13291results about "Physical realisation" 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.

A computer system implements a unified framework integrating an adaptive elastic funnel (AEF) with a convergent intelligence fabric (CIF) for multi-agent AI collaboration. The system provides a universal multi-modal key-value subsystem for sharing partial computations, implements hybrid placement strategies for dynamic memory management, and incorporates quantum-resistant secure enclaves. The architecture integrates hardware acceleration through GPU-FPGA hybrid caching and neuromorphic processors, applies adaptive energy and thermal management across hardware generations, and implements autonomous flash resource orchestration with multi-dimensional wear management. The system orchestrates tensor workflows using hierarchical scheduling, enables cross-agent collaboration with privacy preservation, and supports continuous learning without catastrophic forgetting. This integration delivers unprecedented computational efficiency and security in high-dimensional decision-making environments while supporting incremental adoption through modular interfaces.

Systems and methods for adaptive edge computing using artificial intelligence (AI) include monitoring real-time accuracy of a neural network by using a feedback loop configured to detect changes in inference accuracy and dynamically adjusting the structure of the neural network by adding or removing hidden layers based on monitored error rates and predetermined computational constraints. A Kalman gain computation determines neural network weight adjustments based on monitored error rates. Weight matrices undergo incremental updates derived from these adjustments. Incremental weight adjustments remain stored in memory to enable low-bandwidth model updates. The neural network stores inference results and refined weights in an inference result database. Pre-trained models periodically receive incremental updates based on stored adjustments. Predictive holistic inference logic (PHIL) applied to stored inference results improves the accuracy of the inference results.

The invention discloses a large model reasoning efficiency dynamic optimization and hardware sensing compression method. The method comprises the following five steps: S1, generating an input complexity signal representing calculation complexity; s2, synchronously monitoring a hardware resource index of the operation platform, and generating a hardware state signal reflecting a real-time load; s3, inputting the input complexity signal and the hardware state signal into a dynamic strategy selector, and generating a compression control signal through a pre-trained decision model; s4, according to the compression control signal, dynamic reconfiguration operation is executed on the large model weight and the activation value of the current reasoning task; and S5, performing reasoning calculation by using the reconfigured large model, and feeding back a hardware resource index to the step S2 in real time in the calculation process to form a closed-loop optimization link. According to the large model reasoning efficiency dynamic optimization and hardware perception compression method, the problems of low resource utilization rate, delay fluctuation and energy efficiency imbalance caused by a static compression method in dynamic input and heterogeneous hardware environments can be solved.

The invention discloses a parallel strategy search method for efficient training of an artificial intelligence large model, and relates to the technical field of artificial intelligence, and the method comprises the steps: generating a three-dimensional hardware topology field through training a NeRF model; analyzing the calculation graph of the large model to be trained, mapping operator nodes into the three-dimensional hardware topology field, and generating a labeling calculation graph; constructing a mixed integer programming model based on the annotation calculation graph, and outputting a candidate strategy set; inputting the candidate strategy set into a double-layer strategy network, and outputting a parallel strategy set; loading the three-dimensional hardware topology field in the simulator, verifying and adjusting the parallel strategy set, and outputting an optimal strategy; and deploying the optimal strategy to the real cluster, and triggering a re-screening strategy. According to the method, a physical-level hardware communication field is modeled based on a neural radiation field, mixed integer programming and deep reinforcement learning are fused, parallel strategy global-local collaborative optimization is realized, and multi-target balance of training efficiency, video memory utilization rate and system robustness is realized.

The invention relates to the technical field of data processing, and discloses a heterogeneous computing acceleration method and system based on a deep learning framework network. The method comprises the following steps: acquiring performance parameters of heterogeneous computing equipment to obtain an equipment characteristic data set; receiving a calculation task and analyzing the calculation task into an operation sequence; the operation sequence is coded into a gene sequence, task decomposition is optimized through a genetic recombination algorithm, and a subtask set marked with acceleration characteristics is obtained; performing matching analysis on the sub-task set and the equipment characteristic data set to obtain a task allocation scheme; deploying the subtasks to corresponding equipment according to the allocation scheme to obtain a distributed execution framework; and monitoring the running state of the framework in real time, and dynamically adjusting resource allocation to obtain a calculation result of speed-up ratio improvement. According to the method, a self-adaptive task decomposition and resource allocation and dynamic load balancing mechanism can be realized, and the execution efficiency and the resource utilization rate of the deep learning task are improved.

