43100 results about "Term memory" patented technology

A system and method for implementing a convergent intelligence fabric (CIF) for distributed artificial intelligence operations. The CIF architecture integrates tensor-theoretic foundations, probabilistic cache management, precision-aware memory operations, quantum-resistant security, and neural-based optimization within a unified framework. The system orchestrates asynchronous, multi-hop data flow among computational resources while maintaining data security through per-block encryption and identity-based access control. Key components include a universal multi-model KV cache subsystem, agent-parallel disaggregation pipelines, reinforcement learning-based orchestration, and neuromorphic memory integration. Advanced implementations incorporate graphon-enhanced memory for sparse graph sequences, multi-modal cognitive persistent memory, and quantum-resistant asynchronous multi-domain trust protocols. The system enables efficient cross-agent collaboration, sophisticated knowledge sharing, and secure cross-domain operations while optimizing computational resources and maintaining strict privacy guarantees across distributed AI deployments.

An apparatus comprises a memory and a processor communicatively coupled to one another. The memory may be configured to store a data lake and multiple existing spectrum administration service (SAS) configuration commands. The processor may be configured to perform first SAS operations in accordance with the existing SAS configuration commands, collect multiple channel parameters from one or more communication channels configured to provide connectivity between user equipment and a core network, store the channel parameters in the data lake, monitor the channel parameters in the data lake, and generate optimized SAS configuration commands based at least in part upon the channel parameters. Further, the processor is configured to compare the optimized SAS configuration commands to the existing SAS configuration commands and perform second SAS operations in accordance with the optimized SAS configuration commands.

A coherent, intelligent, packet-switched memory fabric enables predictive, cache-coherent access across distributed compute, accelerator, and memory resources using a Memory-Fabric Transaction Layer Protocol (MF-TLP). MF-TLP defines routable packet formats for read, write, vectorized, atomic, reduction, collective, and predictive-prefetch transactions executed by memory-centric network interface controllers (MC-NICs). Each MC-NIC performs packet parsing, address translation, coherence management, and near-memory arithmetic or tensor operations while coordinating with MF-TLP-aware switches providing hierarchical directory control, multi-path routing, and in-network aggregation. Vectorized and multimodal packets encode multiple addresses or tensor offsets to reduce scatter / gather overhead, and programmable caching and quality-of-service modules manage tiered memory and tenant fairness. MF-TLP supports extension headers for predictive prefetch, collective coordination, and tenant governance, operating across hierarchical leaf-spine topologies using Ultra-Ethernet Transport, InfiniBand, or CXL fabrics. The system delivers scalable, low-latency, memory-centric orchestration for large-language-model training, multimodal AI, and data-intensive analytics.

The embodiment of the invention provides a network flow anomaly monitoring method and device based on a BiLSTM-Att network, and the method comprises the steps: obtaining network flow time sequence data, system resource use data and service call topological data, constructing a multi-modal time sequence feature matrix through employing a sliding time window method, calculating the importance score of each feature in the traffic feature subset, the resource feature subset and the service feature subset through an information gain algorithm, and screening the most discriminative feature based on a score threshold to construct a fusion feature subset; inputting the fusion feature subset into a set bidirectional long short-term memory-attention network, performing multi-level analysis on target features, performing comprehensive evaluation on a detection result by adopting a confidence-based multi-view decision fusion method, and generating a detection report containing an anomaly level, an anomaly type and an anomaly reason; according to the invention, the efficiency and accuracy of network flow anomaly monitoring based on the BiLSTM-Att network can be effectively improved.

Computer-implemented systems and methods for synchronizing data for dataset execution. The system includes a source database that stores a canonical dataset, a secondary database that stores a processed dataset, and a synchronization server that comprises a processor and a memory. The processor is configured to monitor for a publication of one or more source tables and when the publication is detected, identify the processed tables, corresponding to the source tables, to be updated in the processed dataset. The processor determines a tolerance level corresponding to each processed table and updates the processed tables in the processed dataset. The processor 10 determines whether the processed tables in the processed dataset were successfully updated within the tolerance levels and transmits a notification based on determining whether the processed tables in the processed dataset were successfully updated within the tolerance levels.

