471 results about "Encoding (memory)" patented technology

The invention discloses an intelligent optimization method based on lightweight language training data, which comprises the following steps of: constructing a self-attention mechanism by using a Transform architecture, giving an embedding matrix of an input sequence, n being the sequence length, d being the embedding dimension, representing Query, Key and Value matrixes, and controlling the product magnitude by using a scaling factor; a low-rank adapter is added beside an original weight matrix, a parameter update quantity is decomposed into a product of two small matrixes, and the original weight matrix is a sum of two low-rank matrixes introduced by LoRA. According to the method, an end-to-end personalized dialogue memory system is constructed based on a MiniMind framework, dynamic internalization and accurate recall of user habit information are realized on the premise that a local vector knowledge base does not need to be constructed through organic integration of a supervision fine tuning strategy and a LoRA technology, a dialogue memory coding mechanism is designed, and the memory efficiency is improved. User interaction data is directly mapped into an implicit memory unit of a model parameter space, a lightweight memory trigger-retrieval protocol is developed, key features are extracted through dialogue context, and long-term memory response is activated.

The invention provides a long document semantic coherence processing method and device, a storage medium and computer device.The method comprises the steps that block coding is conducted on a long document through a sliding window converter, and the sliding window converter comprises a window dividing unit, a position coding unit and a converter coding unit; the window division unit adjusts block positions according to sentences and paragraph boundaries and sets overlapping areas between adjacent blocks; the encoded document content is stored and managed through a memory management module, the memory management module comprises a chapter-level memory pool and a gating memory unit, the chapter-level memory pool stores the document encoding content, and the gating memory unit maintains entities and event timelines; performing deep semantic understanding on the document content through a semantic processing module, and calculating a multi-dimensional attention weight capture dependency relationship between paragraphs; providing a retrieval result through a retrieval module; and generating semantically coherent output content through a content generation module. According to the scheme, the semantic coherence of long document processing is improved.

The invention provides a self-adaptive recommendation method based on dynamic strategy optimization, and aims to solve the problems of memory bank update lag, retrieval strategy lag and insufficient long-tail article coverage of an existing recommendation system. A time sequence behavior coding module, a dynamic memory storage bank, a strategy control module and a heterogeneous fusion module are designed; and real-time optimized personalized recommendation is realized. Specifically, a time sequence behavior coding module is used for generating a user state vector with time decay, a dynamic memory storage library automatically updates historical records based on a mixed scoring mechanism, and a strategy control module outputs a learnable retrieval number, a weight coefficient and a fusion parameter. The heterogeneous fusion module realizes cross-channel feature enhancement through a dual-channel attention mechanism; according to the method, the problems of preference drift and long-tail recommendation are effectively solved, meanwhile, the online updating efficiency is improved, and more accurate and diversified recommendation results are provided for users.

A system and method for persistent cognitive computation with temporally synchronized multimodal processing implements a geometric approach to artificial intelligence through typed latent entities within a dynamic manifold substrate. The system maintains a latent manifold incorporating heterogeneous data modalities where local curvature reflects semantic density and typed entities are stratified according to structural properties. Temporal synchronization coordinates asynchronous multimodal data streams through generation of temporal alignment fields within the manifold that preserve semantic coherence across modal boundaries. Type-aware geometric operations enforce operation legality based on entity type and local manifold geometry, enabling structured recombination, compression, and traversal while preventing semantic distortion. The system executes synchronized manifold reorganization during idle periods through coordinated optimization operations including perturbation analysis and topological surgery. This architecture enables persistent memory through geometric encoding where frequently accessed concepts develop high-curvature regions and cognitive patterns emerge from usage-based manifold evolution.

