53 results about "Encoding (memory)" patented technology

This invention provides a method and system for constructing a cross-device, cross-modal user memory graph. The method includes acquiring multimodal data from multiple devices and applications, uniformly parsing it into standardized memory atomic events with timestamps, calculating causal relationships between adjacent events during real-time processing, linking highly correlated event pairs into instantaneous event chains, performing deep encoding of the event chains using an offline dual-stream temporal semantic embedding model, and inferring candidate graph substructures using a graph neural network. When fusing these substructures into the user's main graph, relationship conflicts are resolved through confidence comparison, retaining highly reliable results, identifying stable subgraphs, and using a multi-head attention graph summarization network to abstract the subgraphs into higher-order memory concept nodes, thereby achieving continuous optimization and updating of the graph.

This application discloses a method and apparatus for managing video memory blocks, belonging to the field of data processing technology. The method includes: acquiring the occupancy status and access activity curves of video memory blocks within multiple consecutive address segments; for each consecutive address segment, performing frequency domain transformation and encoding based on the occupancy status and access activity curves of the video memory blocks within the segment to generate a video memory convolution spectrum characterizing the fragmentation degree and temporal access activity of that segment of video memory; determining candidate migration regions based on the video memory convolution spectra of each consecutive address segment; determining the differential tensors of candidate video memory blocks within the candidate migration regions; performing a comprehensive evaluation based on the differential tensors of all candidate video memory blocks under preset bandwidth and quality of service constraints to generate a video memory migration plan; and calling the underlying interface to execute the batch migration of candidate video memory blocks according to the video memory migration plan.

This invention discloses an adaptive frame abstraction interface management method and system for multiple encoding formats, relating to the field of data processing technology. The method includes: constructing a format configuration state diagram and a page transition time sequence; identifying the page transition time interval for corresponding format jumps; extracting structural difference parameters of the preceding and following formats and making them dimensionless to obtain fragmentation expectation parameters; making the actual time dimensionless and adding it to the expectation parameters to obtain reassembly resistance parameters; calculating the pre-allocated memory capacity based on the resistance parameters and the real-time bit rate; and performing physical isolation and decoding context preloading. This invention solves the problem of latency accumulation and screen stuttering caused by insufficient underlying memory reassembly and state machine switching during dynamic switching of multiple encoded video streams, achieving smooth transitions and low-latency decoding of video streams.

The application discloses a kind of dual-dimension cognitive function quantitative evaluation system and method based on multi-modal physiological signal.System includes brain electrical signal input module, force touch task presentation module, force feedback operation terminal, synchronous alignment unit, dual-dimension feature extraction module and cognitive evaluation engine.When user executes continuous tracking task, system uses operation failure timestamp as anchor point, jointly extracts ERP latency and frequency band power ratio, constructs attention stability index and resource consumption rate;In memory task, the intensity of encoding, extraction accuracy and scene consistency index are quantified.Cognitive evaluation engine integrates features into attention and memory comprehensive index, and maintains task in the most sensitive cognitive interval through adaptive difficulty adjustment mechanism.The present application embeds neurophysiological evaluation in real operation scene, significantly improves the ecological validity, sensitivity and individual adaptability of evaluation.

Various embodiments describe an apparatus, a method, and a computer program product. An example apparatus includes at least one processor; and at least one memory storing instructions that, when executed by the at least one processor, cause the apparatus at least to perform: encoding an input picture by using a first encoder or first encoding parameters; encoding the input picture by using a second encoder or second encoding parameters; generating a first reconstructed picture based on the encoding of the input picture by using the first encoder or the first encoding parameters; and generating a second reconstructed picture based on the encoding of the input picture by using the second encoder or the second encoding parameters.

The application discloses a large language model weight compression method, an inference operation method, a system and a medium, and belongs to the technical field of large language model weight compression and inference operation optimization. The method comprises the following steps: acquiring a weight matrix of a large language model and constructing an exponential high-frequency window and a benchmark value, classifying and encoding each weight data into a fixed-length code word and splitting the fixed-length code word into an independent bitmap, synchronously generating a compact value stream and a rollback value stream, and packing the compact value stream and the rollback value stream into weight compression data; loading the weight compression data, generating a state mask and a channel prefix mask in a register based on the bitmap, calculating a reading offset by using a population count instruction, reading a compressed data segment from a corresponding data stream, reconstructing weight data in the register in combination with the code word, and finally directly inputting the weight data into a matrix multiplication unit for operation; therefore, by implementing the application, the problem of efficiency reduction caused by control flow divergence and redundant memory access in GPU inference of existing variable-length encoding can be solved, and the model inference efficiency under lossless compression can be improved.

