107 results about "Simultaneous learning" patented technology

The present disclosure relates to the technical field of artificial intelligence, and provides a multimodal large language model training method, a correlation calculation method, and a label generation method. The multimodal large language model training method comprises: on the basis of a sample text feature vector, a sample image feature vector, and a sample first multimodal feature vector which are obtained by processing image information and text description information of a sample commodity by a pre-trained multimodal large language model, training the pre-trained multimodal large language model to obtain a multimodal large language model; and processing a sample search word, the image information and the text description information on the basis of the multimodal large language model to obtain a sample search word feature vector, a first multimodal feature vector and a sample second multimodal feature vector, and training the multimodal large language model to obtain a trained multimodal large language model. The trained multimodal large language model of the present disclosure can simultaneously learn image features related to the search word and the text description information, and thus the generated second multimodal feature vector is more accurate.

The invention discloses a multi-label electrocardiogram classification method based on self-supervised pre-training and multi-modal semantic alignment, which belongs to the technical field of artificial intelligence, and comprises the following steps: realizing self-supervised pre-training of unlabeled data through a single-modal contrast enhancement network, generating global and local contrast views by adopting a multi-scale random cutting strategy, and classifying the global and local contrast views in a multi-scale random cutting mode; in combination with a teacher-student network architecture, the potential invariance features of the ECG signals are learned while negative sample dependence is avoided, the problem of annotation data scarcity is effectively relieved, and the feature robustness is improved. A multi-modal fusion mechanism based on label semantic guidance is provided, a time domain signal and a frequency domain time-frequency graph are mapped to a unified semantic space through fine-grained semantic alignment, local feature enhancement and cross-modal complementary information fusion are realized by using a cross attention mechanism, and the problem of semantic difference caused by modal heterogeneity in a traditional method is overcome. A multi-label comparison loss function based on a disease co-occurrence relation is proposed, a category discrimination boundary is dynamically optimized by modeling a label co-occurrence probability, the feature separability of a tail category is improved while the head category discrimination ability is enhanced, and the problem of sample category imbalance in a multi-label scene is remarkably relieved.

The invention discloses a wind power plant wind speed prediction method, medium and equipment based on a space-time diagram neural network in the technical field of wind power generation, and the method comprises the steps: carrying out the frequency enhancement of an original wind speed sequence, and obtaining a frequency-enhanced wind speed sequence; inputting the frequency-enhanced wind speed sequence into a time convolution module to obtain time sequence characteristics; constructing a wind speed sequence containing unsteady state information according to the original wind speed sequence, and inputting the wind speed sequence containing the unsteady state information into a graph learning module to construct a space graph structure; performing dynamic graph convolution on the time sequence feature and the space graph structure to obtain a first spatial-temporal feature; constructing a time-space unified graph containing time sequence features and space features, and simultaneously learning the time sequence features and the space features in the time-space unified graph through a Fourier graph neural network to obtain second time-space features; and fusing the first spatio-temporal feature and the second spatio-temporal feature, and obtaining a wind speed prediction value in combination with the prediction time step. According to the method, the generalization ability of the model is improved, and hidden ground frequency information in the data can be better learned.

The invention discloses an artificial intelligence speech recognition system, and the system comprises a multi-modal feature extraction module which employs an improved Conformer architecture to synchronously extract the time-frequency features and text embedding vectors of speech signals; the joint training module is used for performing joint optimization on ASR and NMT loss functions through an adversarial training strategy, learning voice recognition and machine translation tasks at the same time through joint training, and completing direct mapping from voice features to a target language; the context perception translation engine is used for integrating an attention mechanism of a pre-training language model, carrying out deep coding on the extracted speech features and generating cross-language semantic representation; the self-adaptive post-processing module is used for dynamically optimizing an output result by adopting a reinforcement learning framework, dynamically adjusting the output result according to a reward function, and optimizing translation quality and a speech synthesis effect; the dynamic language recognition module is a real-time language classifier based on a Wave2Vec 2.0 framework and is used for recognizing the language of the input voice in real time; and the incremental field adaptation module is used for quickly updating a field term library by using a LoRA fine tuning technology.