The invention relates to the field of data processing, in particular to a memory retrieval method based on a large language model and a related device. The method comprises the following steps: in response to an interaction instruction sent by a target user, performing context analysis processing on the interaction instruction to obtain a real-time interaction intention of the target user; performing memory retrieval on a short-term memory module and a long-term memory module based on the real-time interaction intention to obtain target retrieval information; and in combination with a personalized knowledge base of the target user, carrying out personalized processing on the target retrieval information to obtain response information matched with the use habit of the target user, and using the response information to respond to the interaction instruction. According to the method, the memory capacity of the large language model in the interaction process is improved, the information forgetting rate is reduced, the information retrieval efficiency is improved, effective classification and organization of personalized information are achieved, and the user experience is improved.

A digital product network system includes a set of digital products each having a product processor, a product memory, and a product network interface. The digital product network system includes a product network control tower having a control tower processor, a control tower memory, and a control tower network interface. The product processor and the control tower processor collectively include non-transitory instructions that program the digital product network system to generate product level data at the product processor, transmit the product level data from the product network interface, receive the product level data at the control tower network interface, encode the product level data as a product level data structure configured to convey parameters indicated by the product level data across the set of digital products, and write the product level data structure to at least one of the product memory and the control tower memory.

The invention provides a chip production logistics optimization scheduling method and system based on deep reinforcement learning, and relates to the technical field of deep reinforcement learning. An improved bidirectional long-short-term memory network and a graph neural network are used for extracting a material flow time sequence rule, equipment operation cycle characteristics and logistics network topology characteristics, a fusion state representation is generated through a contrast learning network, a layered and collaborative deep reinforcement learning architecture is constructed, an improved encoder-decoder structure is used for generating a strategy vector, and the strategy vector is constructed. And performing hierarchical joint training through a multi-agent execution network, inputting a real-time state and a strategy vector into the trained deep reinforcement learning architecture, generating a scheduling instruction, and performing hierarchical model prediction control in combination with a Bayesian neural network ensemble model and a recurrent neural network prediction model to realize optimized scheduling of chip production logistics.

The invention relates to the field of power equipment data processing, in particular to a biomass power generation combustion parameter deep learning method and system, and aims to solve the problem of phase mismatch caused by sampling frequency difference and clock reference offset of multi-source heterogeneous time sequence data. A parallel multi-scale convolution and bidirectional long-short-term memory network hybrid model is constructed, transient fluctuation and long-period trend features are extracted, and combustion stage feature weights are dynamically distributed through a gating attention mechanism. The optimization control module generates a multi-target constraint condition, an operation instruction is output in combination with a fuzzy inference engine, and a digital twin platform simulates an extreme working condition to enhance model robustness. A closed-loop feedback mechanism dynamically adjusts model parameters through combustion efficiency monitoring data and simulation results, and a two-stage fault-tolerant strategy realizes sensor abnormity compensation and historical control strategy backtracking. The problem of asynchronous data stream feature misalignment is effectively solved, and the combustion efficiency prediction precision and the control decision reliability are improved.

Systems and methods for computing time series predictions are provided, including a two-stage classifier-and-regressor processor. The classifier is trained on the complete data set while the regressor is trained on a pruned dataset. The classifier includes an Extreme Gradient Boosting classifier. The regressor includes an attention mechanism and a Long Short-Term Memory (LSTM) neural network. For a series of successive time period computations, a current output of the LSTM neural network is recursively fed back as an input to the attention mechanism for a subsequent time period computation. The output of the regressor is scaled by the output of the classifier to adjust for overfitting caused by the pruned training dataset.

The invention discloses a big language model reasoning optimization method and system based on dynamic video memory compression and memory isomerism, and intelligent management of video memory resources is realized by integrating a dynamic compression strategy of KV Cache and a memory parallel architecture. The method comprises the following steps: 1) analyzing the spatial-temporal characteristics of the KV Cache in real time, adaptively selecting a quantization compression algorithm, a rarefaction algorithm or a low-rank decomposition algorithm, performing hierarchical storage based on attention head importance scores, keeping high precision of a core head, and implementing low-bit quantization on a secondary head; (2) the compressed inactive data are divided into a plurality of data blocks to be stored in a system memory, a parallel data channel group is established according to the number of physical channels, the compressed blocks are concurrently read through multiple channels during loading, and parallel decompression of a sparse matrix is accelerated through a GPU tensor core; and 3) constructing a KV Cache multiplexing mechanism and a parallel channel, and parallelizing a compression / decompression process and model calculation by adopting a hardware acceleration compression and asynchronous pipeline mechanism.

