21 results about "Stochastic gradient descent" patented technology

A computer-implemented method for training a machine learning model (27) using a training dataset (28) having a set of distributional parameters θ is disclosed. The method comprises selecting (S205) stochastically a subset (28a) from the training dataset (28), calculating a gradient of the subset (28a) by encoding (S210) a cost function representative of the gradient into a quantum circuit, amplifying (S220) the amplitude of the quantum circuit, constructing a likelihood function (S230), and minimising the cost function in a variational quantum circuit to optimise the distributional parameters, extracting the optimised distributional parameters, and entering the optimised distributional parameters into the machine learning model (27).

This application discloses a design method for an optoelectronic imaging system with laser protection and privacy protection functions, relating to the field of optoelectronic imaging. The method includes: determining an optimized point spread function based on preset point spread distribution characteristics, target energy divergence, and point spread radius using an optical-algorithm joint optimization framework; determining an initial phase distribution based on imaging parameters using the Gerchberg-Saxton algorithm; adjusting the initial phase distribution using a stochastic gradient descent algorithm, guided by the difference in energy divergence of the point spread function, to determine the optimized phase distribution corresponding to the optimized point spread function; and using the optimized phase distribution to load the phase modulation component of the optoelectronic imaging system. This application can achieve high-quality optoelectronic imaging with laser protection and privacy protection capabilities.

This invention relates to a knowledge graph-based method for recommending manufacturing resources, comprising the following steps: establishing a supply and demand information model for the manufacturing domain; constructing a manufacturing domain ontology model to represent the concepts and relationships of demand information and manufacturing resources, serving as the schema layer of the manufacturing domain knowledge graph; performing knowledge extraction; utilizing the knowledge graph for visualization; implementing knowledge graph embedding, during training, obtaining erroneous triples (i.e., negative samples) by randomly replacing head entities; continuously optimizing the loss function using stochastic gradient descent to obtain qualified embedding vectors; after training the vector representations of entities using the TransE model, using vector value calculation to measure the degree of conformity of each resource, and employing cosine similarity calculation to calculate the similarity between vectors; and adding resource QoS service matching based on feature matching.

The application discloses a semi-supervised image hashing method and device based on joint prediction probability and relates to the technical field of image processing. The method comprises the following steps: performing end-to-end joint inference by using a joint prediction semi-supervised hashing network according to a labeled image dataset and an unlabeled image dataset; performing neighbor aggregation on an unlabeled hash code according to labeled data features and unlabeled data features; calculating a total loss according to the labeled image dataset, a labeled hash code, the unlabeled hash code, a first-class probability distribution, a third-class probability distribution, a pseudo-similarity label and an aggregated hash code; and performing parameter optimization on the joint prediction semi-supervised hashing network by using a stochastic gradient descent method according to the total loss to obtain an optimized joint prediction semi-supervised hashing network. The application is an efficient and accurate semi-supervised image hashing method for combining the prediction probability of known class data and unknown class data.

The application discloses a financial data risk analysis method and device based on reinforcement learning and a medium, relates to the technical field of data risk analysis, and comprises the following steps: performing time sequence processing on splicing indexes of financial related data to determine a financial time sequence vector; performing time period division feature extraction on the financial time sequence vector to obtain a financial feature vector; determining a financial plan risk value based on the financial feature vector by learning the financial plan risk analysis of an intelligent agent; obtaining a trained reinforcement learning intelligent agent through parameterized learning of a stochastic gradient descent update according to the financial plan risk value; inputting the financial time sequence vector into the reinforcement learning intelligent agent to determine a risk value function, and determining a financial data risk analysis report through plan feature analysis based on the risk value function. The application solves the technical problems of insufficient timeliness and adaptability and dependence on manual processing in the prior art by using the above method.

The application discloses a kind of attention enhancement partial least squares prediction methods for strong-flavor base liquor sensory evaluation, comprising the following steps: 1 obtains the concentration data of key flavoring substances of strong-flavor base liquor and derived feature dataset;2 build feature extraction and latent variable iteration module, through residual update and weight recursion, complete multiple rounds of flavor characteristics and sensory attribute association information mining;3 introduce attention weight calculation module, the self-adapting weighting of key latent variable information is carried out, and the focusing ability of model to flavor-sensory relationship is strengthened;4 build loss function and optimize model by stochastic gradient descent algorithm, realize the accurate prediction of strong-flavor base liquor sensory attribute.The method of the application can realize the objective, quantitative prediction of strong-flavor base liquor sensory attribute, and provides stable and reliable technical support for base liquor sensory quality evaluation.