The invention discloses a multi-mode emotion continuous recognition method for medical treatment, belongs to the technical field of artificial intelligence and medical treatment information, and mainly aims to simulate the dynamic change process of emotion by establishing a Neural ODEs framework and overcome the static property and discreteness of emotion modeling in a traditional method. Through a causal inference technology, emotional features are separated from individual-independent physiological differences, and the generalization ability across individuals is improved. A self-supervised learning method is utilized, the synergistic effect between the EEG and the eye movement signal is improved through cross-modal contrast learning, and the emotion recognition precision is enhanced. The calculation complexity is reduced through a dynamic sparse attention mechanism, and meanwhile, focusing is performed on a key time slice in emotion recognition. Through multi-task joint learning, the model learns multiple tasks such as emotion intensity regression and tested identity recognition during emotion classification, and the personalized emotion recognition capability is improved.

The invention discloses a pipeline defect magnetic flux leakage detection method and device based on multi-scale data driving deep learning, and relates to the technical field of pipeline defect detection. A multi-scale magnetic flux leakage signal data set containing defect global distribution and local details is generated, and multi-level primary features are automatically extracted by using a convolutional neural network; and the defects of poor generalization and easy key information omission of a manual method are overcome. And then dynamically enhancing and performing weighted fusion on multi-scale features by means of a multi-scale convolution branch and an attention mechanism, so that the model can learn global and local features of the defect at the same time, the problem that the global and local features of the defect are difficult to consider in traditional deep learning is solved, and through transverse connection and up-sampling fusion of a feature pyramid, a multi-scale feature is obtained. And the features have high semantic information and high spatial details. And finally, independently carrying out multi-task prediction by virtue of a task decoupling module, and calibrating result consistency by virtue of a task alignment module, so that the detection accuracy in a complex scene is improved, and more accurate and robust pipeline defect magnetic flux leakage detection is realized.

The present invention belongs to the field of image processing, and discloses a method and system for quickly and efficiently identifying mesoscale convective systems in mid - low latitude regions, including obtaining cloud - top brightness temperature data; constructing a dataset for identifying mesoscale convective systems based on the cloud - top brightness temperature data; constructing a model for identifying mesoscale convective systems; training the model using the constructed training set; after reaching the specified number of training rounds, testing the model with the test set to obtain the final result of identifying mesoscale convective systems. The present invention designs a piece - by - piece learning strategy, enabling the model to simultaneously learn the characteristics of mesoscale convective systems in different regions; constructs a multi - scale feature extractor with semi - residuals to prevent the loss of small - scale features during the feature extraction process; designs a semantic consistency regularization to constrain redundant feature extraction and the feature asymmetry during the feature recovery process. The present invention solves the problem of slow identification speed using traditional threshold methods in mid - low latitude regions.

The invention provides a wind power plant power prediction method based on data distribution fitting, which comprises the following steps of: when a bidirectional GRU neural network model of a weighted triple attention mechanism is constructed, firstly fitting Gaussian mixture distribution of errors through maximum likelihood estimation, embedding an error distribution parameter theta into an error loss function of a PINN neural network model, and then fitting the error distribution parameter theta into the error loss function of the PINN neural network model; a PINN neural network model can learn data rules and error distribution characteristics at the same time, the problem that distribution hypothesis is not matched with actual data characteristics of the wind power plant in a traditional method is solved, and the accuracy of wind power plant power prediction is improved; and then connecting an output layer of the triple attention mechanism model with a bidirectional GRU layer of a PINN neural network, and realizing dynamic feature enhancement of input fan feature data based on triple attention-bidirectional GRU dynamic feature selection. The wind power plant power predicted by using the bidirectional GRU neural network model of the weighted triple attention mechanism constructed in the invention is high in precision, and the problem of model response lag when the wind speed suddenly changes in a traditional mode is solved.

The invention discloses a coal mine disaster early warning method based on multi-source data fusion and a time-space diagram network. The method comprises the following steps: (a) collecting multi-source heterogeneous monitoring data; (b) constructing a geologic-physical dual constraint anisotropic space-time diagram; (c) training the graph by using a geologic-physical anisotropic space-time graph neural network, wherein a graph convolutional layer of the network is designed into a propagation mode which is learned and defined at the same time; and (d) predicting a disaster risk by using the trained model and generating an early warning. According to the method, geological and physical priori knowledge is taken as constraints to be incorporated into the topological structure of the GNN model, so that the model can follow physical laws, disaster propagation under anisotropic geological conditions is predicted, and the accuracy and timeliness of early warning can be improved.