An example operation may include one or more of receiving content from a conversation between a user on a user device and a second user on a second user device that is connected to the user device via a network, identifying a topic of the conversation based on execution of a generative artificial intelligence (GenAI) model on the received content from the conversation, identifying a call list that is previously stored in memory that is associated with the topic of the conversation based on keywords included in the identified topic of the conversation, and adding an identifier of the user to the call list stored in memory.

A system comprises a communications module; at least one processor coupled with the communications module; and a memory coupled to the at least one processor and storing processor-executable instructions which, when executed by the at least one processor, configure the at least one processor to provide, via the communications module, a first encryption key of an encryption key pair to a client device; receive, via the communications module and from a conversation agent server, a fulfillment request based on a natural language input transmitted from the client device to the conversation agent server; determine that the fulfillment request includes a request for personal data; obtain the requested personal data; encrypt the personal data with a second encryption key of the encryption key pair; and provide, via the communications module and to the conversation agent server, the encrypted personal data for transmission to the client device.

The invention relates to the field of power equipment control data processing, in particular to a boiler combustion optimization control method based on data driving, which comprises the following steps of: acquiring multi-source data such as temperature field distribution, air and smoke pressure, smoke components and coal quality characteristics, and eliminating noise interference by adopting sliding window mean filtering; generating a standardized feature matrix in combination with principal component analysis and a dynamic time warping algorithm; constructing a dynamic coupling model fusing a gradient boosting decision tree and a long short-term memory network, analyzing a nonlinear relationship between pulverized coal particle size distribution and a wind-coal ratio, and predicting combustion efficiency, pollutant concentration and temperature field uniformity; and model parameter self-correction and weight dynamic adjustment are triggered through actual combustion data feedback, and a closed-loop control link is formed. According to the method, accurate modeling of the multi-physical field coupling characteristic of the combustion system is achieved, the time sequence generalization ability under the dynamic working condition is improved, and the purposes of heat efficiency improvement and pollutant emission reduction are effectively balanced.

Multi-modal models learn a joint latent space for relating data points across different modalities. To more effectively learn multi-modal models with reduced training requirements and greater benefit from limited multi-modal training data, a multi-modal model may be trained with fixed or pre-trained unimodal encoders that generate data representations in respective latent spaces. The multi-modal model is trained to learn a shared latent space while fixing the unimodal encoders, enabling training without storing the unimodal encoders in memory. Limited multi-modal data may also be augmented by generating synthetic data between commonly-labeled pairs in the respective modality's latent spaces. The effect of data diversity can also be determined by generating a diverse data set with respect to the data points in latent space, enabling measurement of performance of the multi-modal model on limited training data.

The invention provides an inventory optimization method and system based on a deep learning model, and relates to the technical field of deep learning, and the method comprises the steps: obtaining the multi-dimensional data of a commodity, constructing a dynamic convolution kernel through employing a graph neural network, constructing a soft clustering distribution matrix in combination with a multi-layer perceptron, and obtaining a long-term dependence feature fusing a space dimension and a time dimension. And fusing the short-term and long-term features, inputting the fused short-term and long-term features into a bidirectional long-short-term memory network to generate a sales prediction sequence, and calculating a prediction confidence interval to obtain a final sales prediction result. And calculating a safe inventory level based on a sales prediction result, generating a replenishment decision by using a dynamic planning model based on deep reinforcement learning in combination with an inventory cost function, and finally sending the decision to a supply chain execution system. The sales volume is accurately predicted through the deep learning model, the inventory management is optimized, the storage cost, the stockout cost and the overstock cost are reduced, and the supply chain efficiency is effectively improved.

The embodiment of the invention provides a chip platform resource dynamic scheduling method, which comprises the following steps of: deploying a resource monitoring probe to acquire a chip platform running state, constructing a resource use index database to record CPU utilization rate and memory bandwidth data, analyzing a time sequence characteristic and a change rule of the running state, establishing a load model, setting a resource threshold value based on the system bottleneck identification model, formulating a scheduling rule through a strategy generator, and creating a resource configuration scheme according to the resource threshold value and the scheduling rule; analyzing a system state through a system bottleneck identification model, selecting a corresponding optimization strategy according to a scheduling rule, constructing a feedback acquisition queue, evaluating a scheduling effect of an execution instruction by using a data analysis technology, inputting the scheduling effect into an evaluation model for verification, and updating the scheduling rule based on a verification result of the evaluation model; according to the invention, the resource management efficiency and performance of the system can be effectively improved, and more flexible cross-platform support is realized.

