278 results about "Stochastic gradient descent" patented technology

Approximate closed-form bounds of Bayes security against record-level inference attacks, in particular membership and attribute inference attacks, are provided. In various embodiments, these approximate bounds of Bayes security are used in conjunction with training neural-network models by differential-privacy stochastic gradient descent to create trained models that achieve a desired level of privacy.

The invention relates to a coronary artery plaque rupture risk assessment method based on a non-local physical information neural network, and belongs to the field of artificial intelligence, and the method comprises the following steps: S1, introducing a near-field dynamics theory, reconstructing a mechanical constitutive equation of plaque rupture, and deducing a crack propagation criterion; s2, constructing a non-local physical information neural network, taking geometric structure information and material attributes of the coronary artery plaque as input, and taking a displacement field and a stress field of the plaque as output; taking the plaque mechanics basic equation as a physical constraint condition, and adding a loss function; s3, constructing a data set; s4, training the physical information neural network by using the training set, and adjusting network parameters by using a stochastic gradient descent optimization algorithm to minimize a loss function; optimizing and testing the model; and S5, inputting a coronary artery medical image number to be evaluated into the trained non-local physical information neural network, and outputting a rupture risk evaluation result of the coronary artery plaque.

The invention relates to the technical field of GIS (Geographic Information System) equipment denoising, in particular to a GIS local interference source denoising method, which comprises the following steps of: constructing a hybrid architecture of a quantum convolutional neural network and a quantum recurrent neural network, extracting features by using a variable component sub-circuit as a convolution kernel in the quantum convolutional neural network, processing time sequence data by using quantum bit entanglement in the quantum recurrent neural network, and extracting a local interference source in the quantum recurrent neural network; normalization, coding and enhancement are carried out on electrical, electromagnetic and other multi-source data, a quantum stochastic gradient descent algorithm is used for training, a multi-modal fusion network based on an attention mechanism is also constructed, features of all modals are extracted through an independent convolutional layer, then splicing and weighted fusion are carried out, and finally a quantum reinforcement learning environment is constructed. An intelligent agent strategy network and a reward function are designed based on a quantum neural network, parameters are adjusted according to quantum Q learning, quantum calculation, deep learning and a multi-modal fusion technology are integrated, the GIS local interference source denoising effect is effectively improved, and stable operation of equipment is guaranteed.

The invention discloses a medical auxiliary diagnosis method and system based on a convolutional neural network, and relates to the field of medical big data analysis, and the method comprises the steps: collecting various types of medical data, which comprise medical image data, electronic medical record text data and vital sign monitoring data; constructing a convolutional neural network model with a multi-branch structure; preparing a training data set containing labeled samples, wherein labeled information comprises disease diagnosis results and disease severity grades; training the convolutional neural network model by using the training data set, dynamically adjusting the learning rate in the training process by adopting a stochastic gradient descent optimization algorithm and combining a learning rate attenuation strategy, and introducing a regularization method; inputting to-be-diagnosed medical data into the trained convolutional neural network model, and outputting disease diagnosis probability distribution and disease severity prediction results through forward propagation calculation of the model; and generating a medical auxiliary diagnosis report according to the output prediction result.

Aspects of the disclosure are directed to implementing a projection-based stochastic gradient descent technique that maintains label differential privacy when training one or more machine learning models. The technique includes denoising gradients by exploiting projections when training the machine learning models to improve performance of the trained machine learning models while maintaining label differential privacy. For instance, the projection-based stochastic gradient descent technique can improve performance of machine learning models in higher-privacy regimes, such as digital content management.

The invention provides an electromechanical equipment fault prediction method and system based on multi-source information fusion, and the method comprises the steps: collecting operation data, including vibration data, temperature data and current data, during the operation of electromechanical equipment; performing feature extraction on the operation data based on a principal component analysis algorithm to obtain a fusion feature vector; inputting the fusion feature vector into a fault prediction model based on a deep belief network, and outputting a prediction result; wherein the deep belief network adopts a small-batch stochastic gradient descent algorithm combined with an adaptive learning rate adjustment strategy during training; and judging whether the electromechanical equipment has a fault hidden danger or not according to the prediction result. According to the method, the relevance between different types of data is mined, and the defect of low prediction precision is overcome.