A system for dynamic latent space adaptation using spatiotemporal kernel context for multiscale rendering with hierarchical and Lorentzian autoencoders. The Spatiotemporal Kernel Estimator (SKE) analyzes media through motion field, temporal recurrence, frequency band, and scene semantics analyzers to generate adaptive kernel parameters encoding content-specific importance distributions. The system dynamically adapts latent manifold geometry by modifying metric tensor properties according to kernel context, enabling content-aware compression that allocates representational capacity based on visual significance. A multiscale cache implements kernel-adaptive retention policies prioritizing important regions. An adaptive renderer provides intelligent level-of-detail selection based on zoom level and kernel-estimated importance, optimizing processing allocation. The self-optimizing architecture continuously refines kernel context and geometric adaptation based on user interaction and performance feedback, achieving superior compression ratios and perceptual quality. Applications include bandwidth-efficient video streaming, virtual reality, scientific visualization, and cognitive video analytics requiring intelligent context-aware visual processing.

Apparatuses, systems, and techniques are presented to predict gaze of an observer. In at least one embodiment, a network is trained to predict a gaze of one or more users based, at least in part, on one or more gazes corresponding to objects not always visible to the one or more users.

The invention relates to the technical field of low-power-consumption AI processors, and discloses a novel real-time data processing method and chip for calculating a low-power-consumption AI processor, and the method comprises the steps: carrying out the neural mimicry classification processing of a calculation task inputted by the low-power-consumption AI processor, and obtaining a calculation task feature matrix; performing multi-dimensional mapping operation on the calculation task feature matrix to obtain a heterogeneous calculation space vector; performing core resource allocation based on the heterogeneous calculation space vector to obtain a dual-mode processing unit configuration matrix; constructing a storage and calculation integrated data path based on the dual-mode processing unit configuration matrix to obtain a parallel calculation data flow diagram; executing dynamic voltage frequency mapping on the parallel computing data flow diagram to obtain a power consumption optimization control vector; and inputting the power consumption optimization control vector into the trained pulse neural network model to carry out computing power load balancing calculation and generate computing power load balancing parameters, and the method improves the real-time data processing capability of the low-power AI processor.

Generally, the present disclosure is directed to efficient neural network architectures, techniques for constructing new efficient neural networks, and approaches to low-latency execution of neural networks. In an example aspect, the present disclosure provides a more powerful configuration of an inverted bottleneck block that maintains an efficient execution profile. An example search system can parameterize a universal inverted bottleneck block with a first parameter that categorically activates a spatial mixing operation in the unexpanded state. In this manner, for instance, a large variety of different network architectures can be explored using a relatively compact search space that admits high levels of parameter sharing. Further, the present disclosure introduces an efficiency-optimized multi-query attention block.

An example method includes receiving input requests to process a plurality of input sequences using the machine-learned sequence processing model to generate a plurality of output sequences respectively corresponding to the plurality of input sequences; generating a plurality of initial attention tensors respectively for the plurality of input sequences, wherein: one or more respective initial attention tensors are generated for each respective input sequence in parallel over input elements of the respective input sequence; and the one or more respective initial attention tensors are generated in one or more batches having a first batch size using a prefill system that comprises one or more prefill computing devices and executes one or more layers of the machine-learned sequence processing model; and autoregressively generating, using the plurality of initial attention tensors, a plurality of output elements for each of the plurality of output sequences in one or more batches having a second batch size, wherein: the plurality of output elements are generated using a generation system that comprises one or more generation computing devices.

The invention relates to an urban road carbon emission prediction and optimization control method and system, and belongs to the technical field of road traffic carbon emission monitoring, and the method comprises the steps: obtaining real-time monitoring data of a to-be-detected road, and carrying out the preprocessing; performing space-time alignment on the preprocessed real-time monitoring data, extracting dynamic traffic features and static road features, and generating a space-time feature matrix; inputting the spatial-temporal characteristic matrix into a pre-trained carbon emission prediction model to obtain a prediction result matrix of the to-be-detected road; on the basis of the prediction result matrix, identifying a road grid of which the carbon emission intensity prediction value exceeds a preset threshold value as a high emission area; acquiring real-time traffic state data of the high emission area; and based on the real-time traffic state data and the prediction result matrix of the high emission area, constructing a multi-objective optimization function and carrying out solving to obtain an optimization control strategy of the high emission area. According to the invention, traffic carbon emission optimization control and real-time traffic condition synchronization can be realized.