This invention discloses a federated learning communication optimization method and apparatus based on local and global dual pruning. The method includes: calculating the cosine similarity between the global parameters from the previous iteration and the parameters after local iteration; making adaptive bit-width decisions based on the similarity, updating parameters locally using pruned stochastic gradient descent, and reporting the bit-width label and quantization gradient to the server; grouping and aggregating clients according to the bit-width label on the server side, using robust aggregation within each group, and then performing weighted global fusion of the results from each group to obtain the global gradient; and updating model parameters using pruned stochastic gradient descent. Compared with existing technologies, this invention, through a local similarity adaptive quantization-group robust aggregation-global pruning end-cloud linkage mechanism, effectively reduces communication overhead, suppresses aggregation bias and oscillation, and improves convergence stability and accuracy in scenarios with non-independent and identically distributed systems and local multi-step processes.

This invention discloses a source-grid-load-storage coordinated scheduling method based on divergence regularization and bibliometric optimization, belonging to the field of power system optimization. Addressing the problems of random output fluctuations, distribution shifts, insufficient robustness of traditional methods, and inefficiency in solving problems under high-proportion wind and solar grid integration, this invention first constructs a continuously differentiable comprehensive loss function containing multiple uncertainties and soft and hard constraints. Then, it uses a generalized Sinkhorn distance with χ²-divergence regularization and Gaussian reference measure to construct a fuzzy set, equivalently transforming the original Min-Max problem into a nested two-layer continuous dual model. Finally, it employs a nested stochastic gradient descent algorithm, with the inner layer estimating the dual multipliers and the outer layer updating the scheduling strategy and projecting it onto the physical feasible region, achieving a rapid solution. This invention overcomes the limitations of historical data support sets, can prevent risks from unknown extreme weather events, avoids the dimensionality curse, is compatible with first-order AI algorithms, and balances grid operation safety and economy, making it suitable for robust scheduling of large-scale, high-proportion renewable energy power systems.

This invention discloses a method for reconstructing complex substation engineering scenes based on a 3D Gaussian distribution function, belonging to the field of substation engineering scene reconstruction. Addressing the shortcomings of existing methods in reconstructing complex substation engineering scenes with occlusion and changing viewpoints, this method describes the 3D distribution of scene data points using a 3D Gaussian distribution function and optimizes the scene representation using a camera coordinate system. It integrates stochastic gradient descent, Gaussian function adaptive control, and a fast rasterization algorithm to achieve rapid iterative scene reconstruction. Object-level key points are extracted and their scale calibrated. The distance between key points is minimized by matching scene data points with the corresponding surface normal vectors of the CAD model using angle constraints. Optimized equipment and structures are embedded, and multi-source data is fused to obtain the final scene. This method achieves rapid and accurate structural reconstruction of complex substation engineering scenes, improving reconstruction efficiency and accuracy, providing a new solution and laying a foundation for the development of this field.

The present invention relates to a Bernoulli sampling-based interpretable CNN training method and apparatus, and a medium. The method comprises the following steps: inputting an image into a CNN to obtain response feature maps of filters; for the response feature maps, performing Bernoulli sampling to obtain a binarized assignment matrix; calculating average weight matrices of the filters for image categories on the basis of the binarized assignment matrix, and calculating the sum of pairwise differences; calculating a Hadamard product of an assignment vector of the binarized assignment matrix and the response feature maps to obtain masked feature maps; separately inputting the response feature maps and the masked feature maps into a CNN fully-connected layer to respectively obtain classification prediction probability vectors, and respectively calculating cross entropy losses between the prediction results and a ground-truth label; and on the basis of the sum of the pairwise differences, and the cross entropy losses, using a stochastic gradient descent method to implement network training to obtain an interpretable CNN for image classification. Compared with the prior art, the present invention has the advantages such as high adaptability and high interpretability.

