95 results about "Backward propagation" patented technology

The invention belongs to the field of computer vision and cross-modal retrieval, and particularly relates to an image-text pedestrian retrieval method based on image block replacement and cross-modal identity alignment, which comprises the following steps: acquiring a public data set of image and text description, and constructing a pedestrian re-recognition model PRCIA; inputting the data set into a PRCIA model for training and verification, and performing iterative updating on a training weight file through forward and backward propagation to obtain a trained PRCIA model; constructing a reasoning stage model, reserving double encoders and fusing global features in a reasoning stage, and ensuring the calculation efficiency; and inputting the test set into the reasoning stage model to obtain a detection result, thereby realizing text-based pedestrian detection. According to the method, fine-grained association between the image blocks and the text phrases is established through the PR module, cross-modal identity feature expression is enhanced through the CIA module, the accuracy and robustness of text-to-image pedestrian retrieval are remarkably improved, and the requirements of actual scenes are more easily met.

The invention discloses a distributed AI training-oriented RDMA transmission mode intelligent selection method, and the method comprises the steps: obtaining multi-dimensional data features in a deep learning training process, the multi-dimensional data features comprising a tensor structure feature, a data access mode feature, a data timeliness feature, a data dependency feature and a data change rate feature; identifying a current training stage, wherein the training stage comprises a forward propagation stage, a back propagation stage and a parameter updating stage; based on the multi-dimensional data features and the training stage, an optimal RDMA transmission mode matched with the current training scene is selected through an adaptive decision matrix, and the optimal RDMA transmission mode comprises one or more combinations of RDMA Write operation, RDMA Read operation, Send / Receive operation and Atomic operation; and switching the RDMA transmission mode in real time according to the dynamic change of a network environment, ensuring that a training data flow is not interrupted in a switching process by adopting a progressive flow migration strategy, and optimizing a subsequent mode selection decision through a historical performance learning mechanism.

The invention discloses a bidirectional adaptive video super-resolution method based on frame difficulty index, and belongs to the field of computer vision. The invention provides a frame-level dynamic reconstruction method for solving the problems that simple frame calculation is redundant and difficult frame reconstruction is insufficient due to the fact that an existing model adopts a fixed calculation strategy for video frames with different difficulties. According to the method, a motion detail decoupling propagation network is constructed, motion information is efficiently transmitted by utilizing a shallow forward propagation branch, and a deep backward propagation branch focuses on recovering texture details so as to decouple a time sequence propagation task; meanwhile, a frame reconstruction difficulty evaluation network is introduced to generate a global difficulty index, so that the receptive field weight of the adaptive time sequence fusion network and the refining depth of the dynamic refining network are regulated and controlled. According to the method, through explicit modeling frame-level reconstruction difficulty, adaptive matching of the model capacity and the video frame feature complexity is realized, and the video reconstruction performance is remarkably improved under limited computing power.

Method comprising: monitoring whether a MTLF receives a first state of an environment on which a RL training is to be performed; performing a ML model forward propagation on a first model of the environment having the first state for each of plural actions to obtain a respective expected reward for each of the plural actions: informing a service consumer on the plural actions and their respective expected reward; supervising whether the MTLF receives a RL training result information after the informing the service consumer on the plural actions, wherein the RL training result information comprises an indication of one of the plural actions, a second state of the environment, and a reward feedback; conducting a ML model backward propagation on the first model of the environment having the second state for the one of the plural actions using the reward feedback to obtain a second model of the environment.

This invention relates to a closed-loop data system and method for embodied intelligent robots. The system includes: a data acquisition module for acquiring multimodal operational data of the embodied intelligent robot during task execution and constructing multiple embodied state nodes based on the multimodal operational data; a construction module for constructing event correlation relationships between embodied state nodes based on temporal and physical constraints, forming an embodied state event graph; an anomaly identification module for identifying anomalous embodied state nodes in the embodied state event graph and marking them as failure source nodes; and a propagation analysis module for determining failure propagation paths forward or backward from the failure source nodes based on event correlation relationships. Using the above scheme, the operational data during the task execution of the embodied intelligent robot can be effectively analyzed, anomalous states can be accurately identified, and data-driven closed-loop optimization can be achieved.