The invention discloses a physiotherapy instrument control system and method based on artificial intelligence, and relates to the technical field of intelligent medical treatment, and the method comprises the steps: collecting an electromyographic signal in real time, carrying out band-pass filtering and TK energy operator processing, and outputting a muscle activation feature vector; based on the muscle activation feature vectors, spatial-temporal features are extracted through a convolutional neural network, and a physiotherapy parameter combination is constructed; a PID controller is adopted to adjust the physical therapy parameter combination in real time, and optimized physical therapy parameters are output; performing electromagnetic thermal coupling simulation on the optimized physiotherapy parameters through finite element analysis to obtain a physiotherapy instrument control instruction; and updating the weight of the convolutional neural network by adopting a stochastic gradient descent algorithm according to the physiotherapy efficiency evaluation report, and outputting an optimization execution instruction. PID parameters are optimized through a quantum annealing algorithm, stimulation parameters are dynamically adjusted in combination with myoelectricity beta wave band phase locking, and treatment parameters and physiological response are synchronized.

The invention discloses a pre-training method and system for a drainage pipeline defect classification and grading model. The method comprises the steps of obtaining a drainage pipeline image without a label; performing updating training on the weight of a feature encoder by utilizing the obtained drainage pipeline image without the label, combining a composite transformation technology, designing an asymmetric branch comprising the feature encoder, reconstruction loss, similarity loss and a stochastic gradient descent algorithm, and completing pre-training of the feature encoder; performing feature extraction and region division on the drainage pipeline image by using a pre-trained feature encoder, generating a target object mask by combining a correlation technology of an image region feature vector, taking the target object mask as a supervision information training feature encoder weight, and selecting a part of the weight as an initialization weight of a backbone network; pre-training of the backbone network is completed; according to the method, the drainage pipeline image without a label can be effectively utilized for pre-training, and the adaptability and precision of the defect detection model are improved.

This invention discloses a SAR target detection and recognition method based on a multi-level enhancement network. This method primarily addresses the problems of existing technologies in complex environments, such as poor robustness, high false alarm and missed detection rates, and low detection and recognition accuracy. The method involves labeling and partitioning SAR measured data to obtain training and test sets; constructing a multi-level enhancement network consisting of a cascade of data-level enhancement modules, feature-level enhancement modules, region proposal modules, and decision-level enhancement modules; training the multi-level enhancement network using the training set based on a stochastic gradient descent algorithm; and inputting the test set images into the trained multi-level enhancement network to obtain SAR target detection and recognition results. This invention significantly improves SAR target detection and recognition performance in complex environments and can be used for battlefield reconnaissance and situational awareness.

The invention discloses a high-elasticity antibacterial yarn production control method and system, and relates to the technical field of textile manufacturing. The method comprises the following steps: acquiring tension, temperature, antibacterial agent distribution and other multi-dimensional state information, and generating a comprehensive performance evaluation index based on a feature projection and gating residual fusion model; the index is compared with target performance, and key influence parameters such as fiber arrangement unevenness and antibacterial agent application distribution density are identified; calling a process response model to generate a parameter adjusting strategy, and driving an adjusting device to adjust process variables such as tension and temperature through a control instruction; and finally, based on a stochastic gradient descent algorithm and in combination with an expert knowledge base, adjusting rules are periodically updated, and performance self-adaptive control is achieved. According to the method, the intelligence and robustness of regulation and control of the elasticity and the antibacterial performance of the yarn are improved, and the method is suitable for industrial production of high-end textiles such as medical yarn and intelligent wearing fabric.