Disclosed in the present invention are a self-supervised image denoising method and system. The method comprises: by means of an input interface, acquiring a noise image and transmitting same to a memory; configuring, in the memory, a search space based on a U-Net framework; a graphics processing unit reading configuration information, executing coarse-grained population initialization on the basis of a decimal modular encoding strategy, and writing population data into the memory as a contiguous memory block; the graphics processing unit executing distance-guided parent selection and modular crossover and mutation operations in parallel; decoding offspring individuals into network structures, then executing self-supervised denoising processing in parallel, and calculating PSNR values; and a central processing unit executing environment selection and controlling an iteration process, finally selecting an optimal network structure, and outputting a denoised image. By means of the optimization of search space design and the collaboration of heterogeneous computing architectures, the present invention simultaneously realizes high-quality image detail restoration and efficient denoising processing, without requiring paired data.

Systems and methods for end-to-end pretraining of multimodal retrieval-augmented visual language models. In some examples, multimodal information may be encoded into key-value pairs and stored in a unified memory, which the model's retriever can access via multimodal query encodings in order to identify relevant information within multiple knowledge sources. The model may include an attentive fusion layer so that automatically-generated retrieval scores for multiple simultaneously-considered documents may b used in calculating attention scores, and gradients from the final task may be used to train the entire model (including the retriever) end-to-end and update the unified memory. In such cases, the retriever may thus be trained with the rest of the model without the need for ground-truth scores indicating which knowledge entries are most helpful in answering a given query, and the model's parameters may thus be focused on understanding queries and conducting reasoning rather than simply memorizing the training data.

The application relates to an OpenGL multi-class object uniform drawing method based on VBO internal coding, which comprises the following steps: constructing a mixed data structure containing vertex coordinates, color components, texture coordinates and object type marks, and storing the mixed data structure in the memory of a VBO in an interlaced arrangement; defining a uniform vertex shader and a fragment shader; judging whether texture mapping is needed according to the object type marks during a rendering process; and deleting the VBO object and the shader program after the rendering is completed, and releasing the display memory and the memory resources. The application links map elements with different rendering logics, realizes the uniform drawing of heterogeneous objects under a single pipeline, fully utilizes the parallel processing capacity of a GPU, avoids frequent switching of the shader, effectively saves the communication bandwidth between a CPU and the GPU, simplifies the development difficulty of multi-map element mixed rendering in a complex scene, and facilitates the horizontal expansion of a system in large-scale geographic information visualization.

A memory controller for a (dynamic random-access memory) DRAM die having m DRAM banks. The memory controller includes: an ECC engine configured to provide error correction for the plurality of DRAM banks, wherein the ECC engine includes: an encoding system configured to provide error correction coding (ECC) coding redundancy to a data component and generate an ECC codeword; and a partitioning system configured to partition the ECC code word into m segments for storage in the m DRAM banks, wherein a first subset of the m segments each include a portion of the data segment and a portion of the ECC coding redundancy, and a second subset of the m segments each include only ECC coding redundancy.

The present disclosure provides an abnormality detection method and device of traffic, a computer device and a storage medium, the method comprising: generating a dynamic heterogeneous graph of a target communication network; using a pre-trained encoding model to perform encoding processing on the dynamic heterogeneous graph to obtain node encoding features corresponding to each node in the dynamic heterogeneous graph; based on normal behavior prototypes of multiple normal communication behaviors learned by a neural memory network, performing addressing operation on the node encoding features corresponding to each node to obtain attention weights corresponding to each node and different normal behavior prototypes, respectively, and performing reconstruction processing on the node encoding features corresponding to each node according to the attention weights to obtain node reconstruction features corresponding to each node; and determining a traffic abnormality detection result corresponding to each node based on the difference information between the node reconstruction features and the node encoding features, and the attention dispersion degree corresponding to the node determined based on the attention weights.