In some implementations, there is provided receiving a query request including a top k query operator for query plan generation, optimization, and execution; generating a query plan that includes at least one pipeline of a plurality of operators, wherein the at least one pipeline is associated with a directed acyclic graph; detecting in the query plan a first operator in the at least one pipeline that causes an early exit; and in response to the early exit by the first operator during query execution, processing back through the directed acyclic graph to identify at least one preceding operator that should or should not run given the early exit of the first operator.

A compute cluster includes a compute mesh of interconnected compute elements. A clock mesh includes reconvergent elements, and the reconvergent clock mesh distributes a clock signal to the compute elements. The clock mesh has a topology of nodes connected by branches. Repeaters are located on the branches between nodes. As a result, the reconvergent clock mesh is unidirectional. The repeaters limit the clock signal to forward propagation through the clock mesh and prevent backward propagation through the clock mesh. The clock mesh is reconvergent, in that the clock signal may branch out along different paths from one node and then these paths reconverge at a later node.

A specialized integrated circuit for artificial neural networks is integrated with high bandwidth memory. The neural network includes a systolic array of interconnected processing elements, including upstream processing elements and downstream processing elements. Each processing element includes a pair of input / output ports for concurrent forward and backward propagation. The processing elements can be used for convolution, in which case the pair of input / output ports can support fast and efficient scanning of a kernel over activations.

The application discloses a medical image class incremental learning method and system based on feature principal direction guided projection, and belongs to the technical field of medical image analysis; the method acquires medical image features through a pre-trained feature extractor; random projection is performed on the features to obtain first-view features; a covariance matrix is calculated based on a feature cache area and is decomposed to extract the first K principal directions, construct a guided projection matrix, and obtain second-view features; the double-view features are spliced to obtain final representation; the classifier weight is directly calculated through a ridge regression closed-form solution without backward propagation training; the application adaptively captures the core structure of the feature space through data-driven principal direction projection, retains the detailed information in combination with random projection, effectively improves the feature separability of fine-grained classes, adopts an analytical classifier updating mechanism, realizes stable class incremental learning without storing historical data, improves the model updating efficiency, and is suitable for efficient continuous learning tasks in various medical image scenes.

The invention relates to a potential function model construction method, system and device and a storage medium, and belongs to the technical field of digital data processing. The construction method of the potential function model comprises the following steps: determining a neighbor atom set for each atom in an atom system according to a truncation radius; constructing a neighbor environment descriptor of each atom based on the neighbor atom set of each atom; inputting the neighbor environment descriptor of each atom into an independently configured neural network corresponding to the element type of the neighbor environment descriptor, and performing forward propagation calculation from an input layer to an output layer of the neural network to obtain energy of each atom; performing back propagation calculation and descriptor back calculation on the energy of each atom from the output layer to the input layer through each independently configured neural network to obtain partial derivative of the energy of each atom to the position of the adjacent atom; and obtaining the stress of each atom based on the partial derivative of the energy of each atom to the position of the adjacent atom. The potential function model has both precision and calculation efficiency.

The invention belongs to the technical field of spiking neural networks. The invention provides a construction method of a reversible pulse neural network supporting parallel computing. According to the embodiment of the invention, the reversible calculation structure is introduced, the neuron state of an intermediate layer does not need to be explicitly stored in the forward propagation process of the network, and the reconstruction of the activation value can be realized through the reversible mapping relation in the backward propagation stage. In the design stage of the reversible structure, hardware mapping characteristics are fully considered, and forward calculation and reverse reconstruction processes of reversible mapping are simultaneously incorporated into a parallel execution framework of a configurable calculation array. Through collaborative design of a data path and control logic, forward propagation and reverse reversible reconstruction can be efficiently and parallelly operated in the same array, and inter-stage calculation idle time and data waiting time are remarkably shortened. On the premise of ensuring reversibility, streamlined and parallel execution of forward and reverse calculation is realized, and the throughput rate and the training speed of the system are integrally improved.