The application discloses a kind of robot process online regulation and control methods suitable for material uncertainty, and it is related to intelligent control technical field.The application collects the process parameter sequence of the component to be processed, initial topography data and post-processing topography data;Through the differentiable topography evolution agent model, the derivable predicted topography data is output;A two-parameter inversion optimization model is constructed, the actual and predicted topography difference is quantified by mean square error, and the minimum is taken as the goal, the material parameter prior L2 norm constraint and process parameter boundary penalty constraint are embedded;Based on the reverse automatic differentiation algorithm, the gradient vector is solved, and the optimal material parameter is obtained by the random gradient descent algorithm with momentum iterative optimization, and then the future time step process parameter sequence is optimized, and the optimal solution of material parameter and process parameter is output after constraint verification and convergence judgment.The application realizes the accurate online regulation and control of robot process under material uncertainty, and has wide application value.

Methods, systems, and apparatus, including computer programs encoded on computer storage media for training a machine learning model. The method includes obtaining a training data set comprising a plurality of training examples; determining i) a stochastic gradient descent step size schedule, ii) a stochastic gradient descent noise schedule, and iii) a stochastic gradient descent batch size schedule, wherein the stochastic gradient descent batch size schedule comprises a sequence of varying batch sizes; and training a machine learning model on the training data set, comprising performing stochastic gradient descent according to the i) stochastic gradient descent step size schedule, ii) stochastic gradient descent noise schedule, and iii) stochastic gradient descent batch size schedule to adjust a machine learning model loss function.

The application discloses an active noise control method and device, electronic equipment and a computer readable storage medium, comprising: collecting a noise source signal, modeling the impulse noise component in the noise source signal, and generating a reference signal; based on the error signal collected by an error microphone, inputting the anti-impulse adjustment parameter and the error signal into a preset function to construct an instantaneous cost function represented by the mathematical expectation of the square of the preset function; calculating the gradient of the instantaneous cost function, and deriving an update formula of an adaptive filter weight vector by using a stochastic gradient descent method; at each sampling time, filtering the reference signal by using the weight vector at the current time to generate a controller output control signal, generating a counteracting sound wave through the secondary path, collecting a residual error signal, calculating the weight vector at the next sampling time according to the update formula, and iteratively executing the above operation to complete noise cancellation. The application realizes efficient suppression of the impulse noise.

A first calibrator detects mismatches in the capacitor array and updates a look-up table (LUT) with calibrated weights that are copied to a positive LUT and a negative LUT, which are then adjusted for non-linear errors by a second calibrator using a least mean square (LMS) algorithm. The binary code in a successive approximation register (SAR) is complemented to generate a complement with a sign bit. When the sign bit is positive, entries from the positive LUT with complemented data bits = 1 are read and summed, a first offset is added, and the sum is normalized to obtain a corrected code. When the sign bit is negative, entries from the negative LUT with complemented data bits = 0 are read and summed, a second offset is added, and the sum is normalized to obtain a corrected code. A multivariate stochastic gradient descent generates polynomial coefficients that further correct the corrected code.

This application discloses a humanoid robot motion control method with joint constraints of text and trajectory, belonging to the field of humanoid robot control technology. The method first obtains the text description and motion trajectory corresponding to the preset control action. Through motion synthesis with joint constraints of text and trajectory, a SMPL digital human pose sequence conforming to dual constraints is generated. An optimization strategy combining stochastic gradient descent and momentum methods is used to complete motion redirection, obtaining pose data adapted to the target humanoid robot. A teacher-student model is constructed based on a proximal policy optimization algorithm, using the redirected pose data as prior guidance. Policy distillation is performed through a dataset aggregation algorithm to train the motion control model. Finally, the model is deployed to the robot, and motion control is achieved using the joint target angles output by the model as driving signals. This application provides high-quality prior guidance through dual constraints, effectively improving the semantic consistency, trajectory accuracy, anthropomorphism, and scene robustness of humanoid robot motion, significantly reducing training costs, and possessing excellent engineering application value.