The present invention provides an encoding device that can efficiently perform motion compensation using affine motion compensation. [Solution] The encoding device 100 comprises a circuit 160 and a memory 162. The circuit 160 uses the memory 162 to perform motion compensation on the target block in the affine motion compensation prediction process in the inter prediction process of the target block by limiting the range in which motion search or motion compensation is performed. In the affine motion compensation prediction process, the range in which motion search or motion compensation is performed is limited so that the variation between the motion vector of the control point at the upper left corner and the motion vector of the control point at the upper right corner of the target block in the affine motion compensation prediction process falls within a predetermined range. The variation is a value based on the difference between the motion vector of the control point at the upper left corner and the motion vector of the control point at the upper right corner of the target block.

A memory die executes enhanced foggy-fine program operations to recover foggy programmed data including a distribution that has extended into multiple states. The die includes blocks to store data in various formats. A memory controller on the die receives data from a storage device and generate parity data and enhanced foggy data based on the data received from the storage device. The memory controller stores the parity data and the enhanced foggy data to a cache prior to performing a foggy program operation to a block. The memory controller uses the parity data and the enhanced foggy data to recover the data when a foggy distribution associated with the foggy program operation crosses beyond an adjacent distribution. The memory controller then writes the recovered data to the block with a fine program operation.

The application discloses a kind of big model training data deduplication method based on semantic clustering.The method is first filtered by Hash matching and MinHash structure to completely repeated and locally repeated text;Subsequently, a pre-training semantic encoding model is used to generate a deep semantic vector, and a density clustering method is used to construct a semantic cluster, converting global high-complexity comparison into local retrieval within the cluster;Then, combined with the efficient neighbor search based on FAISS, the semantic similar samples within the cluster are grouped and deduplicated;Finally, an information entropy sensing mechanism is introduced, and representative samples are selected according to the content complexity of the samples, to significantly reduce redundancy while maintaining corpus diversity.The application can effectively reduce the risk of model memory caused by data duplication, improve training efficiency, and reduce data processing cost, suitable for various large-scale corpus construction and privacy-sensitive scenarios.

Embodiments of the present application provide an image encoding processing method, an image decoding processing method, an electronic device and a program product. The key image data is extracted from the video data, each key image data is processed according to the time sequence of the key image data by the recurrent memory network, and the partition feature map of each key image data is obtained, the partition feature map includes the region of interest and the non-region of interest; the position parameter of the region of interest is corrected based on the adaptive dynamic offset, and the focus feature map is obtained; the global attention weight corresponding to the region of interest and the local attention weight corresponding to the non-region of interest in the focus feature map are determined; the focus feature map is weighted calculated based on the global attention weight and the local attention weight, and the weighted feature map is obtained; the region of interest and the non-region of interest are encoded based on the weighted feature map, and the code stream of the key image data is obtained. According to the method provided by the embodiments of the present application, the efficiency and quality of video compression are improved.

A method performed by means of an apparatus, the apparatus including a memory comprising a plurality of storage units and a processor. The method includes: the processor acquiring, by using the plurality of storage units, a plurality of anomaly scores corresponding to a plurality of image regions segmented from an inspection image of a target object; the processor acquiring device anomaly values ​​by performing automatic encoding on the plurality of anomaly scores for each of a predetermined number of inspection images accumulated over a time series; and the processor detecting anomalies of the inspection device based on the acquired device anomaly values.

Methods and devices for processing multimodal data using a Large Multimedia Model. The method includes an encoding process to receive multimodal data, generate multimodal tokens, and store them in a cache; a prefill process to combine multimodal and text tokens, generate attention states and an initial output token, and store them in memory; and a decoding process to retrieve stored data, generate successive output tokens in an autoregressive manner, and produce an output response. A communication process facilitates independent operation of these processes by mapping memory blocks across graphics processing units, and thereby enables inter-process data transfer. The encoding, prefill, and decoding processes operate asynchronously, independently, and in parallel.

This invention discloses an intelligent generation method and system for power project material lists based on multimodal document parsing and knowledge graph semantic mapping, belonging to the fields of artificial intelligence and power engineering. It includes: a first stage of unified parsing and semantic encoding of text documents, engineering drawings, and tabular data to generate a semi-structured raw information set; a second stage of named entity recognition and type labeling of the raw information set based on a bidirectional long short-term memory network and a conditional random field hybrid model to generate a structured entity set; on this basis, a knowledge graph for the power material domain is constructed, a standard key point rule base is established, and semantic mapping and information completion are performed through a multi-level matching mechanism and knowledge reasoning; after quality control, a standardized material list is generated. This invention achieves intelligent and automated generation of power project material lists, improving the accuracy and efficiency of material list generation.