The invention provides a large language model long-term memory method based on human cognition inspiration, which comprises the following steps of: calculating semantic similarity distribution of historical interaction text vectors according to a sliding window, and calculating local information entropy of the window; determining an event boundary based on the local information entropy difference of the adjacent windows; taking each discrete event as the memory of the model, and performing partition management on the memory; when the event memory partition is full, the memory event with the low retention rate is transferred to a long-term memory area; when the interactive content relates to historical information, retrieving related historical events in the event memory partition by using a contextualized memory retrieval method; and the large language model generates content for the user in combination with the current context and the retrieved related historical events. According to the method, memory coding and storage, multi-level dynamic memory management and event-level situational memory retrieval methods based on the event cutting theory are adopted, and therefore the problem that an existing large language model long-term memory method lacks dynamic memory and situational memory is solved.

A system and method for controllable lossy data compression employing a joint learning framework to efficiently compress and reconstruct input data while balancing compression ratio and reconstruction quality. The system comprises an encoding system, a temporal modeling system, and a decoding system, which are jointly optimized to minimize a combined loss function. The encoding system, such as a Vector Quantized Variational Autoencoder (VQ-VAE) compresses the input data into a compact representation, while introducing a controllable degree of lossy compression based on adjustable compression parameters. The temporal modeling system, such as a Multilayer Perceptron Long Short-Term Memory captures temporal dependencies in the compressed representation. The decoding system, such as a VQ-VAE decoder, reconstructs the input data from the compressed representation. By providing control over the trade-off between compression ratio and reconstruction quality, the system offers flexibility for diverse applications.

The invention discloses a pulse neural network training method based on adaptive threshold integer activation neurons, and the method specifically comprises the following steps: inputting a neuromorphic data stream or a static image, generating a pulse sequence through time coding, and taking the pulse sequence as the input of a model; in the backbone network, performing down-sampling and extracting multi-scale features by using an SNN module based on LIF neurons activated by an adaptive threshold integer; the visual tasks are divided into image recognition, image detection and image segmentation, and are respectively accessed to different processing modules for processing and outputting; according to the invention, on the basis of a Meta-SpikeFormer element module, the spiking neural network based on LIF neurons activated by an adaptive threshold integer value is used, so that the memory overhead during training is reduced, the network performance of direct training is improved, the application range of the spiking neural network is expanded, and the network is better applied to image classification, recognition and segmentation tasks.

According to one aspect, a computing system is provided that includes processing circuitry configured to receive input data from multiple interaction modalities of a user, generate a multi-interaction-modality user interaction history from the input data, and extract memories from the multi-interaction-modality user interaction history using a trained memory-extracting generative model. The memories include natural language text descriptions of interactions in the user interaction history generated by the trained memory-extracting generative model. The processing circuitry is further configured to store the memories in file storage having an associated database with a vector search interface configured to receive memory retrieval queries. The computing system may further be configured to receive a user message via an interaction interface, retrieve relevant memories, generate a response-generating prompt with the user message and relevant memories, and use the prompt to generate a response to the user message with a generative language model.

The invention provides a quadruped robot foot end terrain prediction method and system based on a long short-term memory network and reinforcement learning, and relates to the technical field of machine learning. Firstly, a teacher-student strategy model for predicting the foot end terrain of the quadruped robot is constructed, and the teacher-student strategy model comprises a teacher strategy model and a student strategy model; body sensing and privilege information and radar point cloud information of the quadruped robot are obtained, and a teacher strategy model is trained through the body sensing and privilege information, the radar point cloud information and reinforcement learning; the teacher strategy model is utilized to guide and train the student strategy model, and in the training process, the student strategy model uses a long short-term memory network to fit the privilege information coding value and the radar point cloud coding value in the teacher model; and the trained student strategy model is deployed on the quadruped robot for terrain prediction, stable walking in the unstructured terrain is realized, and the prediction robustness is improved.