Computer systems and computer-implemented methods modify a machine learning network, such as a deep neural network, to introduce judgment to the network. A “combining” node is added to the network, to thereby generate a modified network, where activation of the combining node is based, at least in part, on output from a subject node of the network. The computer system then trains the modified network by, for each training data item in a set of training data, performing forward and back propagation computations through the modified network, where the backward propagation computation through the modified network comprises computing estimated partial derivatives of an error function of an objective for the network, except that the combining node selectively blocks back-propagation of estimated partial derivatives to the subject node, even though activation of the combining node is based on the activation of the subject node.

This invention belongs to the field of bank erosion control, specifically a machine vision-based bank erosion control monitoring system and equipment. The monitoring system includes the following steps: Step 1: Building an outdoor intelligent data acquisition platform, its network equipment, and a GPS positioning module; remotely controlling the monitoring equipment to monitor the entire construction process, collecting construction videos and location information, and uploading them to a data center; Step 2: Preprocessing the construction videos, constructing a construction material dataset, inputting the dataset into a YOLOv5-based neural network for training, and optimizing the cost function through forward and backward propagation algorithms; Step 3: Using the trained neural network, identifying important process events, recording key frames of important events occurring at key embankment sections using time as an index, and saving them to the data center; Step 4: Saving the construction videos and key frames collected in Steps 1 and 3 in the database, reading key frames and related information from the data center, and displaying them in the form of charts and reports.

The application discloses an online task screening-based meta-learning training method and system for small sample image classification model training. The existing task-level self-paced meta-learning training adopts a double-pass process, which needs to traverse the complete training round to count all task losses and then replay the training, causing the problems of repeated traversal and high training time consumption. The method of the application comprises the following steps: sampling a main task and performing one gradient-free detection forward propagation to obtain a detection loss; maintaining a sliding loss buffer and an exponential moving average value through an online loss tracker; determining a target retention rate and constructing a dynamic threshold through a course scheduler; comparing the detection loss with the threshold to decide whether to retain or skip; and performing only one backward propagation and optimizer update for the retained task. The double-pass process is transformed into online single-pass screening, the repeated traversal and replay overhead in the same training round is reduced, the additional training time consumption is reduced, and the additional propagation and parameter update overhead of the old multi-step update is reduced in consistency stabilization.

The application provides a keyword reverse propagation algorithm based on a citation network structure, and comprises the following steps: a spring charge model is established, force-directed layout processing is performed, and a force-directed layout graph is established; a keyword propagation model is established by using a reverse propagation algorithm, and a keyword weight change contrast curve is obtained; a citation network model is constructed by using the force-directed layout graph; iterative calculation is performed on the force-directed layout graph until the energy state in the force-directed layout graph reaches a minimum value; while the citation network is iteratively calculated, the keyword weight in the citation network model is adjusted, and a converged citation network layout graph is calculated. The application has the beneficial effects that the entire network is taken as a main body, keywords owned by cited documents are selected at a certain probability and are propagated backward along the network to the citing documents, the weight of the keywords is increased while the text clustering idea is retained, and clear visual data effects are obtained through force-directed layout.

