88 results about "Co-training" patented technology

The use of data from multiple data source provides inferred air quality indices with respect to a particular pollutant for multiple areas without the addition of air quality monitor stations to those areas. Labeled air quality index data for a pollutant in a region may be obtained from one or more air quality monitor stations. Spatial features for the region may be extracted from spatially-related data for the region. The spatially-related data may include information on fixed infrastructures in the region. Likewise, temporal features for the region may be extracted from temporally-related data for the region that changes over time. A co-training based learning framework may be further applied to co-train a spatial classifier and a temporal classifier based at least on the labeled air quality index data, the spatial features for the region, and the temporal features for the region.

A semi-supervised polarization SAR terrain classification method based on full convolution GAN is implemented by the following steps: performing refined Lee filtering; inputting polarization SAR data;constructing a generation network G; constructing a discrimination network D; training and testing the networks and calculating classification accuracy rate. Compared with the semi-supervised Co-training classification method in the prior art, the invention improves the classification accuracy of polarization SAR data of the polarized synthetic aperture radar, solves the problem of insufficient utilization of polarization information, enhances the robustness of the model to the input data, and has the effect of end-to-end classification. The invention can be applied to the terrain classification of polarized synthetic aperture radar polarized SAR data.

Systems and methods are described that facilitate categorizing a group of linked web pages. A plurality of web pages each contains at least one link to another page within the group. A feature analyzer evaluates features associated with the one or more web pages to identify content, layout, links and / or metadata associated with the one or more web pages and identifies features that are labeled and features that are unlabeled. A graphing component creates a vector associated with each web page feature wherein vectors for unlabeled features are determined by their graphical proximity to features that are labeled. A co-training component receives the graph of vectors from the graphing component and leverages the disparate web page features to categorize each aspect of each feature of the page. A page categorizer receives aspect categorization information from the co-training component and categorizes the web page based at least upon this information.

The invention discloses a semi-supervised medical image segmentation method based on adversarial collaborative training. A neural network segmentation model is trained by using a small amount of labeled medical image data and a large amount of unlabeled medical image data, so that the model performance is improved. The model uses two decoder branches with different structures, the two decoder branches share the same encoder, and the two decoder branches can learn each other through a cooperative training method. Meanwhile, the model also uses an adversarial learning method to train a discriminator, and the discriminator can learn the high-order continuity between the segmentation result and the real label, so that the output of the segmentation network is closer to the real label visually.And meanwhile, the discriminator can also select the part with higher confidence in the unlabeled data pseudo tags to train the segmentation model. The method provided by the invention is not limitedby diseases and focus types, can be used for medical image segmentation of diseases of various parts such as the liver and the oral cavity, and has very good universality and universality.

The present invention provides a multi-category multi-view data generation method for generating a confrontation network based on the depth convolution. The method comprises the following steps: 1. carrying out picture center cutting; 2. overlaying multiple views in the channel dimension; 3. extracting multi-view category tags; 4. co-training the DC-GAN network by using the overlaid multiple views, category tags, and high-dimensional random noise; 5.introducing the high-dimensional random noise and custom tags into the trained network, and generating multi-view overlaid data; and 6. cutting and filling the background to obtaining the multiple views satisfying the original size. According to the method provided by the present invention, the method for generating the confrontation network bymulti-view overlaying and tag training realizes the function that the multi-category multi-view can be generated through a model only by modifying the input, and the generated data can be used as extension of trained data to increase the diversity of trained data.

The present invention relates to a semi-supervised biomedicine event extraction method based on co-training. Automatic extraction of biomedicine events greatly interests people with rapid increasing of biomedicine literatures. The scale of the marked biomedicine event corpus is small to influence the performances of the classification algorithm and even cause overfitting. The method provided by the invention identifies more accurate positive instances from unmarked data to enlarge a marked training set. The method comprises the steps of: designing abundant features for usage of an SVM; learning short sentences based on word embedding from Word2vec and Pubmed; further extending the short sentences to dependent short sentences between triggering words and parameters, and inputting the dependent short sentences into a CNN; and finally, performing backfill of samples, meeting conditions, predicted by the SVM and the CNN in the unmarked corpus into the training set, incrementally extendingthe training set. Lots of experiment results show that the new semi-supervised biomedicine event extraction method can effectively extract events.