The embodiment of the invention relates to the technical field of text processing, in particular to an event causal relationship identification method, device and equipment of a Chinese text. According to the scheme, the to-be-recognized Chinese text is obtained for word-by-word recognition and labeling, each word in the Chinese text is coded by adopting a new coding mode, the text code is generated, and then through the convolutional neural network and the bidirectional long-short-term memory network, the capture capability of the model on local features and context features in a long text is improved, and the recognition efficiency is improved. And fusing the local feature and the context feature vector to construct a full-text comprehensive feature representing the Chinese text, and finally generating a causal relationship labeling sequence according to the full-text comprehensive feature through a conditional random field to perform causal relationship prediction. The specific causal relationship identification weight is enhanced, the identification capability of the model on the causal relationship is improved, and the completeness and precision of the model on causal feature extraction are greatly improved.

One or more embodiments of the invention provide a memory resource dynamic allocation method and device, electronic equipment and a storage medium. The method comprises the following steps: acquiring memory behavior characteristic data, wherein the memory behavior characteristic data reflects behavior characteristics of a system or a process in the aspect of memory use; encoding the memory behavior characteristic data to obtain a state code of the memory behavior characteristic data; dynamically selecting a target prediction model according to an actual memory use operation at a historical moment, the actual memory use operation representing information which is generated after a memory allocation strategy at the historical moment is executed and reflects a strategy effect and a system state; obtaining a memory allocation strategy according to the state code and the target prediction model; and dynamically allocating memory resources according to the memory allocation strategy. According to the invention, the memory can be reasonably allocated, the resource utilization rate is effectively improved, and the cost and energy consumption are reduced.

The invention discloses a heterogeneous intelligent computing power optimization management scheduling system for accelerating a large model reasoning task, and relates to the technical field of computing power optimization management scheduling. The video memory fragmentation problem in a long sequence scene is converted into a controllable block migration task, and the performance bottleneck of a traditional video memory exchange mechanism is broken through; based on an operator-level scheduling strategy of a hardware capability fingerprint database, position coding and other compute-intensive tasks are accurately matched with vector instruction set hardware, and resource mismatch loss caused by black-box scheduling is eliminated; an expert selection process is reconstructed by an integer routing and counting sorting algorithm, near-lossless reasoning is realized at a limited node of an instruction set, and the potential value of an old computing power pool is activated.

The invention provides a multi-modal explicit memory system and device, a storage medium and a program product, and relates to the technical field of multi-modal data retrieval. The multi-modal explicit memory system comprises a multi-modal coding module used for converting multi-modal data such as text, visual and auditory data into vectorized multi-modal features; the unified memory representation module is used for realizing cross-modal alignment and constructing a multi-modal layer explicit memory bank; the self-adaptive rarefaction module is used for pruning or quantifying the multi-modal features; the hierarchical retrieval and fusion module is used for performing hierarchical retrieval and cross-modal fusion in the memory bank; and the cross-modal reasoning and integration module is used for integrating hierarchical retrieval results by using a large language model to obtain a reasoning result of comprehensive multi-modal information. According to the method, by processing text, image and audio information, a reasoning result integrating visual clues, auditory features and text contexts is output, and the data processing efficiency and the multi-modal correlation degree are greatly improved.

The invention belongs to the technical field of dialogue generation, and relates to a digital intelligent carrier pigeon WeChat intelligent dialogue generation method based on deep learning. Dialogue data are obtained, and semantic anomaly detection is carried out through a regular expression and a BERT model in combination with a public corpus; organizing dialogue data in a mode of constructing a tree structure and introducing a three-dimensional timestamp; for emotion information of emoticons, text and visual features of expressions are combined through a hybrid encoder, a dynamic attention mechanism is adopted, weighting and memory length adjustment are carried out according to importance of historical statements, after user portrait features are introduced, user features and dialogue codes are fused, and therefore, the emotion information of the expressions is obtained. In the generation stage, through a Transform-based model and a Beam Search algorithm, a diversity penalty term is introduced in the generation process, and generation parameters are dynamically adjusted according to a user portrait, so that the diversity and security of generated dialogues are ensured, the precision and coherence of a dialogue generation model are improved, and the technical problems of context coherence and personalized generation are effectively solved.