Described herein are a system, method, and device for configuring a generative machine learning model using a syntactic interface. A system may include a user interface, a memory, and a processor configured to, using a syntactic interface displayed using the user interface, receive a syntactic interface input from a user; identify an electronic medical record (EMR) by generating an EMR database query as a function of the syntactic interface input, querying an EMR database using the EMR database query, and receiving, from the EMR database, an EMR database response; generate a prompt as a function of the syntactic interface input; generate a first generative model output as a function of the prompt and the EMR using a trained generative machine learning model; and using a conversational interface displayed using the user interface, display the first generative model output to the user.

Multi-modal models learn a joint latent space for relating data points across different modalities. To more effectively learn multi-modal models with reduced training requirements and greater benefit from limited multi-modal training data, a multi-modal model may be trained with fixed or pre-trained unimodal encoders that generate data representations in respective latent spaces. The multi-modal model is trained to learn a shared latent space while fixing the unimodal encoders, enabling training without storing the unimodal encoders in memory. Limited multi-modal data may also be augmented by generating synthetic data between commonly-labeled pairs in the respective modality's latent spaces. The effect of data diversity can also be determined by generating a diverse data set with respect to the data points in latent space, enabling measurement of performance of the multi-modal model on limited training data.

A data storage system can implement a key-value engine configured for tunable read, write, and space amplification. The key-value engine can support multi-versioning, synchronous and asynchronous key updates, and read snapshots. The key-value engine is highly scalable and can support generalized parallel, in-memory computation. Experimental results demonstrate that a key-value engine consistent with disclosed embodiments can outperform a state-of-the-art production LSM-based key-value store in a wide range of metrics.

Apparatuses, systems, and methods for half-page modes. A memory device may be operated in a full-page mode where all the memory cells along a word line are used for data or a half-page mode where less than all of the memory cells are used for data. In some memory devices, each half of the memory cells may be separately activated by different word line portions. In some half-page modes, data may be stored along a selected portion of the memory cells and additional information such as metadata or module parity may be stored along the non-selected portion of the memory cells. The additional information may be provided along additional data terminals so as not to increase the data burst length.

The invention discloses a deep learning-based deep fusion network production line fault prediction method, which comprises the following steps of: firstly, acquiring SCADA (Supervisory Control And Data Acquisition) system or sensor data of an automatic production line, and carrying out data preprocessing, feature extraction and fault prediction label construction; then a deep fusion network 1D-CNN-LSTM-Attention-DCE model combining a one-dimensional convolutional network, a long short-term memory neural network, a fusion attention mechanism and a dynamic cross entropy loss function DCE is established, the preprocessed data is input into the constructed model for training, and finally the trained model is used for fault prediction to obtain a fault prediction result. According to the method, a 1D-CNN-LSTM-Attention-DCE fault prediction model is provided, local features are extracted from time sequence data by using the 1D-CNN, then a long-time dependency relationship is captured through LSTM, and finally weight distribution of key features is enhanced through an attention mechanism, so that the processing capability of the fault prediction model on complex time sequence data is comprehensively improved, and the fault prediction efficiency is improved. And the fault of the production line can be accurately predicted within a certain time length range.

A codeword read from memory includes data blocks including data and supplemental blocks including error correction code (ECC) symbols for detecting and correcting data errors. Metadata can be stored in the supplemental blocks to increase memory utilization but using bits of the supplemental blocks for metadata leaves too few bits remaining for the ECC symbols. To maintain error protection, the supplemental blocks include ECC symbols to protect a first data portion of the codeword and parity bits configured to protect a second data portion of the codeword. Errors in the first data portion can be located and corrected using the ECC symbols. Errors in the second data portion can be detected by the parity. For example, the first data portion is encoded based on the second data portion, so locations of parity errors correspond to locations of symbol errors, and parity errors can be corrected.

Methods and apparatuses for a system error-correction code function are presented. The apparatus includes a memory configured to communicate with a host via at least one data connection and at least one non-data connection. The memory includes a memory array. The memory array includes a first portion and a second portion. The memory is further configured to, in a first mode, store and output data in the first portion and the second portion of the memory array. The first portion is addressable by a first address, and the second portion is addressable by a second address. The memory is further configured to, in a second mode, receive ECC of the data from the host via the at least one non-data connection, store the data in the first portion of the memory array, and store the ECC of the data in the second portion of the memory array based on the first address.