The invention discloses a multi-view self-representation and clustering joint optimization deep clustering method, which comprises the following steps of: firstly, acquiring data samples of a plurality of views, and selecting a most representative sample from each view as an anchor point by adopting a VDA algorithm; constructing and pre-training an auto-encoder network; on this basis, a self-representation module is introduced, shared self-representation and view unique self-representation are learned at the same time, and self-representation of each view is constructed through the similarity between an anchor point and a sample; constructing comprehensive self-representation based on sharing and unique self-representation, constructing a bipartite graph affinity matrix, and obtaining an initial clustering result of samples and anchor points by adopting a bipartite graph clustering algorithm; a clustering result is fed back to the self-representation module, and the representation is iteratively corrected; and finally, sharing and view unique self-representation are fused, comprehensive self-representation is obtained and used for spectral clustering, and a final clustering result is obtained. According to the method, the consistency and diversity characteristics between the views are effectively combined, and the problems that a traditional method is insufficient in structure modeling and low in optimization efficiency are solved.

The invention discloses an Internet of Things intrusion detection method based on a Transform and GNN hybrid model, a model, a training method of the model and electronic equipment, belongs to the technical field of deep learning, and aims at solving the problem that statistical characteristics, time sequence characteristics and topological characteristics of traffic are learned at the same time to utilize network space characteristics for Internet of Things intrusion detection. The method is characterized by comprising the following steps of: carrying out position coding on network flow data in the Internet of Things, inputting the network flow data into an encoder of a Transform model, and outputting a first embedding embedding vector; the network flow is converted into a first graph, the host is a node of the first graph, and the network flow is an edge of the first graph; the first graph is converted into a second graph through a Line Graph model, the network flow is a node of the second graph, and the relation between the host pairs is an edge of the second graph; inputting the second graph into the GNN model, and outputting a second embedding embedding vector; and splicing the first embedding embedding vector and the second embedding embedding vector to obtain a spliced vector, inputting the spliced vector into a classifier, and outputting an Internet of Things intrusion detection result by the classifier, so that topological information can be captured, and relevance between different flows can be captured.

A task prioritized experience replay (TaPER) algorithm enables simultaneous learning of multiple RL tasks off policy. The algorithm can prioritize samples that were part of fixed length episodes that led to the achievement of tasks. This enables the agent to quickly learn task policies by bootstrapping over its early successes. Finally, TaPER can improve performance on all tasks simultaneously, which is a desirable characteristic for multi-task RL. Unlike conventional ER algorithms that are applied to single RL task learning settings or that require rewards to be binary or abundant, or are provided as a parameterized specification of goals, TaPER poses no such restrictions and supports arbitrary reward and task specifications.

A method for learning a shared machine learning model while preserving privacy of individual participants is provided. The method includes: receiving, from each of a group of users, an encrypted user input; when a number of user inputs is greater than or equal to a threshold, transmitting, to each user, a list of the group of users; receiving, from each user, a message indicating a mutual agreement regarding a shared secret among the group; and when a number of received messages indicating the mutual agreement is greater than or equal to the threshold, determining information about the shared machine learning model by combining the received encrypted user inputs. The shared machine learning model facilitates a secure multi-party computation of a function that generates an updated version of the shared machine learning model.

The invention discloses a cross-modal Hash learning method oriented to high-order structure similarity and label correlation, which comprises the following steps of: inputting cross-modal information into a joint learning framework, and learning high-order similarity structure information and label correlation between samples, consistency potential representation and Hash codes at the same time; and finally generating a discrete hash code capable of fully retaining high-order similarity structure information and label correlation between samples. The learning framework uses a graph diffusion-fusion mechanism to adaptively learn the high-order structure similarity between samples, and uses an asymmetric embedding mechanism to migrate the consistent similarity and semantic relationship between the samples to a to-be-learned hash code; and obtaining high-order label correlation through label correlation adaptive learning, integrating the high-order label correlation into a framework, and obtaining high-level semantic information of a multi-modal sample to guide learning of a hash code. The variables of the Hash code learning model are subjected to interactive optimization through an iterative optimization algorithm, and the iterative optimization algorithm adopts an augmented Lagrangian multiplier method to iteratively update the variables to solve the problem.