The application discloses a kind of based on adaptive flow parallel segmented federal learning method and device. The method comprises: for each device: the device divides bottom submodel into multiple segmented models, sequentially respectively to each segmented model is carried out forward propagation, and the intermediate activation data corresponding to the segmented model after forward propagation is uploaded to server;After server receives intermediate activation data, forward propagation and back propagation are carried out using top submodel, and the gradient of corresponding segmented model is obtained and is issued to corresponding device;For each device: after the gradient is received, the gradient is updated using its set weight, and the set weight is inversely proportional to the update frequency of the device in the current round, and the local bottom submodel is updated using the updated gradient;Forward and backward propagation, data upload, gradient issuing, model updating are carried out in the way of pipeline parallelly. The application accelerates the training of large model on resource-limited edge device and maintains accuracy.

A three-stage semi-supervised instance segmentation training method and a system. In a first stage, a teacher model and a student model are trained based on labeled data. In a second stage, the teacher model performs prediction on unlabeled data and generates pseudo labels according to a prediction result. The student model learns labeled data and unlabeled data based on the pseudo labels. A soft label filter is used to filter for obtaining high-quality pseudo labels, a positive sample loss function is used to eliminate a problem of incorrect model convergence due to the pseudo labels with incomplete information, and the parameters of the student model are updated via a backward propagation process. In a third stage, the parameters of the student model are transferred to the teacher model through an exponential moving average operation, so that the teacher model and the student model are under a same architecture.

This application discloses a multimodal model task execution method, visual reasoning method, apparatus, and device. The method includes: performing a visual reasoning task on first visual data and first text data using a first multimodal model to obtain a first hybrid feature representation; determining a text loss based on the discrete text feature representation; determining a visual loss based on the continuous visual feature representation; and training the first multimodal model based on the text loss and the visual loss to obtain a second multimodal model. Specifically, the forward inference process retains the continuous visual feature branch, eliminating the need to forcibly compress the visual signal into discrete text tokens, thus improving the prediction accuracy of the trained multimodal model in the visual reasoning task. The backward propagation process decouples the learning of the discrete text feature representation from the learning of the continuous visual feature representation, effectively stabilizing the update of the continuous visual features without hindering text convergence, thereby improving the training stability of the multimodal model.

The application provides a flow field reconstruction and deduction method based on K-A theorem, and belongs to the technical field of ship hydrodynamics. First, multi-dimensional data such as the direction, amplitude, period, frequency and flow velocity of irregular waves are collected and integrated into high-dimensional wave vectors. Then, the KAN network with B-spline curve as a learnable function is used to map the high-dimensional input into a low-dimensional parameter list, obtain the amplitude, angular frequency and initial phase of the dominant equivalent regular wave, calculate the decoupling target function by constructing an equivalent wave superposition model, compare the decoupling target function with the true value to obtain a loss function, and then optimize the network parameters by backward propagation until convergence. Finally, based on the network output parameters, a wave superposition model is constructed to represent the real wave in the form of a series of interpretable regular waves superimposed in time sequence, so that efficient and accurate flow field reconstruction and deduction are realized.

Embodiments of the present application provide a recalculation control method and device, a storage medium and a program product, relating to the technical field of artificial intelligence chips. The method comprises: obtaining a first data amount of activation values generated in a forward propagation process of a to-be-tested model, and obtaining an expected offloading data amount of the activation values that can be offloaded through a target bandwidth between a central processing unit and an artificial intelligence chip; when the first data amount is greater than the expected offloading data amount, it indicates that the target bandwidth cannot offload all the activation values to the memory, so a difference data amount between the first data amount and the expected offloading data amount is used to select part of candidate operators from a plurality of candidate operators included in the to-be-tested model to perform recalculation in a backward propagation process; and the other part of the candidate operators performs activation value offloading in the forward propagation process, which greatly reduces the memory requirement of model training, and avoids bandwidth shortage caused by activation value offloading, thereby improving the overall performance of the artificial intelligence chip during model training.