The invention discloses a field specialized large model fine tuning and deployment method and system based on adaptive optimization, belongs to the technical field of artificial intelligence and machine learning, and aims to solve the technical problem of how to improve the rapid adaptation capability of a large model on small-scale field data. In order to meet the requirements of special fields of law, medical treatment, finance and the like on professional terms and complex contexts, the technical scheme adopted by the invention comprises the following steps of: data processing and sample generation: performing cleaning, feature extraction and small sample expansion on field data, and generating a high-quality training sample through a field feature guide mechanism; small-sample fine tuning and migration: realizing rapid field adaptation of a large model through a small-sample learning technology and dynamic Prompt optimization, and reducing dependence on large-scale annotation data in combination with cross-field migration learning; compressing and optimizing the model; performing automatic deployment and collaborative reasoning; and monitoring and adaptive optimization are carried out.

Various embodiments provide for electro-photonic networks, including a plurality of processing elements connected by bidirectional photonic channels, suited for implementing neural-network models. Weights of the model may be preloaded into memory of the processing elements based on assignments of neural nodes to processing elements implementing them, and routers of the processing elements can be configured to stream activations between the processing elements based on a predetermined flow of activations in the model.

A parallel processing unit comprises a plurality of processors each being coupled to a memory access hardware circuitry. Each memory access hardware circuitry is configured to receive, from the coupled processor, a memory access request specifying a coordinate of a multidimensional data structure, wherein the memory access hardware circuit is one of a plurality of memory access circuitry each coupled to a respective one of the processors; and, in response to the memory access request, translate the coordinate of the multidimensional data structure into plural memory addresses for the multidimensional data structure and using the plural memory addresses, asynchronously transfer at least a portion of the multidimensional data structure for processing by at least the coupled processor. The memory locations may be in the shared memory of the coupled processor and / or an external memory.

A key-value cache management system processes inference requests of a generative model. An inference request requests a starting virtual memory address and a number of layers in the generative model. In response to receiving an inference request, contiguous virtual memory space is reserved for the key-value cache in accordance with the inference request by assigning the starting virtual memory address in the key-value cache and calculating memory pointers for each layer in the model and each block so that the one or more blocks are written sequentially. The generated memory pointers are outputted to the generative model so that each self-attention layer writes computed key-value pairs at physical addresses in the key-value cache specified by the memory pointers.

The invention discloses a monocular vision inertial positioning method based on fine pre-integration and adaptive vision inertial weight, which comprises the following steps of: firstly, constructing key frames by using parallax by adopting a sliding window method, and processing the output of an inertial measurement unit between adjacent key frames by using fine pre-integration; initializing a visual inertial odometer to obtain external parameters and internal parameters of an initial sensor and scale and speed information of the odometer; normalizing the acceleration and the angular velocity of each pre-integration quantity by adopting an L2 norm, and weighting the components of the covariance matrix constrained by each fine pre-integration by taking the normalized acceleration and angular velocity as adaptive inertia optimization weights; based on the quality of feature matching after optical flow tracking, a visual weight is obtained according to the number of times of feature continuous tracking and changes of gray values before and after the feature continuous tracking, and the size of an information matrix of a visual re-projection residual error is adjusted in a self-adaptive mode; and finally, performing optimization iteration on the camera pose by adopting a sliding window and a factor graph to minimize a residual term based on mahalanobis distance measurement and obtain a final positioning result.

Techniques for optimizing and deploying deep neural network (CNN) machine learning models for inference using static analysis are described. A method includes obtaining a deep neural network (DNN) machine learning (ML) model, generating an intermediate representation for the ML model, the intermediate representation including one or more nodes corresponding to one or more operators utilized by the ML model, identifying, for at least one node of the intermediate representation, an optimized schedule for at least one operator corresponding to the at least one node using a static analysis that is based on a hardware-specific cost model, generating an optimized intermediate representation using the optimized schedule that is optimized for execution on a hardware platform, and generating code corresponding to the ML model based at least in part on the optimized intermediate representation, wherein the code is specific to the hardware platform.