The invention discloses a method for detecting acetone gas in mixed gas, which specifically comprises the following steps of: (1) periodically acquiring response intensity data of gas to be detected through a sensor array, and preprocessing the response intensity data; and (2) extracting response intensity data in a response stage or a stable stage, inputting the response intensity data into the gas detection model, and carrying out qualitative detection and quantitative detection on acetone gas in the gas to be detected by the gas detection model. According to the method, a model architecture sharing bottom layer features is constructed through multi-task learning, gas qualitative and quantitative tasks are processed at the same time, on one hand, data requirements and labeling cost are reduced, and limited data resources can be utilized more efficiently by sharing a feature extraction layer; and on the other hand, collaborative optimization among tasks is realized through multi-task learning, so that the model can learn the variety characteristics and the concentration characteristics of the gas at the same time, and the overall performance is improved.

The invention provides a general information extraction method based on contrast supervision and cross-stage distillation. The general information extraction method comprises the following steps: selecting a pre-trained model as a base model and performing initialization; initializing a first low-rank decomposition matrix and a second low-rank decomposition matrix for the weight matrix of the base model; training the first low-rank decomposition matrix and the second low-rank decomposition matrix; combining the weight of the base model, the first low-rank decomposition matrix and the second low-rank decomposition matrix to obtain a reasoning model; inputting natural language text information into the inference model; and assigning a task type and an output format through a natural language instruction, and outputting a structured information extraction result. The method has the beneficial effects that the LoRA matrix is trained in stages and the parameters are compounded, so that the target task is learned while the auxiliary task knowledge is reserved, and the loss of cross-stage knowledge migration is remarkably reduced; the natural language instruction drives the same model to process different tasks, and model architectures do not need to be switched.

The invention discloses a building apparent disease detection and evaluation method based on a visual language model, and the method comprises the steps: receiving a high-resolution image and related text description of a building surface through an input module, and converting the high-resolution image and related text description into a format for subsequent processing, so as to guarantee that the model fully obtains visual and language information; the encoder module extracts the multi-scale features of the building image and the semantic features in the text description, thereby enhancing the feature expression capability. The double-branch multi-mode fusion device effectively integrates visual features and text features so as to improve the performance of defect detection. The prior experience learning module optimizes the model performance by storing historical experiences and generating dynamic soft tags. According to the multi-task training mechanism, the comprehensive performance and robustness of the model are improved by learning a plurality of related tasks at the same time, training is carried out in two stages, and effective feature extraction and further optimization are ensured. The efficiency and accuracy of the disease detection method are improved, and the method is suitable for industrial large-scale application and popularization.

The invention provides a micro-service anomaly detection method based on multi-modal feature fusion, and belongs to the technical field of computers. The method comprises the following steps: firstly, performing semantic feature extraction on a log template by combining BERT and a knowledge distillation technology, and learning category features of the log template while enhancing log semantic information; secondly, performing time sequence feature extraction on the log template sequence and the KPIs data by using LSTM and TCN respectively; and then, designing a double-path cross-modal Transform module, respectively extracting collaborative information and redundant shared information between the log and the KPIs, and carrying out weighted fusion through an improved gating unit to obtain a final fusion feature vector. And finally, reconstructing the fusion feature vector by using a diffusion model, calculating a reconstruction error, and comparing the reconstruction error with an abnormal threshold value to realize abnormal detection. Experimental results show that the micro-service anomaly detection method is superior to the existing mainstream method in various indexes of micro-service anomaly detection, and has relatively high detection precision and robustness.

The present application discloses a main melody extraction model training method and component, a singing voice detection method and component in the field of computer technology. The present application uses labeled song audio frames marked with main melody and unlabeled song audio frames without main melody to perform supervised training and semi-supervised training, which not only increases the model training volume, but also enables the model to learn the processing capabilities of labeled data and unlabeled data at the same time, and no additional data labeling work is required. Therefore, the model training efficiency is high and the model generalization ability is also improved. In addition, the scheme can directly process the song audio without the need to separate the background sound and the human voice, thereby improving the training efficiency and the main melody extraction efficiency. Correspondingly, the main melody extraction model training component and the singing voice detection method and component provided by the present application also have the above-mentioned technical effects.