The invention discloses a semi-supervised text sentiment classification method based on random feature subspace. The method comprises steps as follows: 1 acquiring and pre-processing comment text data and constructing a global feature set; 2 expressing all comment texts in a vector form; 3 marking part of comment texts and acquiring a marked sample set and an unmarked sample set; 4 calculating feature weight of all feature words in the global feature set; 5 constructing random subspace; 6 performing cooperative training with unmarked samples to obtain Z classifiers finally; 7 integrating the Z classifiers in a main voting manner to obtain a final integrated classifier. The method solves the problems that a large number of misclassified samples exist in the training process with traditional cooperative training algorithms and the difference between classifiers is small in semi-supervised text sentiment classification methods, so that the accuracy of the text sentiment classification method is improved.

The invention discloses an orthogonal frequency division multiplexing (OFDM) received frame synchronizing method for co-training sequence mutual-correlation information and is applicable to all OFDM systems adding training sequences in front of data frames. The OFDM received frame synchronizing method specifically includes 1), designing a frame structure of transmitting end; 2), performing mutual-correlation calculation and energy calculation synchronously by means of delay mutual-correlation algorithm at a receiving end, and obtaining regular measuring results; and 3) selecting energy threshold value and regular measuring threshold value of an energy window, and judging synchronous time. By accumulating a plurality of segment correlation values and uniformizing, operation load is reduced. The problems about peak platform and a plurality of high and side peak values are solved, and synchronization accuracy is increased. Channel influence is eliminated by receiving segmental mutual correlation of the sequence, and high frequency offset can be resisted. Dynamic range of receiving data is enlarged after uniformization, and selection of regular judgment time threshold value and synchronous detection can be simply and accurately realized. Meanwhile, adaptability is improved, frame synchronization in various channels can be completed accurately.

The invention discloses a false comment identification method based on rolling type cooperative training, and relates to the technical field of false comment identification. The method comprises the steps: obtaining a comment text, and determining the content features of the comment text according to the comment text; obtaining reviewer information corresponding to the comment text, and determining reviewer behavior characteristics according to the reviewer information; performing false comment identification according to the comment text content characteristics and the reviewer behavior characteristics; outputting false comment recognition results. According to the false comment recognition method, unlabeled samples are effectively utilized to assist model learning, multiple features suchas emotion and text representation are fused at the same time, the recognition performance of the model is improved through multi-feature fusion collaborative training, and the accuracy is improved by 3.5% compared with a traditional false comment recognition method.

The invention discloses a junk comment identification method based on collaborative training. Junk comments are classified into explicit junk comments and implicit junk comments, the explicit junk comments are screened out by adopting a rule-based method, identification training is conducted on one of the implicit junk comments by adopting two classifiers AdaBoost and the SVM based on an automatic identification method, and finally whether the comment is a junk comment or not is further judged through Co-Training. Therefore, the classification accuracy is improved, and meanwhile the classification efficiency of a junk comment classification method is also ensured.

The invention discloses a text abstract generation method and device, equipment and a medium, the method belongs to the field of computer vision, and the method comprises the following steps: obtaining text content; inputting the text content into a coding layer to obtain a hidden layer embedding vector of the text content; wherein the coding layer is obtained based on collaborative training of anextraction type abstract generation model and a generation type abstract generation model; and inputting the hidden layer embedding vector into the extraction type abstract generation model or the generation type abstract generation model for processing, and outputting the text abstract. According to the method and the device, the respective advantages of the extraction type abstract generation mode and the generation type abstract generation mode can be integrated, so that the finally output text abstract can better summarize the content characteristics of the text content.

The invention discloses a geographic information service metadata text multi-level multi-label classification method, which comprises the following steps: 1) obtaining a geographic information servicemetadata text set to perform text preprocessing, and dividing each data sample into text feature word combinations; 2) setting a primary classification directory, and generating a typical word list semantically associated with classification categories; 3) screening the text feature words according to the typical word list; 4) selecting ML-KNN as a base model for collaborative training; 5) establishing a topic prediction model ML-CSW as another base model for collaborative training; 6) designing a collaboration mechanism, matching a multi-label topic for the metadata text, and taking the multi-label topic as a primary coarse-grained topic classification result. According to the method, domain features and text semantics of the geographic information service metadata are considered, only asmall number of marked data samples are depended on, and the classification result is better in overall performance compared with a traditional multi-label classification method.