The invention discloses a multi-intersection traffic signal cooperative control method driven by a cross attention neural network, and the method employs a local cooperative Transform architecture, integrates a decision converter and a shared memory mechanism, and achieves the efficient modeling of a space-time dependence relation of a multi-intersection traffic state. The method comprises the following steps: firstly, through a memory head module, extracting a hidden state of each agent in a sequence modeling process, and updating global shared memory for supporting information interaction and strategy collaboration among multiple agents; and then, a cross attention module is adopted to carry out cross calculation on the local representation and the shared memory of each agent, so that dynamic perception and efficient modeling of the global state of the traffic system are realized. A backbone network of the model is based on Transform, and the understanding ability of time and space traffic characteristics is enhanced through position coding, self-attention and cross attention mechanisms. In the fine tuning stage, only the inserted Adapter module and the output layer are subjected to parameter updating.

The invention provides a large model connection context dialogue method based on an LSTM (Long Short-Term Memory) model, which belongs to the field of intelligent deep learning and comprises the steps of data preprocessing, input sequence coding, hidden state and memory cell initialization, decoding response generation, output sequence decoding, training and optimization, evaluation and adjustment. The method has high accuracy and response speed when processing a complex dialogue scene, and can be widely applied to the fields of intelligent customer service, intelligent assistants and the like.

This specification discloses systems and methods for enhancing the efficiency of transformer models during inference and training by precomputing and storing in memory a significant portion of operations in the first transformer layer. The stored precomputed outputs are retrieved from memory during runtime, reducing computational complexity and memory bandwidth requirements. This approach results in decreased latency, increased throughput, and lower cost-per-token. The disclosed techniques are particularly advantageous for transformer models that incorporate positional encodings within the attention mechanism, such as Rotary Position Embedding (RoPE) and other relative position encoding schemes. The method of offline precomputing involves calculating the outputs of the eliminated operations and components for each of the original vocab_size embedding-vectors, where vocab_size is the size of the embedding vocabulary. One embodiment of the invention removes the feedforward network and the attention query, key, and value projections from the first transformer layer of the encoder and the decoder stacks.

The invention relates to the technical field of complex equipment fault identification processing, in particular to a two-way decoupling and gated memory fused borescope image damage detection method, which is characterized in that a two-way decoupling and gated memory fused borescope image damage detection model is established, and training is performed only by adopting a normal image without a label; establishment of the two-way decoupling and gated memory fused borescope image damage detection model comprises feature decoupling, feature fusion and feature reconstruction, and in the feature decoupling stage, double-trunk encoders are adopted to serve as an effective information encoding branch and a redundant information encoding branch respectively; in the feature fusion stage, an adaptive gating and memory guiding feature fusion module is arranged, a structure perception reconstruction decoder is established in the feature reconstruction stage, fused representation features are sent to the structure perception reconstruction decoder for image reconstruction, and the structure perception reconstruction decoder adopts a joint optimization loss function.

The invention discloses a method and system for generating an EBT scene of a pilot based on a text and flight data, and the method comprises the steps: extracting threat error items in the operation of the pilot from a multi-source text through employing a scene perception word frequency-inverse document frequency method, and generating text analysis; a method of combining time feature attention, convolutional self-encoding and a long short-term memory network is utilized to comprehensively analyze flight data of daily routes and simulator training, dynamic features and potential risk factors in pilot operation are extracted, and flight data analysis is generated; in combination with text data analysis, flight data analysis and a scene element library, training scene elements are generated by adopting a hierarchical rule structure condition generative adversarial network method, and a teacher performs screening and combination to generate a personalized training scene. Through multi-source data fusion and data-driven analysis, a pilot personalized training scene is generated, the pertinence and effectiveness of training are improved, the use of training resources is optimized, and the overall flight safety and training quality are improved.

