of an electroencephalogram deep neural network, which uses the weights of the depthwise convolutional layer to complete the selection of target electrode channels, so as to avoid large-scale artificial design feature calculations or the design of dedicated channel selectors, and make full use of the characteristics of the electroencephalogram deep neural network EEGNet itself. At the same time, for the fully connected layer, the present invention uses a binary tree to complete the centroid category to which the input vector belongs, greatly simplifying the calculation process of the fully connected layer. Furthermore, in a manner combining target channel selection and product quantization of the fully connected layer, the compression of the electroencephalogram deep neural network model is completed. The method for compressing the model of the electroencephalogram deep neural network provided by the present invention can greatly reduce the computational complexity of EEGNet, reduce the deployment difficulty of the model on wearable, embedded and other devices, and thus better support the relevant research on brain-computer interface applications.

The invention discloses a product quantization method based on hierarchical fuzzy clustering, and the method comprises the steps: carrying out the spatial decomposition of a to-be-compressed high-dimensional vector of image or video data into a plurality of low-dimensional subspaces through employing a hierarchical decomposition strategy; in each low-dimensional subspace, a hierarchical fuzzy clustering method is adopted, so that each data point belongs to a plurality of fuzzy clustering centers at different membership degrees; and on the basis of a hierarchical fuzzy clustering result, a product quantization coding mechanism is adopted, so that the fuzzy clustering center of each subspace forms a corresponding codebook, data points are coded and compressed according to membership degrees of the data points in different clustering centers, and data compression of a high-dimensional vector of the image or video data is realized. The invention provides a compression method combining a hierarchical structure and a fuzzy clustering technology, and the efficiency, the accuracy and the robustness of vector compression are remarkably improved.

The application discloses a similar text retrieval method based on vector representation. The method first processes the text at the article level by using the word embedding + pooling method, obtains the text vector representation through Bert coding, constructs all text vector representations into a ball tree index by using the ball tree method, performs product quantization on all text vector representations and ball tree center points, and builds a PQ index; during retrieval, the query text is also converted into a vector representation, a distance matrix is calculated through product quantization, KNN retrieval is performed based on the distance matrix and the ball tree index structure, and k similar texts are obtained. The similar text retrieval method based on vector representation constructed by the application supports long text retrieval, can accurately and quickly retrieve massive text vector representations, has good effect on complex or multi-semantic text queries, is very friendly and helpful to users who are not familiar with the jargon of a certain field or want to do exploratory search, and finally, the search is more comprehensive and accurate compared with the IndexIVFPQ method in Faiss.

The present application relates to the technical field of power grid maintenance, and provides an intelligent generation method and electronic equipment for a power grid maintenance work ticket, comprising: in the case that the data amount of a business logic vector library is greater than a preset threshold, performing approximate nearest neighbor retrieval in the business logic vector library based on an inverted file index and a product quantization manner and a query vector to obtain a plurality of logic chain segments related to a current power grid maintenance task; in the case that the data amount is less than or equal to the preset threshold, performing accurate nearest neighbor retrieval based on a flat inner product index manner to obtain the plurality of logic chain segments; inputting the plurality of logic chain segments into a business logic reasoning model to perform task decomposition, step framework building and equipment correlation analysis, and obtaining a business logic thinking chain including equipment, equipment correlation constraints and operation rules, and combining original information to input work content generation model to obtain a work ticket. The present application solves the problem that related technologies cannot efficiently generate reliable work tickets.

The invention belongs to the technical field of information, and relates to a product quantization optimization method and system based on optimal subspace allocation. The coding process of the method comprises the following steps: dividing N input D-dimensional vectors into M subspaces, wherein each subspace comprises N D / M-dimensional subspace vectors; respectively mapping all the subspace vectors to d subspaces; executing a K-means clustering algorithm on the mapped subspaces to obtain approximate estimation code words of each subspace vector in d subspaces to which the subspace vector is mapped; for a subspace vector and d approximate estimation code words thereof in different subspaces, reserving a mapping position with a minimum residual absolute value to obtain a code of the subspace vector; and combining the codes of the subspace vectors to obtain a compression result of the input vector. The invention aims to improve and optimize a product quantization algorithm, can reduce quantization errors as much as possible while ensuring the decoding speed, and can be widely applied to the related fields of approximate nearest neighbor search and large-scale quantization models.

The invention discloses a question and answer method and device based on a large language model and electronic equipment, relates to the technical field of artificial intelligence, and aims to solve the problem that a question and answer result output by the large language model is poor in accuracy. The method comprises the steps that vectorization processing is carried out on a query text to obtain a query text vector, and the query text is a text obtained by carrying out keyword extraction on a to-be-queried problem; based on the query text vector, a plurality of first text vectors are determined from a vector database by using an inverted index algorithm, and the vector database comprises a plurality of pre-stored text vectors; determining a second text vector from the plurality of first text vectors by using a product quantization algorithm based on the query text vector; generating a target question based on the second text vector and the to-be-queried question; and inputting the target question into a pre-trained large language model for processing to obtain a target answer. According to the embodiment, the accuracy of the question and answer result output by the language large model can be improved.