The invention provides a multi-task interaction enhanced electronic text event extraction method, which is a method for extracting an event from an electronic text, combines three tasks of an entity recognition task, a trigger word recognition task and argument role prediction, and additionally introduces an auxiliary event sentence recognition task. Therefore, the attention of the extractor is more focused on the event sentence. According to the method, two sets of parameters are adopted for modeling an entity recognition task and a trigger word recognition task respectively, and informationinteraction is added between sub-tasks, so that a plurality of sub-tasks can be cooperatively trained to realize excellent performance. According to the combined event extraction method, a neural network technology is adopted, a more direct, complete and end-to-end event extraction model is formed, and the effect of an event extraction task is greatly improved.

The invention discloses an entity recognition method and device, electronic equipment and a storage medium, relates to the field of natural language processing, and solves the problem that entity recognition needs to label corpus samples on a large scale, and the method comprises the following steps: learning a text to be labeled based on a BERT model to obtain a text vector; carrying out preliminary training on each neural network model by utilizing the labeled text; respectively obtaining N-1 groups of entity label sequences of each unlabeled text based on other N-1 preliminarily trained neural network models in the model set, training each preliminarily trained neural network model mi based on each unlabeled text and the N-1 groups of entity label sequences of each unlabeled text to obtain a cooperatively trained neural network model Mi; calculating the text vector based on a plurality of cooperatively trained neural network models and CRF models to obtain a plurality of candidate labeling sequences; and selecting a group of candidate annotation sequences as annotation results of the to-be-annotated text based on voting rules.

The invention discloses an image scene classification method and system combined with semi-supervised clustering, and the method comprises the steps of redefining an objective function of semi-supervised Kmeans through employing a labeled sample, and supplementing and defining an objective function of SVM, and obtaining semi-supervised Kmeans clustering and a base learning device based on SVM classification; enabling the two base learners to carry out cooperative training, and forming a selection and iterative training scheme of a pseudo label sample; and finally, according to the confidence coefficient, fusing results of the two learners to obtain a scene image category to which the sample belongs. According to the invention, different types of methods in the image scene classification field are used to construct a base classifier and carry out cooperative training. Meanwhile, a pseudo label sample is introduced to expand a training set, so that the problem of insufficient label samples is effectively solved. Furthermore, clustering is carried out on the label-free samples to obtain the distribution characteristics of the label-free samples, and the concept drift problem is solved. Finally, the labeling cost of the scene image is reduced, concept drift is solved, and the image scene classification accuracy is improved.

The invention discloses a social network cross-platform malicious user detection method based on longitudinal federated learning. The method comprises the following steps: step 1, constructing a social network cross-platform malicious user detection hierarchical architecture based on longitudinal federated learning; 2, dividing participants into active parties and passive parties, and performing preprocessing operation on sample data of the active parties and the passive parties in a data preprocessing layer to obtain structured data; step 3, mapping common sample data of the active party and the passive party by the structured data processed by the data preprocessing layer; 4, cooperatively training a global model under the definition of machine learning, and encrypting and decrypting data of an active party and a passive party by using homomorphic encryption to complete federal learning layer training; step 5, enabling the active party and the passive party to update own local model training parameters and output prediction results; and step 6, transmitting a prediction result obtained by the federal learning layer back to each participant in the data application layer, and realizing a high-quality malicious user detection effect.

At present, a decoder-encoder structure is commonly used in neural machine translation, and a good effect is obtained under the condition that parallel corpora are sufficient. However, for the Mongolian language as a small language, mongolian and Chinese parallel corpus resources are limited and are very difficult to obtain, therefore, the invention provides a semi-supervised Mongolian-Chinese neural machine translation method based on collaborative training. Three translation models are constructed by using a semi-supervised classification generative adversarial network: a Mongolian-Chinese translation model M-mo-ch, an English-Chinese translation model M-en-ch and a Korean-Chinese translation model M-ko-ch; the three translation models are used for marking a multi-source end parallel corpus Mongolian and Korean to a target end, namely Chinese, the marked corpus with the best quality is selected by using a language model LM-ch trained by Chinese monolingual training to expand an original corpus, and a better translation model is trained again. According to the method, collaborative training and the semi-supervised classification generative adversarial network are combined and applied to Mongolian-Chinese neural machine translation, and the quality of the Mongolian-Chinese neural machine translation model is improved.