The invention discloses a pavement crack recognition and classification method based on multi-feature scale fusion Faster-R-CNN, and particularly relates to the technical field of road detection, and the method comprises the steps: obtaining pavement image data of a research region through an automobile carrying a high-definition camera, and carrying out the preprocessing of the pavement image data to form a required image data set; a ZFNet is used as a crack feature extraction module to perform multi-feature fusion and enhance feature information, an RPN full convolutional network is combined to perform detection by optimizing an anchor box generation mode, and back propagation and stochastic gradient descent are used to optimize and extract a disease candidate region to realize disease classification and frame regression; according to the Faster-R-CNN target detection algorithm, the detection precision is improved through a Soft-NMS algorithm, the Soft-NMS algorithm is the same as an NMS algorithm in the execution process, but function operation is used for original confidence score, the objective is to reduce the confidence score, reduce the omission ratio and improve the detection precision, and construction of the Faster-R-CNN target detection algorithm is achieved.

The invention discloses a road crack deep learning segmentation method for a complex scene, and belongs to the technical field of image processing. Firstly, multi-illumination and multi-material road surface images are acquired at high altitude through carrying multi-spectral equipment by an unmanned aerial vehicle, and generalization is improved by adopting adaptive threshold graying, CLAHE illumination equalization and GAN data enhancement. Secondly, deformable convolution, a self-adaptive space attention mechanism and an object context representation module are introduced into the HRNet backbone network, and the problems of irregular fracture shapes and background interference are solved; and finally, class imbalance is relieved by adopting a focus loss function, and robustness is improved by combining stochastic gradient descent and a verification set dynamic weight storage strategy. In the test stage, edge equipment deployment is carried out, a pixel-level segmentation result is output, crack parameters are quantified, and decision support is provided for road maintenance. The method can significantly improve the crack segmentation precision in a complex scene, and especially has an engineering practical advantage in a fine crack and noise interference scene.

The invention discloses a machine olfaction system drift compensation method based on dynamic domain adversarial learning. The method comprises the following steps: step 1, collecting sample data through a machine olfaction system; step 2, constructing a DRDAN, and inputting the sample data into the DRDAN to obtain a predicted gas category and a domain classification label of the sample data; the DRDAN comprises a feature extractor, an integrated gas classifier and a domain discriminator, the integrated gas classifier is connected with the feature extractor, the integrated gas classifier outputs samples and finally predicts gas categories, the domain discriminator is connected with the feature extractor through a gradient inversion layer, and the domain discriminator outputs domain classification labels of sample data; and obtaining an optimal model parameter by minimizing a DRDAN learning objective function, and performing gradient updating through stochastic gradient descent. The method has the advantages that cross-domain joint distribution alignment is achieved, and class distinguishability is improved.

The invention discloses a lumbar disc herniation postoperative recurrence prediction system based on a multi-modal medical image, and belongs to the field of medical images. The lumbar disc herniation postoperative recurrence prediction model based on the multi-modal image is trained by using the lumbar vertebra image set, the feature true value corresponding to each image and the recurrence prediction label corresponding to each patient, and in the training process, model parameters are updated by using a stochastic gradient descent algorithm; and a trained lumbar disc herniation postoperative recurrence prediction model based on the multi-modal image is obtained. According to the method and the system, due to rich image features provided by the multi-modal image, the prediction network trained based on the multi-modal data shows higher accuracy in postoperative recurrence prediction of the lumbar disc herniation.

The invention belongs to the technical field of production control, and provides a combined optimization method and system for yarn combined weaving parameters. The method comprises the following steps: acquiring sample data containing first parameters such as yarn twist and elasticity and corresponding interaction characteristics, and sample labels such as pattern definition and cloth cover flatness; and inputting the data into the initial neural network, and calculating a loss value by combining a loss function which contains a mean square error and reflects a physical interaction weighted item through feature transformation of an input layer and a hidden layer. And by using a stochastic gradient descent algorithm, when the loss value is greater than a threshold value, performing back propagation to update the parameter training model until the loss value of the model meets the requirement and is not over-fitted in the verification set. Physical constraints such as influence of yarn twist and elastic interaction on pattern definition and influence of fabric density and twist on flatness are embedded through the mechanism sensing layer, effective interaction characteristics are constructed, and accurate prediction of silk fabric indexes is achieved.