Apparatuses and methods for training a neural network accelerator using quantization precision data formats are disclosed, and in particular for storing activation values from a neural network in a compressed format having lossy or non-uniform mantissa for use during forward and backward propagation training of the neural network. In certain examples of the disclosed technology, a computing system includes a processor, a memory, and a compressor in communication with the memory. The computing system is configured to perform forward propagation for a layer of a neural network to produce first activation values in a first block floating point format. In some examples, the activation values generated by the forward propagation are converted by the compressor to a second block floating point format having a non-uniform and / or lossy mantissa. The compressed activation values are stored in the memory, where they can be retrieved for use during backward propagation.

The present disclosure provides a method for model training, a host, and a storage apparatus. The method may include instructing, by a host apparatus to a graphics processing unit (GPU), to load first intermediate data of a current layer of a model from a dynamic random access memory (DRAM) of a storage apparatus into a memory of the GPU, wherein the current layer is a layer of the model for which computation is being performed; and notifying, the host apparatus to the storage apparatus, to prefetch second intermediate data of a layer of the model to be computed from NAND of the storage apparatus into the DRAM of the storage apparatus based on a space capacity of the DRAM, wherein the first intermediate data of the current layer is used by the GPU in backward propagation of the model to perform the computation for the current layer.

A data processing method and device, a processor, an electronic device, and a storage medium. The data processing method comprises: performing a padding operation on an output gradient matrix loaded from a memory according to a padding mode of a forward propagation process to obtain a first gradient matrix and cache the first gradient matrix to a cache component of a processor, wherein the padding value filled in the padding operation is 0; reading the first gradient matrix from the cache component, and performing a backward propagation operation of average pooling on the first gradient matrix to obtain an input gradient matrix, wherein, during the backward propagation operation, a sliding window convolution operation is performed on the first gradient matrix using a first kernel to obtain the input gradient matrix, the shape size of the first kernel is the same as that of a convolution kernel of the forward propagation process, and both are RXS, the element value of each element in the first kernel is 1 / RXS, and R and S are positive integers. The data processing method uses calculation instead of the original judgment jump, reduces a large number of branch instructions, and greatly improves the processing performance.

The application discloses a model training method and device, a storage medium and a program product, and belongs to the technical field of data processing. The method comprises the following steps: performing multiple iteration training on a neural network model by using training samples. In one iteration training in the multiple iteration training, the following steps are performed: in a forward propagation process, sample data is processed according to a weight value in the neural network model and a current truncation value of a network layer, and output data is obtained; in a backward propagation process, the weight value in the neural network model is adjusted according to a loss value between the output data and sample labels, and the truncation value of the network layer is adjusted according to the loss value, the current truncation value of the network layer and an activation value. In the application, the truncation value in the neural network model is obtained through training, that is, the upper and lower limits for quantizing the activation value can be adaptively adjusted during the model training process, so that the quantization error is reduced, and the performance of the neural network model is improved.

This invention discloses a hybrid distributed dense-state fine-tuning method and system for models based on a trusted data space. The method includes: acquiring plaintext data to be trained; splitting a large model into a first model and a second model, and deploying them respectively on a first computing node group and a second computing node group; the first computing node group performing forward computation of the first model and dense-state processing of the results based on the plaintext data to be trained, and transmitting the dense-state processed data to the second computing node group; decrypting the dense-state processed data, and the second computing node group performing forward computation and backward propagation of the second model based on the decrypted data to obtain feature map gradients, performing dense-state processing, and sending the gradients to a trusted execution environment; decrypting the dense-state gradient data and performing backward propagation of the first model based on it, and obtaining the updated parameters of the first model based on the propagation results. This provides an efficient computational solution for dense-state fine-tuning of large models with hundreds of billions of parameters in dense-state scenarios.

The application discloses a method for training a neural network in a longitudinal federated scenario based on multi-bias interaction, an electronic device and a storage medium, and belongs to the technical field of privacy computing. In order to improve the efficiency under the premise of safe transmission, the model structure of each training party is a full connection layer and a Dropout layer, including training of data of a training initiator and data of a training assistant by a forward propagation method to obtain a training result of the forward propagation method; the training result of the forward propagation method is trained by a backward propagation method, model parameters of the training initiator and the training assistant are updated respectively, one round of training is completed, and the next round is entered until the training result reaches a precision requirement or a stop condition, and the neural network training in the longitudinal federated scenario based on the multi-bias interaction is completed. The application increases information interaction between training parties, and the filtering function of features is guided by multi-party data information, so that the result is more convincing.