The invention discloses a bionic moving target sensing, storage and calculation integrated photoelectric detection system based on continuous photoconduction, and relates to the field of bionic intelligent vision and the field of brain-like calculation, and the system comprises a synaptic bionic photoelectric detector array and a brain-like memristor array which are interconnected. When the synaptic-like bionic photoelectric detector array detects visual information of a moving target, a fixed bias voltage is applied to a device in each pixel, space illumination information corresponding to each pixel is converted into a continuously attenuated photocurrent signal, illumination information of a multi-frame dynamic target is continuously received in the continuous attenuation process of the photocurrent, and the visual information of the moving target is detected by the synaptic-like bionic photoelectric detector array. After the last frame of illumination information is input, a light current signal of each pixel is read, and a feature map memorizing all historical moment information is determined; the brain-like memristor array is based on the Ohm's law and the Kirchhoff's law, performs deep neural network operation on the feature map, outputs a current corresponding to a task result label, identifies a dynamic target, and can meet the requirements of a deep neural network and a complex dynamic visual task.

The invention provides an end-side multi-modal large model accelerated reasoning method and system, and the method comprises the steps: carrying out the two-stage screening and rearrangement of visual tokens based on the CLS attention and text-to-visual attention in a visual encoder and pre-filling stage, and constructing a sparse attention and sparse key value cache; in a decoding stage, an important neuron set is judged according to activation gating or historical statistics, only a corresponding feedforward network weight is pulled and calculated, missed weights are loaded on demand through asynchronous I / O, and hot neurons are maintained in a high-speed memory to utilize model sparsity, so that video memory / memory occupancy and calculation overhead are remarkably reduced on an end side; throughput and time delay performance are improved. According to the method, the internal memory and computing resources required by reasoning of the multi-modal large language model are reduced from two dimensions by utilizing the endogenous sparsity of the end-side large language model in input and the model, so that a higher reasoning speed is achieved by utilizing fewer resources on the premise of keeping the size of the model unchanged, and the performance of the whole system is improved.

Apparatuses, systems, and techniques to perform a neural network. In at least one embodiment, an application programming interface schedules two or more graph nodes to be performed by two or more parallel processing pipelines based, at least in part, on an order of layers in the neural network.

Provided in the present invention is a parameter-efficient large-language fine-tuning federated learning framework, comprising the following steps: performing modeling on LoRA adapters of different edge clouds; since different weights exhibit different average performances on the LoRA adapters, using singular values to quantify the importance of the weights, and therefore, before each round of independent training of the LoRA adapters using N edge clouds, using a matrix singular value to decompose a BA matrix in the LoRA adapter for each trainable weight; configuring heterogeneous LoRA adapters on the basis of the importance of the weights; and using different numbers of quantization bits to quantize a pre-trained model, and performing high-precision inverse quantization on the pre-trained model only when matrix multiplication is executed, wherein the pre-trained model is quantized to the maximum number of quantization bits on the basis of the memory budget of the edge clouds. The present invention has the following beneficial effects: the present invention determines the optimal fine-tuning model structure, thereby improving the performance of LLM fine-tuning, and adapts to heterogeneous and resource-constrained edge clouds.

The invention discloses a dynamic optimization updating method of a credible data intelligent detection system, which relates to the technical field of credible data detection, and comprises the following steps: deploying heterogeneous feature acquisition nodes, acquiring data association features, operating behavior fingerprints and setting a dynamic credible evaluation threshold value; constructing a gradient aggregation channel based on differential privacy encryption, performing security fusion, generating a global model optimization vector, and generating an adversarial sample set; carrying out weight distribution, configuring a dynamic update step length, and generating an update instruction set; and activating the adversarial sample generation network, generating an incremental update package, triggering asynchronous hot deployment of the edge computing nodes, reconstructing a model topological structure of the federated incremental learning framework, and generating an update log. The technical problems of poor adaptability and low real-time performance in the aspects of data credibility evaluation, privacy protection and security defense in the prior art are solved, and the technical effects of improving the data credibility, enhancing the privacy protection capability and improving the attack defense capability are achieved.

The invention provides a power grid dispatching strategy optimization method and system, and the method comprises the steps: deploying a device based on a quantum entanglement technology between transformer substations, and collecting the state information of a power grid, and encrypting and transmitting the state information to a control center through a quantum network; a pulse neural network processor is used in a control center to extract spatial-temporal characteristics, and a quantum game theory model is used to generate an optimized scheduling strategy. Multi-modal verification data is collected, including visual deformation, abnormal sound monitoring, and operational resistance data. A genetic algorithm and deep reinforcement learning are utilized to optimize a decision model, and a dynamic weight distribution mechanism is included. And through a photon-quantum hybrid computing architecture execution model, a scheduling strategy is collaboratively optimized, and a result is fed back to a physical power grid. According to the method, the quantum technology and the spiking neural network are combined, new energy fluctuation is captured in real time, the multi-target weight is dynamically adjusted through the quantum game theory model, and the scheduling strategy robustness is improved. And a photon-quantum hybrid computing architecture is adopted, so that the feature extraction speed and the optimization efficiency are improved.