A method encodes compositional data structures by receiving component sparse distributed representation arrays having a predetermined array length and target sparsity level, applying position-specific permutation transformations to encode ordinal position information, combining position-encoded arrays through bitwise union to generate an intermediate array, and processing the intermediate array through a dual-phase sparsity reduction procedure comprising a coarse additive phase and a fine subtractive phase controlled by a sparsity overshoot threshold to generate a composite encoded array. The dual-phase procedure converges in substantially constant iterations for 4 or more components with final sparsity tightly controlled around the target. A decoding method applies inverse position-specific transformations to generate position-decoded arrays, computes overlap scores with candidate components through bit counting operations, and determines component identities based on threshold comparison. Scalable decoding may use triadic associative memory. Applications include searchable compression, privacy-preserving analytics, and efficient neural network embeddings.

Disclosed are apparatuses for controlling data input and output of a neural network circuit. The control apparatus includes a memory, an encoding circuit configured to receive a data sequence, generate a compressed data sequence in which consecutive invalid bits in a bit string of the data sequence are compressed into a single bit of the compressed data sequence, generate a validity determination sequence indicating valid bits and invalid bits in the bit string of the compressed data sequence, and write the compressed data sequence and the validity determination sequence to the memory, and a decoding circuit configured to read the compressed data sequence and the validity determination sequence from the memory, and determine bits in the bit string of the compressed data sequence that are set for transmission to the neural network circuit based on the validity determination sequence, such that the neural network circuit omits operations regarding non-consecutive invalid bits.

This invention belongs to the field of artificial intelligence and model compression technology, and discloses an additive quantization method for large language models. Without changing the additive quantization encoding format and codebook structure, a sampling strategy that jointly considers sub-vector energy density and spatial coverage is introduced in the residual K-means initialization stage. This allows high-energy-density regions to obtain more reasonable codeword coverage, significantly improving the codebook initialization quality. In the block-by-block end-to-end fine-tuning stage, a direction alignment loss based on negative log-cosine similarity is introduced in addition to the mean square error. This explicitly constrains the directional consistency between the quantized block output and the full-precision block output, making the direction constraint and amplitude constraint complementary. This method requires no quantization-aware training or large-scale fine-tuning, does not introduce additional trainable parameters, has low computational and storage overhead, is simple to implement in engineering, maintains near-full-precision inference performance under extremely low 2-bit quantization conditions, and significantly reduces memory usage and energy consumption during model deployment.

This invention discloses a multi-entity recognition method for Chinese resumes based on a BERT-BiLSTM-CRF joint model, belonging to the field of natural language processing and information extraction technology. The method performs deep semantic encoding on the original resume text using a pre-trained language model and generates offset mapping information. Then, it utilizes a bidirectional long short-term memory network to capture long-distance contextual dependencies in the text, enhancing sequence features. Next, it performs global decoding based on label transfer rules through a conditional random field layer to obtain the optimal word-level entity label sequence. Finally, based on the offset mapping information, the sequence is mapped and merged into a character-level entity recognition result. This invention effectively solves the problems of entity type ambiguity, insufficient capture of long-distance dependencies, and inaccurate entity boundary localization in Chinese resume parsing, and can accurately jointly identify multiple key entities in the three modules of educational background, work experience, and job expectations.

This application provides a method, device, and medium for accelerating a self-attention mechanism neural network, relating to the field of artificial intelligence model optimization technology. The method includes: constructing a Transformer network structure and initializing the embedding layer and positional encoding layer to obtain an initialized Transformer network; training and parameter fixing based on training data; and accelerating inference of the trained Transformer network based on channel importance parameters, sparse attention parameters, low-rank decomposition parameters, and projection matrix sharing relationships. This application addresses the technical problem in existing technologies where the high computational complexity and large memory consumption of Transformers lead to low inference efficiency and difficulty in real-time deployment in embedded environments, further affecting operational efficiency. By reconstructing the Transformer network, efficient deployment in embedded environments is achieved, improving operational efficiency.