The invention discloses a particle swarm neural network PID parameter optimization system based on an FPGA, and relates to the technical field of automatic control, the system comprises a particle swarm optimization module, a BP neural network module, a PID control module and an FPGA processing module, the particle swarm optimization module carries out global optimization on an initial weight and a threshold value of a BP neural network, and the PID control module carries out PID parameter optimization on the initial weight and the threshold value of the BP neural network; the BP neural network module receives optimized network parameters as initial parameters, performs forward propagation calculation and back propagation training according to control error signals and outputs PID control parameters, the PID control module calculates control output quantity according to system errors, and the FPGA processing module executes operation tasks of all the modules in a hardware parallel mode. According to the method, the particle swarm optimization is combined with the BP neural network, so that the global search capability of the PID control algorithm is enhanced, the problem that the BP neural network falls into a local minimum value is avoided, and the calculation speed and the real-time performance of the system are improved.

Provided in the embodiments of the present application are a formation imaging system, a formation imaging method, a device and a medium. The system performs formation imaging on the basis of a reverse time migration (RTM) formation imaging algorithm. The system comprises: at least one group of nodes, and a CXL 2.0 memory pool in communication with the at least one group of nodes, wherein the at least one group of nodes are used for inputting computation process data into the CXL 2.0 memory pool in a forward propagation simulation stage of the RTM formation imaging algorithm, and reading the computation process data from the CXL 2.0 memory pool in a backward propagation simulation stage of the RTM formation imaging algorithm; and the CXL 2.0 memory pool is used for dynamically allocating memory to the at least one group of nodes. By means of the embodiments of the present application, the dynamic requirements of the RTM formation imaging algorithm in a cluster for memory can be met, and the effective management of a public memory pool in the cluster is realized, thereby reducing the consumption of hardware resources.

The application discloses a kind of neural network monitoring method, equipment and medium based on functional topology grouping, belong to neural network technical field, including after the backward propagation execution of standard training cycle, before parameter update execution inserts monitoring collection point, according to the tensor operation type indicated by parameter name each trainable parameter gradient vector is automatically classified into multiple function isomorphic groups etc.Step.The present application is classified into function isomorphic group by function topological relationship to the trainable parameter of neural network, and in group, multiple mathematical orthogonal gradient dynamics signals are synchronously collected, the scope error caused by the mixed heterogeneous parameter calculation statistics of existing gradient analysis in full parameter level is solved, and the problem that training monitoring only relies on single loss function cannot distinguish different abnormal types such as gradient explosion, complete learning, training stuck, precise diagnosis and abnormal early warning of multidimensional health state in training process are realized.

The application relates to video text cross-modal retrieval and relates to a video text cross-modal retrieval method based on pre-clustering guidance. When an unsupervised video text cross-modal retrieval model is trained on an unlabeled data set, the problem of weak supervision information caused by unreasonable sample pair construction exists. 1) Extracting video and text original features of the training data set; 2) splicing time-aligned normalized original features as unified features, setting a group of threshold values and performing DBSCAN clustering multiple times to generate multiple groups of clustering clusters; 3) calculating the mean value of the video and text features of each clustering cluster as the clustering center; 4) sampling a batch of training samples, for each group of clustering centers, finding the nearest center to construct a pseudo label for the video and text, and constructing a video text positive sample pair; 5) projecting the original features into a common space, calculating the loss of the positive sample corresponding to each group of centers, and updating the projection model through backward propagation; 6) repeating steps 4) to 5) until a predetermined number of training rounds is reached.