The invention relates to a small model side privacy data enhancement method and system based on large and small model collaboration, and the method comprises the steps: determining a generator at a user side and a loss function of the generator according to a data mode type, training the generator at the user side through employing a differential privacy stochastic gradient descent method, generating synthetic data through employing the trained generator, and carrying out the prediction of the synthetic data. The synthesized data is sent to a server side; the server side calculates embedded vectors of the synthetic data by utilizing a large model, clusters the embedded vectors by adopting a clustering mode, and filters clustering results; sending an enhanced data set containing the user itself and the non-user itself and an embedded vector of each data in the enhanced data set to a user side; and after the user side fuses the enhanced data set and the local original data, the generator is updated.

This specification relates to privacy-preserving model training on tabular data. In some aspects, a method includes receiving, by one or more computing devices, tabular data; serializing the tabular data into a natural language string in a natural language format; combining the natural language string and a prompt as an input to a pretrained large language model (LLM) to generate a predicted result, wherein a set of learned vectors are added into the pretrained LLM for fine-tuning the pre-trained LLM; fine-tuning the pretrained LLM using a differential privacy stochastic gradient descent (SGD) process, wherein fine-tuning the pretrained LLM comprises: determining values of the learned vectors that minimize a difference between the predicted result and the ground truth; receiving a request including test tabular data for a predication task; and generating, in response to the request for the prediction task, a prediction result for the test tabular data using the fine-tuned LLM.

The invention provides an electroslag remelting smelting endpoint element dynamic forecasting and process decision optimization method and system, and the method comprises the steps: obtaining a historical data set and a real-time data flow in a smelting process, and carrying out the preprocessing of data; an online stochastic gradient descent and incremental learning algorithm is introduced, an element dynamic prediction model is constructed, sensor data in the smelting process are received in real time, and an element concentration prediction model is dynamically optimized; on the basis of a depth deterministic strategy gradient algorithm, an optimization decision model is constructed, and a smelting operation strategy is adjusted in real time according to element components, smelting working conditions and other environment parameters predicted online by the element dynamic forecasting model; and a real-time feedback mechanism is set, so that the element dynamic prediction model and the optimization decision model work cooperatively, and element concentration prediction and real-time synchronous adjustment of a control strategy are realized. The smelting parameters can be adjusted in real time, element prediction precision is improved, energy consumption is reduced, manual intervention is reduced, and smelting efficiency and quality stability are remarkably improved.

The invention provides a structured feedback risk constraint LQR solving method based on optimal control optimization, and belongs to the technical field of control optimization, and the method comprises the steps: obtaining an initial feedback gain matrix meeting a microgrid communication topology sparsity constraint; a zero-order strategy gradient method is adopted, through an outer layer iteration optimization step and an inner layer sampling estimation sub-step, a gradient direction is estimated by utilizing observation change of a target function value; according to the distributed frequency control method, strategy updating is carried out on the basis of an optimal control algorithm, a sparse feedback gain matrix obtained after N rounds of iteration is output, the sparse feedback gain matrix is directly deployed to a local controller of the micro-grid, and distributed frequency control meeting the risk constraint is achieved. Compared with a classical stochastic gradient descent method, the OCP algorithm obtains a satisfactory optimality gap, the convergence speed is higher, and the stability is also improved through dual redescription of maximum and minimum problems.

Methods, systems, and apparatuses, including computer programs encoded on computer storage media, for training a neural network using differentially private stochastic gradient descent (DP-SGD) with correlated noise. In some implementations, a system determines a correlation matrix by performing an optimization to improve a utility metric that is dependent on a noise multiplier. The system can then train the neural network using the optimized correlation matrix and a privacy-amplifying batching scheme to produce a neural network with high utility while satisfying a data security criterion.