The application discloses an image clustering method based on a composite subspace, and is implemented according to the following steps: step 1, pre-training a convolutional autoencoder; step 2, training a composite subspace image clustering network; and step 3, obtaining a clustering result by applying spectral clustering. The application realizes image clustering in a manner of unsupervised learning, obtains an explicit nonlinear mapping by simultaneously learning self-expression coefficients in an input space and a latent space, embeds input samples into corresponding deep representations, and thus better captures a subspace structure and improves clustering precision.

The invention relates to the technical field of automatic driving, in particular to a multi-client-side automatic driving teaching system which comprises a carla platform and a server side, the carla platform and the server side can be uniformly understood as the server side and both run in the same high-performance server computer, and the client sides run in common computers. The server side and the client side communicate through a network, the client side carries out automatic driving theory teaching and experiment teaching through communication with the server side and resource acquisition, automatic driving teaching can effectively support multi-client-side automatic driving teaching, the cost is low, teaching scenes are diversified, the simultaneous learning requirements of multiple students can be met at the same time, and the teaching efficiency is improved. And the efficiency of automatic driving teaching is improved.

A method for training a student model based on a pre-trained language model (PLM) and a text classification system are disclosed. The method includes: constructing cue-based training samples; adjusting the pre-trained language model using the cue-based training samples to obtain a cue-adjusted teacher model; and training the student model using the processed training samples, wherein during training, the student model simultaneously learns the classification probability vectors output by the cue-adjusted teacher model and the original teacher model. This invention requires the student model to learn from two teacher models simultaneously, thereby alleviating the overfitting problem of the student model in small-sample scenarios by adding a distillation path that learns from the original PLM teacher model with unsupervised data. Furthermore, by transferring the intermediate layer representation of the PLM through knowledge probes and stabilizing the performance of knowledge distillation through comparative learning, the student model can learn higher-order dependencies from the intermediate layer representation of the teacher model, improving the accuracy and efficiency of knowledge distillation.

The application discloses a sintering endpoint intelligent sensing method based on a 3D convolution network and belongs to the field of industrial process soft measurement modeling. The prediction model utilizes a space-time 3D convolution network to simultaneously learn space-time features hidden in data, and adopts an encoding-decoding network to perform multi-step prediction on a present sintering endpoint. Firstly, a random permutation and combination method is used to convert a two-dimensional data structure into a three-dimensional format. Then, a 3D convolution is used to extract time features and space features in the data. Next, a space-time feature calibration module is proposed to learn the importance of different features and finely process the features. Finally, the fine space-time features are input into the encoding-decoding network, and a dynamic multi-step prediction loss function is used to realize multi-step prediction on the sintering endpoint, so that the purpose of intelligent sensing is achieved. Related data collected in a sintering plant in South China verifies the effectiveness and feasibility of the method.

The invention relates to the technical field of information processing, and discloses a technology for acquiring 3D information of a human face and a head from a video. Comprising the steps of learning head 3D information from different appearance videos, estimating a head geometric structure by using a FLAME model, representing modeled head geometry by sharing a deformation field and standardizing spatial nerve voxels, introducing learnable geometry attachment features and an appearance condition density formula, and the like. According to the technology, different appearances of the same object can be learned at the same time, training is carried out in a unified network, a vivid head Avatar image is generated, and controllability is achieved. The method has the advantages that multi-appearance learning, vivid rendering and high controllability are achieved, the method is superior to a reference method in the aspect of image synthesis quality, the model efficiency is high, the method can be well generalized to novel head postures and facial expressions, the accuracy of appearance details is kept, and the method is suitable for meeting the requirements for accurate head three-dimensional data and real-time reconstruction in the fields of medical diagnosis, human-computer interaction and the like.

The present application relates to an online map construction method and system based on inverse perspective mapping and synchronous learning, a perspective image is obtained and preprocessed, a multi-scale inverse perspective mapping image is obtained through transformation, multi-scale strip convolution decoding is performed, different direction strip convolution residual connections are used, layer-by-layer decoding and feature fusion are performed, and global bird's eye view features are obtained; the preprocessed perspective image is subjected to a synchronous learning sequence and a semantic segmentation network to obtain perspective features; the perspective features are converted to a bird's eye view coordinate system, and are spatially aligned with the global bird's eye view features to obtain enhanced perspective features; the global bird's eye view features and the enhanced perspective features are fused through convolution operation; and the fused features are output by a decoder to construct a current frame of a map. In combination with inverse perspective mapping and synchronous learning, the bird's eye view features and the perspective features of multiple frames are fused, the road structure under the bird's eye view is retained, and the perspective image is supplemented with environmental details.