The invention relates to the technical field of data processing, in particular to a storage and calculation integrated parallel processing system and method. The method comprises the following steps of obtaining original stored data and performing topological skeleton projection, generating a projection skeleton structure to realize characteristic hierarchical integration, determining a weight relation matrix based on multi-scale space fusion data, realizing matching of a memory and a calculation unit through gradient optimization adjustment, and obtaining a multi-scale space fusion matrix. And the mining storage and calculation unit activates the chain type interaction relationship of the mapping data and performs recursive self-calibration, generates self-calibration update data, predicts a heterogeneous access channel and reconstructs a data access link, and finally performs task recombination and pipeline scheduling and eliminates parallel conflicts, thereby obtaining an efficient parallel processing result. According to the method, high efficiency and intelligentization of storage and calculation are realized, and reliable support is provided for execution of complex calculation tasks.

A memory device includes a bank array including a plurality of memory cells, a row decoder connected to the bank array through a plurality of wordlines, and a column decoder connected to the bank array through a plurality of column select lines. The bank array may include a first region and a second region different from the first region. First metadata for first normal data stored in the first region is stored in the second region, and second metadata for second normal data stored in the second region is stored in the first region.

The invention relates to the technical field of electronic manufacturing equipment, in particular to an adaptive thermal compensation system and method for a high-precision mounting head of a chip mounter, and the system comprises a temperature-deformation sensing unit, a thermal-mechanical coupling analysis unit, a dynamic compensation control unit, and a closed-loop execution unit. The temperature-deformation sensing unit collects temperature and deformation data of multiple parts of the mounting head in real time, the thermal-mechanical coupling analysis unit reconstructs a three-dimensional temperature field based on a finite element algorithm, the thermal expansion distribution quantity is dynamically calculated, the problem of rough model of traditional single-point temperature measurement is solved, and the measurement precision is improved. The dynamic compensation control unit predicts the thermal drift amount in the future 5 ms through online parameter identification and a long-short-term memory network model, a compensation strategy is adaptively adjusted in combination with the motion working condition, the closed-loop execution unit decomposes the compensation amount into displacement and torsion correction instructions, accurate offset of thermal deformation is achieved, and a whole-process thermal compensation closed loop is constructed. The precision stability of the mounting head in a complex thermal environment is improved, and the production efficiency is improved.

The invention relates to the technical field of resource allocation, in particular to a multi-channel video stream cooperative transmission method based on dynamic priority, which comprises the following steps of: acquiring task context characteristic parameters, system resource states, user behavior responses and computing node load data in real time, and constructing a priority allocation model to carry out pattern recognition to generate real-time task priority. And establishing a task scheduling strategy generation model, and performing resource allocation by adopting a priority weighting efficiency evaluation function, a memory-computing unit occupancy rate prediction matrix and a gradient optimization target under an adjustment resource constraint condition to form an initial scheduling strategy. And performing multi-dimensional parameter fusion analysis on the data through an adaptive optimization decision model, generating a resource redistribution correction vector, dynamically adjusting an initial scheduling strategy by adopting an online iterative optimization mechanism, and outputting a final real-time optimization task scheduling strategy. The task priority dynamic evaluation and the closed-loop optimization of the resource allocation are realized, and the efficient cooperative execution of the multi-channel video processing task is ensured.

The invention provides a power prediction precision improvement method and system for a wind and light station, and relates to the technical field of new energy power generation, and the method comprises the steps: obtaining the historical power, weather and equipment operation data of the wind and light station through a data preprocessing module, and carrying out the time sequence decomposition, spatial interpolation, abnormal value detection and correction, and obtaining the fusion feature data. And constructing a deep learning prediction model based on the fused feature data, extracting spatial features by using a convolutional neural network, processing the spatial features by using a long-short-term memory network to obtain an initial prediction result, introducing an attention mechanism to perform weight allocation to obtain weighted features, and combining to obtain power prediction data. And judging the weather type according to the real-time meteorological condition, and correcting by adopting the matched correction model when the extreme weather is identified. And finally, an error compensation model is constructed based on historical prediction errors for compensation, a final power prediction result is obtained, error correction is performed according to actual power data, and compensation model parameters are updated.

Apparatuses, systems, and methods for shared codeword in two-pass access operations. The memory may use a read read modify write write (RRMWW) cycle to write data and metadata to the array. Metadata and a data codeword are read out as part of two read access passes and combined into a shared codeword. Error correction is performed on the shared codeword, and then the corrected shared codeword is modified with write data and metadata. Updated parity is generated based on the modified shared codeword and the modified data and updated parity and the metadata are written as two write access passes.