The invention discloses a large-scale video access and AI reasoning enhancement system and method, belongs to the technical field of video processing, and aims to solve the technical problem of how to realize large-scale video stream efficient access, resource elastic scheduling and abnormity self-healing in a complex environment. Comprising a resource dynamic scheduling module, a decoding and reasoning control module, a resolution dynamic processing module, a multi-stage shared memory transmission module, an exception self-healing and fault-tolerant module and a task parallel execution module, and resource allocation is optimized and calculated by dynamically binding a CPU core and GPU hardware unit isolation; the resolution is dynamically adjusted based on the scene algorithm precision requirement, and invalid calculation is reduced; a stream pushing and frame rate decoding strategy is controlled by using a Redis mark, and memory occupation is reduced by combining long and short queues; transmission coding and decoding are reduced by adopting a shared memory, and the cross-process interaction efficiency is improved; and task self-healing is realized through dual anomaly detection and process-level heartbeat monitoring.

According to one aspect, a computing system (10) is provided that includes processing circuitry (14) configured to receive input data from multiple interaction modalities (69) of a user, generate a multi-interaction-modality user interaction history (32) from the input data, and extract memories from the multi-interaction-modality user interaction history (32) using a trained memory-extracting generative model (30). The memories include natural language text descriptions of interactions in the user interaction history (32) generated by the trained memory-extracting generative model (30). The processing circuitry (14) is further configured to store the memories in file storage (18) having an associated database (19) with a vector search interface configured to receive memory retrieval queries. The computing system (10) may further be configured to receive a user message (47) via an interaction interface (48A), retrieve relevant memories (45), generate a response-generating prompt (50) with the user message (47) and relevant memories (45), and use the prompt to generate a response (58) to the user message (47) with a generative language model (56).

The invention relates to an online action detection method based on multi-stage memory mapping. The method comprises the following specific steps: S1, feature extraction: extracting frame-level features from a video stream by using a pre-training model; s2, memory division: dividing the extracted features into long-term memory and short-term memory according to time; s3, multi-stage mapping memory encoding: adopting a multi-stage mapping encoder for long-term memory, including segmented compression and global compression branches, compressing and extracting key information, and inhibiting irrelevant backgrounds; s4, cyclic decoding and updating: dynamically updating the short-term memory by adopting a cyclic decoder, and generating pre-judgment information of future actions; s5, model training and optimization: improving a loss function through design, optimizing the fusion effect of local features and global dependence in the model, and improving the overall detection performance; and S6, on-line action detection: detecting actions in real time based on the memory features updated by cyclic decoding, and evaluating performance according to actual action labels. According to the invention, the accuracy and robustness of action detection are improved.

The invention provides a multi-modal semantic fusion image and text relevance dynamic analysis method and system, and belongs to the technical field of multi-modal data processing. The method comprises the steps of image and text data preprocessing, feature extraction, fusion semantic vector generation through a bidirectional cross attention mechanism, dynamic relevance score calculation through a time sequence attention long-short-term memory network and combined loss function end-to-end training. According to the method, accurate alignment of image and text features can be realized through a bidirectional cross attention mechanism, the cross-modal matching accuracy is improved, the dynamic correlation analysis capability is enhanced by using time sequence modeling, and the problems of insufficient feature coding suitability and limited time sequence modeling capability in the prior art are solved.

A computer-implemented method of training an encoder neural network of an autoencoder, comprising: receiving a data stream at the autoencoder, the autoencoder comprising an encoder neural network, a decoder neural network, a first memory layer, and a second memory layer; and incrementally training the encoder neural network. The incremental training comprises: performing an encoding process on the input data by the encoder neural network to obtain a latent representation of the input data; processing the encoded input data and encoded input data stored in the first memory layer from previous iterations of the training steps to create a memory representation; performing a decoding process on the latent representation; processing the decoded input data and the updated memory representation to refine the updated memory representation; and outputting the refined memory representation to the encoder neural network for use in a next training step.

A computer-implemented method includes receiving enrollment audio from a user comprising a wake word to be enrolled for the device, preprocessing the enrollment audio to obtain a vector representation along at least a feature dimension and a temporal dimension, inputting the extracted vector representation to a trained encoding model to generate an embedding representation of the enrollment audio, wherein the encoding model includes a plurality of mixing blocks, and wherein the feature dimension and the temporal dimension of an output of a first layer of each mixing block are flipped for inputting to a second layer of the mixing block, and storing the generated embedding representation in a memory for use in detecting input of the enrolled wake word.