Aspects of the present description refer to wireless communication systems configured to provide techniques for polar encoding of control information in conjunction with combined cyclic redundancy check (CRC) information. The combined CRC information may include a number of CRC bits selected to jointly decode and verify the control information to reduce CRC overhead.

The application discloses a kind of hardware perception type vector quantization calculation fusion acceleration instruction parallel processing system and method.For the memory bandwidth bottleneck problem of asymmetric distance calculation in large-scale vector retrieval, the application is inside processor register, 8-bit quantization encoding is loaded, zero extension to 32-bit floating point precision, affine correction based on quantization offset and scale factor, and distance accumulation operation for query vector are fused into single instruction multiple data stream driven in-register execution pipeline.The method eliminates the intermediate memory copy and register overflow operation in the traditional scheme, improves the throughput capacity of vector retrieval and processor resource utilization under the premise of maintaining the calculation accuracy.

The application discloses a variable working condition industrial process monitoring method and device based on abnormal sensitivity feature coding and knowledge distillation, and relates to the technical field of industrial process fault detection and intelligent monitoring. The method comprises the following steps: collecting variable working condition physical state parameters to obtain time series data, and decoupling the time series data through stationary subspace analysis to obtain stationary features; inputting the abnormal sensitivity long short-term memory network coding to extract abnormal sensitivity high-dimensional features; updating the dynamic memory bank based on the features, and utilizing the cross-attention mechanism to fuse the historical memory features to form cross-working condition fusion features; reconstructing the global time and feature dimension weight through the cross-attention time series distillation mechanism, aligning the current and historical fault sensitive features to obtain target features; inputting the target features into the fault classification head to map the positive abnormal probability, and determining the state and triggering the equipment control according to the positive abnormal probability. The application can accurately perceive the hidden subtle faults in the variable working condition non-stationary industrial process, and realize the continuous inheritance of monitoring knowledge among different production states.

Appropriate correction processing is performed on the predicted image. [Solution] The encoding device comprises a circuit and a memory. The circuit stores motion vector information and a BCW index in association with each other in a FIFO buffer (S2001). It registers one or more prediction candidates, including the combination of motion vector information and BCW index stored in the FIFO buffer, as prediction candidates in a prediction candidate list (S2002). It selects a prediction candidate from the prediction candidate list (S2003). Based on the BCW index of the selected prediction candidate, it performs BCW processing on the predicted image of the block to be processed (S2004). If no prediction candidate with the same motion vector information as the motion vector information stored in the FIFO buffer is registered in the prediction candidate list, the circuit registers the combination of motion vector information and BCW index stored in the FIFO buffer in the prediction candidate list.

This invention belongs to the field of graph representation learning and scientific and technological intelligence analysis, specifically involving a scientific and technological intelligence recommendation method based on temporally heterogeneous graphs. First, temporally context-aware encoding is performed on various types of entities and their relationships in complex temporally sequenced scientific and technological intelligence, such as authors, papers, institutions, and research topics, mapping time scalars to contextual features containing absolute evolutionary cycles and relative time decay. Second, a time-guided dynamic collaborative attention mechanism is used to adaptively allocate the fusion ratio of multiple relationships based on temporal signals. Next, an evolutionary state update module based on gated recurrent units is introduced to iteratively update local instantaneous features into global long-term memory. Then, combined with a temporally smoothing alignment mechanism, joint optimization is performed by penalizing the Euclidean distance between memory states of adjacent time steps. Finally, the learned node representations are used for scientific and technological intelligence association recommendation tasks such as potential collaboration discovery, related results recommendation, and topic association analysis.

The application discloses a video stream processing method and device, and a video stream generation method. The method comprises the following steps: receiving an input video stream, and generating a target feature vector according to multi-dimensional description information of the input video stream, wherein the multi-dimensional description information is used for describing video content recorded by the input video stream from different dimensions; searching for a candidate memory entry in a first memory bank according to the target feature vector; determining a reconstruction cost generated by a reconstruction operation of the candidate memory entry by using each encoding method; determining a target encoding method from multiple encoding methods according to multiple reconstruction costs corresponding to the multiple encoding methods, extracting to-be-encoded information corresponding to the target encoding method from the candidate memory entry and the input video stream, and executing the target encoding method on the to-be-encoded information to obtain an encoding result; and outputting the encoding result to a decoding end, wherein the encoding result is used for searching for a memory entry in the decoding end.