The invention provides a discrete manufacturing production line process decision and optimization method and system based on digital twinning, and relates to the field of production intellectualization and digitalization, and the method mainly comprises the steps: constructing a high-fidelity digital twinning model corresponding to a physical production line through the combination of mechanism modeling and data-driven modeling; establishing a virtual-real consistency evaluation index system, and realizing deviation identification and dynamic consistency maintenance of the digital twinborn model; production disturbance, equipment state change and process deviation information are sensed in real time; a part-process-equipment-quality multi-dimensional correlation model is constructed; an improved stochastic gradient descent algorithm is introduced to carry out dynamic updating and convergence control on intelligent agent strategy network parameters, and a knowledge base self-evolution updating mechanism is constructed; according to the method, dynamic updating and consistency maintaining of the twinborn model are achieved, the dynamic response capability of the digital twinborn model under the dynamic operation condition is remarkably improved, and the systematicness and reusability level of process knowledge are effectively improved.

The invention discloses a semi-supervised radar target identification method based on a hybrid expert network and pseudo label filtering. The method comprises the following steps: S1, carrying out data preprocessing on radar target echoes; s2, constructing a data set X1 with labels and a data set Xu without labels; s3, constructing a hybrid expert network model; s4, constructing a fusion loss function; s5, constructing a semi-supervised model training framework; s6, performing iterative optimization on the fusion loss function of the semi-supervised deep neural network model by using a stochastic gradient descent method in combination with the mixed data set, and finally generating a new semi-supervised deep recognition model Mb; and S7, classifying the test sample set X *: inputting the test sample set X * into the trained semi-supervised depth recognition model to obtain a prediction label to evaluate the performance of the model, and inputting an unknown sample into the model to obtain a prediction result so as to realize automatic classification of radar target echoes.

A computer-implemented method and a distributed computer system for privacy-preserving distributed training of a global neural network model on distributed datasets. The system has data providers communicatively coupled with each having a respective local training dataset and a vector of output labels for training the global model. Further, it has a cryptographic distributed secret key and a corresponding collective cryptographic public key of a multiparty fully homomorphic encryption scheme, with the weights of the global model being encrypted with the collective public key. Each data provider computes and aggregates, for each layer of the global model, encrypted local gradients using the respective local training dataset and output labels, with forward pass and backpropagation using stochastic gradient descent. One data provider homomorphically combines the current local gradients of the data providers into combined local gradients, and updates the weights of the current global model based on the combined local gradients.

The invention discloses a heterogeneous split learning collaborative optimization method and system based on a feature alignment adapter, and belongs to the edge computing and distributed machine learning technology. Based on an edge splitting learning framework, each client constructs a personalized local underlying model according to local computing resources and data distribution characteristics, and performs cooperative training with a top model shared by the server. According to the method, a feature adapter module is introduced into a model splitting layer to align heterogeneous intermediate feature representations from different clients, so that training obstacles caused by inconsistency of a model architecture and output dimensions are eliminated. The method further comprises the steps of constructing an initialization mechanism to accelerate convergence of splitting training, including initialization of an adapter and a feature distillation-based heterogeneous client model initialization strategy, and further jointly optimizing the server model, the adapter and the heterogeneous client model based on a global small-batch stochastic gradient descent algorithm, so as to improve the resolution of the heterogeneous client model. And the split model accuracy and the training efficiency under the equipment resource and data heterogeneous environment are effectively improved.

The invention discloses a lithium ion battery RUL prediction method based on MBGD-WGAN-GRU (Model-Based Graphics Decomposition-Wavelet German-Based Graphics According to the method, firstly, health factors highly related to the aging characteristics of the lithium ion battery are mined from original monitoring data through a grey relational analysis method, a characteristic space reflecting the degradation trend of the battery is constructed, and a data set is divided into a training set and a test set according to the proportion; then, a small-batch stochastic gradient descent optimization generative adversarial network is adopted, and high-fidelity synthetic data is generated; then training a GRU (Gated Recirculation Unit) model by using the extracted health factor and the extended training set, capturing a dynamic evolution law of battery aging, and realizing high-precision prediction of the RUL; and finally, battery RUL prediction and verification of the GRU model are realized through a public test data set. According to the method, the data expansion of the lithium ion training set is realized by optimizing the WGAN through the MBGD, the GRU model is trained by using the health factor sequence and the expanded training set, the dynamic evolution law of the aging of the lithium ion battery is captured, and the high-precision prediction of the RUL is realized.