A method of performing target detection during autonomous driving, comprising: performing 3D target detection in a 3D target detection segment; uploading outputs of a plurality of sensors in communication with the 3D target detection segment into a plurality of point clouds; transmitting point cloud data of the plurality of point clouds to a region proposal network (RPN); independently performing 2D target detection in a 2D target detector, the 3D target detection being performed in parallel in the 3D target detection segment; and obtaining a given input image and simultaneously learning bounding box coordinates and class label probabilities in a 2D target detection network that operates to treat target detection as a regression problem.

The invention discloses a semi-supervised image classification method based on anchor points and adaptive graph learning, belongs to the technical field of traditional machine learning, and provides various and reliable data for semi-supervised image classification by adopting data sets of different scales and types of digital images or color pictures; according to the constructed semi-supervised classification model, prediction sample labels can be converted into prediction anchor point labels through anchor points and a linear propagation mode, then the labels of the samples are obtained through linear combination, and the classification speed is effectively increased through the characteristic that the number of the anchor points is far smaller than the number of the samples; an adaptive graph learning method is adopted to take an adaptive graph as an object of iterative optimization, and the adaptive graph and a sample pseudo-label matrix are learned at the same time, so that the influence of an initial adjacent graph on the classification precision is reduced, and the classification precision is improved.

The present invention discloses a method, device, equipment and storage medium for automatic optimization of multiple beams of phased array antennas, the method comprising: constructing a beam optimization model based on sparse coding and deep neural network; optimizing the constructed beam optimization model by using reinforcement learning strategy, so that the beam optimization model can autonomously adjust the beam control strategy according to the feedback of the actual space environment and working state; establishing a multi-task learning framework, so that the beam optimization model can simultaneously learn the mutual influence and coupling effect between multiple beams to achieve a global optimal solution; deploying the trained beam optimization model to a satellite-borne satellite, so that the trained beam optimization model outputs the optimal beam control code according to the working parameters of the current phased array antenna of the satellite-borne satellite, so that the satellite-borne satellite controls the operation of the phased array antenna based on the optimal beam control code. The present invention can adaptively optimize the multiple beams of the phased array antenna in real time.

The application discloses an image classification method based on adaptive local ordinal preserving analytic dictionary learning and belongs to the technical field of computer vision. The method comprises the following steps: performing feature extraction on images in a data set, dividing the images into a training set and a test set, adopting a support adaptive local ordinal preserving analytic dictionary learning model to simultaneously learn an analytic dictionary and a classifier based on the training set, then calculating coding coefficients of the test set based on the learned analytic dictionary, and finally obtaining class labels of the test set through the classifier and the coding coefficients of the test set. The method introduces an adaptive ordinal local preserving term based on a discriminative convolution analytic dictionary learning model, simultaneously preserves neighborhood correlation between dictionary atoms and distance ordering information of atoms in the neighborhood in the learning process, optimizes the analytic dictionary learning model, enhances the discriminativeness of the dictionary, and improves the image classification accuracy of the model in general scenarios such as face recognition, object recognition and scene recognition.

The invention discloses an industrial chain financial risk prediction method based on RGNN-Crossform fusion architecture. The method comprises a data preprocessing module, a relation extraction module, a feature embedding module and a final prediction module. Firstly, aiming at the problems of high noise and serious loss of financial data, a function type data analysis method is used for preprocessing original data, and then an industry chain upstream and downstream enterprise association network is constructed through a data driving method and a text driving method; and then a routing mechanism and a graph neural network are introduced into a feature extraction part of the deep learning prediction model to help the model to learn enterprise risk association structures from different information sources, and finally a financial risk prediction model RGNN-Crossform considering an enterprise association relationship is constructed. Experimental results show that the prediction effect of the constructed RGNN-Crossform model is obviously superior to that of various leading edge financial risk prediction models. Wherein the routing mechanism and the graph neural network module are adopted, so that the model can learn data internal dependence and external correlation characteristics at the same time, and the performance and robustness of the model are improved.