The invention discloses an electric energy meter fault fusion detection method based on VMD-GRU, and the method comprises the following steps: data collection, data preprocessing, VMD feature extraction, IVF-PQ modal importance analysis, and CNN-GRU prediction. VMD (variational mode decomposition) and IVF-PQ (inverted file-product quantization) technologies are introduced, so that the recognition precision and reliability of the model on fault signals are further improved, the method has good performance on key indexes such as accuracy, accuracy, recall rate and F1 score, and efficient classification of normal signals and fault signals is realized.

The invention discloses an old and ancient book cover image retrieval method based on dynamic partitioning and multi-modal feature fusion, and relates to the technical field of image recognition, old and ancient book cover images are used as query vectors to be input into a retrieval model, and a matched book list corresponding to old and ancient books is obtained according to the query vectors; the training process of the retrieval model is as follows: performing visual feature extraction and text feature extraction on the cover image of the ancient and old book, and partitioning; inputting the blocked visual features and text features into a cross attention mechanism, and generating an original high-dimensional fusion feature vector including a layout text structure and image layout information; performing subspace decomposition and quantization coding on the original high-dimensional fusion feature vector by adopting a product quantization algorithm to obtain a compressed feature piece vector, and constructing an ancient and old book feature database according to the compressed feature piece vector; taking cover images of the old and ancient books as input query vectors, indexing the old and ancient book database by adopting a three-level progressive query architecture, and returning a matched book list; the trainable parameters of the retrieval model are jointly optimized by constructing a total loss function; according to the retrieval method, multi-level understanding and digital reconstruction of ancient book covers are achieved, and rapid retrieval is achieved through efficient index construction.

The invention discloses a vector database-based big data real-time file gathering system, which comprises a data acquisition and access module used for acquiring multi-source heterogeneous data and establishing a data access channel; the data preprocessing module is used for preprocessing the multi-source heterogeneous data; the vector modeling module is used for generating a normalized object feature vector through the improved TabPFN; the file-gathering index construction module is used for constructing file-gathering index structure entries through an improved reverse index structure and a product quantization coding method, and writing the file-gathering index structure entries into a vector database; the file-gathering updating module is used for carrying out similar retrieval on the entity objects to be filed and updating or newly creating file-gathering index entries; the archive output and visualization module is used for constructing a structured data table and a JSON (JavaScript Object Notation) data document and generating a visual archive map; and the incremental feedback learning module is used for executing incremental training on the improved TabPFN network. According to the method, the accuracy, the real-time performance and the processing efficiency of big data file gathering are improved.

The invention discloses an edge deployment-oriented continuous learning industrial anomaly detection method and device, and the method comprises the steps: obtaining industrial normal sample data, and dividing training tasks and corresponding training sets and test sets according to classes; a teacher-student dual-network architecture is constructed, a lightweight convolutional neural network is adopted as a backbone network, dual-network front-segment parameter sharing is set, and only student network rear-segment parameters are updated; inputting samples according to a task sequence, calculating teacher-student feature matching loss based on a difference enhancement knowledge distillation mechanism, and updating student network parameters; a replay mechanism is designed, and catastrophic forgetting is avoided in combination with KNN Greedy and random sampling; compressing and replaying sample features by adopting a product quantization method; after training, test data features are extracted, and anomaly scores are calculated to realize anomaly classification and segmentation. The method can meet the requirements of unsupervised learning, continuous learning and edge deployment at the same time, gives consideration to the detection precision and the adaptability of edge equipment, and is suitable for various industrial detection scenes.

The invention belongs to the technical field of information, and particularly relates to a sliding fence product quantification method and device. The method comprises the following steps: obtaining N input D-dimensional vectors which are regarded as an N * D matrix, and carrying out initial segmentation to obtain a small-value matrix, a large-value matrix and an indication matrix; then recursive segmentation is carried out, a small-value matrix and a large-value matrix which are obtained after recursive segmentation are combined into a new matrix, and subspaces are divided in a sliding window mode; k-means clustering is executed in each subspace, and a cluster center is used as a codebook to encode all vectors in the subspace; and obtaining a final code of each element through a Count-min operation, and combining the final codes to obtain a quantization result. According to the method, improvement and optimization of a product quantization algorithm are realized, the memory overhead can be effectively reduced on the premise of ensuring the precision, and the method can be widely applied to the fields of large language model weight quantization, vector database management, KV cache optimization, image compression, image compression and the like which need efficient matrix quantization.

The invention discloses a multi-modal data approximate query method and system based on a hybrid block chain, and relates to the technical field of block chains. According to the method, a verifiable index structure is constructed and used for describing the mapping relation between the compressed representation of the feature vector of the multi-modal data and the data pointer under the chain of the multi-modal data, the index state change is recorded only by using a very small space on the chain, and huge storage overhead caused by directly storing the original high-dimensional feature vector or complete data in the block chain is avoided. According to the method, feature vectors are compressed into PQ codes through product quantization, a verifiable lightweight cryptographic proof is generated in combination with an MPT structure, authenticity verification of a multi-modal approximate query result is achieved, a Merkle proof is independently verified through on-chain root hash, and it is ensured that the result is not tampered; the client downloads original data, executes full-precision feature extraction and distance calculation, and compensates precision loss caused by PQ compression; and a complete and credible closed loop of'efficient approximate search + accurate result output 'is realized.