This invention discloses a method and system for optical character recognition of printed Burmese text based on deep learning, belonging to the field of optical character recognition technology. The method first preprocesses scanned images of printed Burmese documents, using Hough transform to segment table and character regions and extract target character regions. Then, it constructs an image dataset and builds a Burmese OCR recognition model. The model is trained using a stochastic gradient descent optimizer with momentum, and parameters are optimized through 4-fold cross-validation to obtain a high-precision recognition model. Finally, the segmented character regions are input into the model to complete detection and classification, outputting text information and importing it into a data table. This invention can accurately segment Burmese characters within tables, improve model recognition accuracy, has low deployment costs, and can be used for detecting duplicate voter registration in Burmese elections, effectively improving the efficiency of Burmese document digitization and information extraction.

A computer-implemented method for training a machine learning model using a training dataset having a set of distributional parameters θ is disclosed. The method comprises selecting a subset from the training dataset, calculating a gradient of the subset by encoding a cost function representative of the gradient into a quantum circuit, amplifying the amplitude of the quantum circuit, constructing a likelihood function, and minimising the cost function in a variational quantum circuit to optimise the distributional parameters, extracting the optimised distributional parameters, and entering the optimised distributional parameters into the machine learning model.

The application provides an on-chip training method of an in-memory computing memory artificial neural network, and belongs to the field of artificial neural network algorithm optimization.The application follows the Manhattan rule idea, introduces a probability-based ternary update rule, converts high-precision weight update in an ideal classical error back propagation algorithm BP algorithm into ternary weight update, only applies at most one programming pulse to one device in each training batch, reduces the operation times, the training method converges fast and is stable, the recognition accuracy is high after training, the original BP algorithm is slightly changed, and the performance exceeds the Manhattan and threshold-Manhattan rules from the algorithm perspective;The application can efficiently realize on-chip stochastic gradient descent SGD and mini-batch gradient descent MBGD, does not need to store high-precision weight update values, reduces the additional hardware overhead, and optimizes the design of the inference circuit.

This invention proposes a method for image segmentation of head and neck tumor lesion regions and a computer-readable medium. The invention acquires multiple sets of original head and neck tumor PET-CT images, sequentially preprocesses them to obtain preprocessed images for each set, and labels them with corresponding ground truth classification tags. A lesion image segmentation network is constructed, and lesion segmentation prediction is performed using the input of each preprocessed image set to obtain a head and neck tumor prediction probability map for each preprocessed image set. Combining the ground truth classification tag of each pixel of the head and neck tumor lesion in each preprocessed image set, a cross-entropy Dessell weighted loss function is constructed, and the network is optimized and trained using a stochastic gradient descent algorithm to obtain a trained lesion image segmentation network. Real-time acquired head and neck tumor PET-CT images are then used to predict and segment the lesion region using the trained lesion image segmentation network, and probability thresholds are determined to obtain the real-time pixel range of the head and neck tumor lesion region. This invention utilizes the complementary information between multiple modalities to improve the accuracy of pixel region segmentation prediction for head and neck tumor lesions.

The application discloses a kind of based on light network and data enhancement's microdefect nondestructive testing method, it is related to metal component defect detection technical field.The method first prepares microdefect sample, collects and pre-processes the metal defect image containing uneven reflection, oil stain background noise, constructs microdefect dataset;Again, the improved target detection network consisting of MobileNet v3 main network, feature fusion network and detection head of integrated ELA attention mechanism and CARAFE operator is built;Color disturbance, geometric transformation, image flipping, mixed and segmentation copy and paste data enhancement operation are implemented on dataset, SIoU positioning loss function and random gradient descent optimizer are used to train network;Finally, the image of metal component to be measured is input into the network trained, and the defect position and boundary box are output.

This invention proposes a text-image matching method and system based on semantic segmentation and feature association, relating to the field of semantic processing technology. The method includes: performing semantic analysis on target text and target images in response document samples to obtain response text features for different paragraphs and response image features for each target image; mapping the response text features and response image features to the same dimensional space to obtain text-image combinations, using these combinations as input to an initial association model, and training the initial association model using the cosine similarity between the response text features and response image features as output; optimizing the initial association model using a loss function and a stochastic gradient descent optimization algorithm to obtain a target association model; preprocessing the question text to generate corresponding question text features, inputting these features into the target association model to obtain the corresponding text-image combinations, and displaying the text paragraphs and target images corresponding to the text-image combinations.