AI neural networks are integrated with distributed ledger technology (DLT), including the integration of decentralized oracles, to enhance the functionality, security, and efficiency of systems. The integration incorporates several components and methods comprising novel neural network architectures, data preprocessing techniques, DLT integration approaches, privacy and security measures, consensus mechanisms, and others. The invention enables the development of neural smart contracts, neural consensus algorithms, dynamic neural network training, neural network optimization techniques, hybrid DLT architectures, neural network interpretability methods, and reinforcement learning for neural smart contract execution. By integrating decentralized oracles into the system, the invention has the capability to further enhance the reliability, accuracy, and diversity of data used within the DLT, improving transaction validation, consensus mechanisms, and overall system performance.

A neural network model training method includes: constructing a first neural network architecture, where the first neural network architecture includes M basic unit layers, each of the M basic unit layers includes a plurality of basic units, and the plurality of basic units includes at least a first-type basic unit and a second-type basic unit; and obtaining a target model through training based on datasets respectively corresponding to a plurality of tasks and the first neural network architecture, where the target model includes a plurality of task paths, at least some of the plurality of task paths include N basic units selected from some of the M basic unit layers, and N<M.

A new multi-edge smart community system consisting of communication, computing, and energy harvesting models, where the task execution delay and energy consumption are formalized as the optimization objectives under multiple constraints. For single-edge scenarios, we propose an improved twin-delayed DRL-based algorithm. For multi-edge scenarios, we develop a novel personalized FL-based training framework for DRL. Using the real-world settings and testbed, extensive experiments are conducted to validate the effectiveness of the proposed PFR-OA. The results show that the PFR-OA achieves better trade-offs between delay and energy consumption and exhibits higher task execution success rates than benchmark methods under different scenarios. Notably, the PFR-OA reaches a faster convergence speed compared to advanced DRL-based and FRL-based methods. Moreover, we further verify the practicality and superiority of the PFR-OA via real-world testbed experiments.

A resource set which includes multiple servers with a respective plurality of training computing devices is identified for training a machine learning model. The resource set is subdivided into partition groups, such that each partition group can store a respective replica of state information of the model. The model is trained using the partition groups. The training comprises a multi-stage gathering of a portion of the state information at training computing devices of a particular partition group. Different types of communication channels between training computing devices are used in respective stages of the gathering, including inter-server communication channels in one stage and an intra-server communication channel during another stage. A trained version of the model is stored.

An electronic circuit system implementing and executing machine learning inference engines. While ML inference engines are based on (architectures and parameters defined by) configured, trained and tuned machine learning models, our design has the novel ability to support data driven, on-the-fly-reconfigured model runs. Reconfiguration and tuning operations include dynamic computational graph modifications, define-by-run alterations, changes to network depth (number of layers) and width (neurons per layer), and adjustments to weights, biases, plus activation function parameters. Neural networks supported include Feed-Forward, RNN, CNN, and Hopfield architectures, plus Ensemble, Federated, Cooperating, Adversarial, and Swarm collections. Decision Trees and Forests are also supported, as are more esoteric approaches such as ART and KAN. Our invention is capable of running both standalone and cooperatively, the cooperative processing being local and / or remote / cloud based, interfacing with telemetry applications to feed data, and machine learning software to feed new or updated models.

The embodiment of the invention discloses an in-memory computing circuit based on a magnetic cache, and the circuit comprises at least one magnetic cache unit, at least one in-memory computing unit, and a timer. The magnetic cache unit in the at least one magnetic cache unit is used for caching data output by the corresponding in-memory computing unit as to-be-processed data within the corresponding data retention time; the timer is used for respectively setting data retention time for the at least one magnetic cache unit; and the in-memory computing unit in the at least one in-memory computing unit is used for extracting the data to be processed from the corresponding magnetic cache unit for calculation and outputting the computed data to other magnetic cache units. According to the embodiment of the invention, the invention achieves the flexible adjustment of the data retention time of the magnetic cache unit in various in-memory calculation scenes, and achieves the provision of a high-capacity cache for the data needed by in-memory computing under the lower power consumption.