The invention provides an examination answer code similarity intelligent comparison and analysis method and system, and relates to the technical field of educational informationization, and the method comprises the steps: adaptively determining an optimal instrumentation position, collecting an execution path sequence and a memory access mode, and constructing a dynamic behavior feature vector; extracting static structure features and fusing the static structure features with the dynamic features; constructing a similarity measurement model by adopting a feature coding network and an adversarial discrimination network comprising three sub-networks; constructing a multi-level adversarial sample library based on a heuristic rule for adversarial training; and finally identifying and positioning the suspected plagiarism code snippets. According to the method, various plagiarism deformation strategies can be effectively identified, and the accuracy of code plagiarism detection is improved.

The invention relates to a video processing method applied to a processor of computer equipment. The computer equipment further comprises an acquisition unit, a memory unit, a blocking unit and a coding unit; the memory unit comprises a pre-allocated cache space; the method comprises the following steps: under the condition that an acquisition unit circularly fills acquired source video data into a cache space in a row filling manner, receiving a row interruption instruction which is triggered and generated after the acquisition unit finishes filling the cache space; in response to the row interrupt instruction, updating a first cache address of the cache space to a blocking unit; and after the blocking unit receives the data reading request sent by the coding unit and judges that the data reading request is released based on the first cache address and a second cache address carried by the data reading request, reading the source video data from the cache space based on the first cache address, and coding the read source video data through the coding unit. By adopting the method, the memory occupation and the coding delay can be reduced.

The invention discloses a man-machine interaction method based on an education robot and a system for intelligently generating a psychological counseling strategy based on interaction records, and constructs a progressive interaction path of tool use-emotional interaction-trust deepening by integrating high-frequency interaction scenes such as learning tutoring and home-school interaction. The system collects multi-modal interaction data through an entity and a software education robot, constructs a trust-intimacy two-dimensional evaluation model (T = 0.6 F1 + 0.4 E, I = 0.5 F2 + 0.5 E), and automatically generates a differentiated psychological counseling scheme based on a Becker T-type model and a four-quadrant strategy matching algorithm. The innovation points comprise a triple authorization privacy protection mechanism (block chain evidence storage and dynamic anonymous ID generation), a multi-modal sentiment analysis engine (a bidirectional coded representation model (BERT), a long short-term memory network (LSTM) and an FACS), and a rule engine and XGBoost mixed strategy generation architecture, and the whole process intelligence from interaction and data analysis to precise psychological counseling strategy generation is realized.

A three-dimensional memory array comprising a plurality of NAND strings. Each NAND string is configured to store a storage data having multiple bits and receive a searching input having multiple bits. Each bit of the storage data and the searching input is of a bit value “0”, a bit value “1”, a wildcard bit or an invalid bit. At least one bit of the storage data is configured as data bit(s), at least bit of the searching input is configured as search bit(s). All of the data bit(s) and the search bit(s) are of bit value “0” or bit value “1”. When the position of the data bit(s) in the storage data overlaps with the position of the search bit(s) in the searching input, the NAND string that stores the storage data is turned off; otherwise, the NAND string that stores the storage data generates an output current.

One or more systems, devices, computer program products and / or computer-implemented methods of use provided herein relate to caching key-value vectors for fixed prefixes according to a caching policy. A system can comprise a memory that can store computer-executable components. The system can further comprise a processor that can execute the computer-executable components stored in the memory, wherein the computer-executable components can comprise an accessing component that can access a request comprising a fixed prefix and variable data. The system can further comprise a storage component that can selectively store in a cache, one or more key-value vectors generated via processing of the fixed prefix by a large language model (LLM), based on a temporal moving average of the fixed prefix being greater than a first defined threshold, according to a smart encoding policy.