The invention discloses a coal-rock mass stress field inversion method and system driven by a mechanical mechanism. The inversion method comprises the following steps: obtaining detection monitoring data; constructing a neural network model of a multi-layer perceptron through a loss function, taking multi-point stress monitoring number stress as an initial value condition, combining a boundary condition and a mechanical equation, and performing iterative solution to obtain global stress field distribution data; constructing an acoustic emission signal as input, global stress field distribution as output, a long-short-term memory network as a main structure, and combining a full connection layer to perform feature extraction and mapping; constructing a loss function with a self-training weight according to the properties of the coal and rock mass, the mechanical balance and the constitutive model; and optimizing the deep neural network according to a stochastic gradient descent method and an adaptive learning rate, and training the deep neural network by using back propagation according to a constructed loss function with a self-training weight to obtain an inversion model.

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).

The invention provides a method and system for detecting small target defects in a paper cup based on deep learning, and relates to the technical field of computer vision, and the method comprises the steps: obtaining a surface image of a to-be-detected paper cup; constructing a paper cup small target defect detection model based on a deep learning algorithm; constructing a loss function of the paper cup small target defect detection model; with minimization of the function value of the loss function as a target, optimizing the small target defect detection model in the paper cup in combination with a back propagation algorithm and an SGD stochastic gradient descent method; and inputting the surface image into the optimized small target defect detection model in the paper cup for detection, and outputting a target defect detection result of the to-be-detected paper cup. According to the invention, all products can be detected in time, the phenomena of missing detection and false detection are avoided, and the production efficiency and the product quality are improved.

The invention discloses an unmanned aerial vehicle small sample target recognition method based on spatial information mining, and aims to solve the problems that an existing small sample method does not fully utilize an image structure in a scene, ignores a semantic relationship between support and query samples and is limited in recognition precision. The core is a query-support transformer framework for unified modeling of global and local structural relationships of samples, the framework comprises a cross-scale interactive feature extractor, a sample transformer and a block transformer module, a task-type mechanism is adopted in the training stage, enhanced features are generated by images through the extractor, the sample transformer constructs a global semantic relationship, and the overall semantic relationship is obtained. The method comprises the following steps: realizing local block-level matching by a block transformer, performing weighted summation on similarity of the two to obtain a matching score, combining weighted sum of comparison loss and cross entropy loss, updating parameters by using momentum stochastic gradient descent, constructing a task based on a trained model in a test stage, extracting features, and fusing global and local similarity to complete classification. And the unmanned aerial vehicle image small sample classification performance is effectively improved.

The invention belongs to the technical field of image processing, and discloses a multi-target hierarchical aggregation federated learning optimization method, which comprises the following steps: acquiring a plurality of images; the method comprises the following steps: establishing a four-objective optimization model of federated learning, taking convolutional neural network parameters as decision variables, taking federated learning as a basic training framework, and processing an image, specifically, initializing a global model population; setting an objective function of federal learning training; the plurality of initialized models are distributed to clients, and each client executes a certain number of rounds of stochastic gradient descent based on local data and updates a model copy received by the client; after training is completed, a hierarchical asynchronous aggregation mechanism is introduced to a server side, and average sparse processing is carried out on aggregation updating frequency of deep model parameters in the whole communication round; and obtaining an image processing result. According to the method, the convergence speed and generalization performance of the federated learning model are improved, and the problem of multi-target collaborative optimization in a dynamic network and data heterogeneous environment is solved.