A text retrieval method, apparatus and device, and a storage medium, the method comprising: reading text information in a file set according to a preset code tool to extract original data in the text information to construct a text tree; processing the text tree according to a multi-scale analysis and recombination technology to generate a plurality of structured text fragments; inputting the plurality of text segments into a preset training language model, and obtaining a text embedding vector corresponding to the output of the preset training language model; decomposing and quantizing the text embedding quantity according to a product quantization technology, and constructing a vector index database; according to the obtained to-be-retrieved text and the vector index database, the distance between the to-be-retrieved text and each text fragment in the vector index database is calculated to obtain a target text fragment corresponding to the to-be-retrieved text, and the problems that in the related technology, training depends on a large amount of corpus data, key information is lost, and the retrieval efficiency is improved are solved. And the technical problem that the retrieved text effect is poor is solved.

The application discloses a kind of self-supervised remote sensing image retrieval methods, comprising the following steps: obtaining remote sensing image data;Remote sensing image data is carried out data enhancement, and training data set is obtained;Based on cross quantization and self-adapting error correction label, construct initial self-supervised remote sensing image retrieval model;Using training data set, initial self-supervised remote sensing image retrieval model is trained, and self-supervised remote sensing image retrieval model is obtained;Through self-supervised remote sensing image retrieval model, actual remote sensing image retrieval is carried out.The application further discloses a kind of system for implementing the self-supervised remote sensing image retrieval method.The method model of the application reduces the load of storage and processing data by product quantization compression high-dimensional data, effectively improves the efficiency of retrieval;The application method uses cross learning strategy to improve retrieval precision;The application also designs self-adapting error correction label, effectively solves the sampling bias problem in contrast learning, improves retrieval precision.

A computational storage device for executing DiskANN search is provided by an aspect of the present disclosure. The computational storage device comprises a NAND memory array configured to storage node chunks corresponding to nodes in the DiskANN. The computational storage device also comprises a non-volatile memory array configured to store multiple product quantization (PQ) vectors corresponding to the nodes of the node chunks. The computational store device also comprises a processing unit coupled to the NAND memory array and the non-volatile memory array. The processing unit configured to execute DiskANN search (DiskANNs) among the node chunks in the NAND memory array and the multiple PQ vectors in the non-volatile memory array according to a received search instruction, and output a search result.

This invention provides a computing storage device, a system, and a method for operating the memory device. The computing storage device includes a NAND flash memory array for storing multiple node chunks corresponding to multiple nodes in a disk approximate nearest point. The computing storage device also includes a non-volatile memory array for storing multiple product quantization (PQ) vectors corresponding to multiple nodes in the multiple node chunks. The computing storage device also includes a processing unit coupled to the NAND flash memory array and the non-volatile memory. The processing unit performs a disk approximate nearest point search on the multiple node chunks in the NAND flash memory array and the multiple product quantization vectors in the non-volatile memory array according to a received search instruction, and outputs the search result.

This invention discloses a method and system for approximate multimodal data query based on hybrid blockchain, relating to the field of blockchain technology. The invention constructs a verifiable index structure to describe the compressed representation of multimodal data feature vectors and their mapping relationship to off-chain data pointers. It uses only a minimal on-chain space to record index state changes, avoiding the huge storage overhead of directly storing the original high-dimensional feature vectors or complete data on the blockchain. By compressing the feature vectors into PQ codes through product quantization and combining them with an MPT structure to generate verifiable lightweight cryptographic proofs, the authenticity of the multimodal approximate query results is verified. The Merkle proof is independently verified using on-chain root hashes to ensure the results have not been tampered with. The client downloads the original data, performs full-precision feature extraction and distance calculation, compensating for the precision loss caused by PQ compression; thus achieving a complete and reliable closed loop of "efficient approximate search + accurate result output".

The application discloses a neural network joint quantization method for image classification, solves the problem that a deep neural network for image classification is difficult to be efficiently stored and run on a resource-limited embedded device, and realizes compression and acceleration of the neural network. Scalar quantization is used to convert element values of a mapping codebook obtained through product quantization into 8-bit integers. In combination with 8-bit integer activation values, fixed-point operation of the neural network is realized, and the purpose of calculation acceleration is achieved. In order to eliminate the activation value distribution error caused by different inputs, the saturation truncation method is used to reduce the quantization interval of the activation value, and 8-bit scalar quantization of the activation value is realized. Finally, a pseudo-quantity unit is inserted into the neural network, a scalar quantization error is introduced, parameters are updated through back propagation, and the error is eliminated. The application fully combines the advantages of scalar quantization and product quantization, and can significantly reduce the model size while reducing the network operation amount.