Collaborative bootstrapping with uncertainty reduction for increased classifier performance. One classifier selects a portion of data that is uncertain with respect to the classifier and a second classifier labels the portion. Uncertainty reduction includes parallel processing where the second classifier also selects an uncertain portion for the first classifier to label. Uncertainty reduction can be incorporated into existing or new co-training or bootstrapping, including bilingual bootstrapping.

The invention discloses a semi-supervised collaborative training analog circuit early fault diagnosis method, which can classify and identify analog circuit faults without using a large amount of labeled data so as to realize early fault diagnosis. The method comprises the following steps: firstly, carrying out Fourier transform on a known analog circuit normal signal and a fault signal; carryingout spectrum partitioning, empirical wavelet transform on the frequency spectrum; obtaining AM-FM components, carrying out hilbert transformation on the data; obtaining a plurality of EWF components;superposing the energy of the plurality of EWF components; finally, through a semi-supervised collaborative training algorithm, utilizing a small amount of labeled data fully for learning and utilizing a large amount of unlabeled data fully, thus the accuracy of marking actions is guaranteed to a certain extent, a large amount of marking cost is saved, and therefore early fault diagnosis of theanalog circuit based on semi-supervised collaborative training is achieved.

The invention discloses a CO-training method based on unlabeled-sample consistency determination, relates to CO-training and multi-angle image classification, and aims to solve the problem that existing multi-angle remote sensing increases difficulty of joint analysis on ground objects of the same area, especially difficulty of change analysis of the ground objects, and enables classification accuracy of multi-angle remote sensing images to be low. A process includes: 1, carrying out initial classification: 2, selecting trusted samples of unlabeled samples in shot images; 3, obtaining a retrained classifier of a current view angle until retraining of classifiers corresponding to all view angles is completed; 4, obtaining a classification result of the current view angle until reclassification of the classifiers corresponding to all the view angles is completed; 5, repeatedly executing 2, 3 and 4 until an iteration termination condition is met, obtaining a classification result of eachview angle, carrying out voting, and using labels of highest voting rates as labels of a sample which is in the shot images in 1 and has no label. The method is used in the field of digital image processing.

The invention provides an entity alignment method and device in a knowledge graph, and the method comprises the steps: obtaining a plurality of entities from a plurality of platforms as an entity training set; according to the available information related to each entity in the entity training set, generating features used for cooperative training wherein the features are used for indicating the similarity between the same type of available information in the plurality of entities; training a model based on cooperative training according to the features, and judging whether the entity pairs tobe processed are synonymous or not according to the model obtained through training. Through the method and the device, the problem of poor alignment due to the fact that a knowledge representation learning-based method depends on a large number of annotation data entities in related technologies is solved.

The invention discloses a collaborative training method based on domain self-adaptation, which comprises the following steps: training one party by utilizing high confidence output of the other party,and processing candidate regions with low confidence output of the two parties by utilizing a maximized classifier difference method; besides, in the aspect of feature alignment of the backbone network, the output of the RPN is used for calculating the foreground probability of each point on the feature map, and a larger weight is given to an area with a larger foreground probability during feature alignment; the target detection capability of the model in the label-free field is improved, the requirement of the target detection model for annotation data is reduced, and dependence on human resources is reduced.

The invention discloses a text classification auxiliary labeling method based on collaborative training, and relates to the technical field of text classification. The text classification auxiliary labeling method comprises the following steps: labeling sample data; constructing a data set; training two classifiers; classifying and recording classification results; extracting correct features and error features of the classifiers; performing sample optimization; retraining the replacement data set until the accuracy reaches a confidence threshold; and meanwhile, performing classification by using the two classifiers, and outputting a result under the condition that classification results are the same. According to the text classification auxiliary labeling method, most of simple data can be automatically labeled with high quality, so that the labeling accuracy and efficiency are greatly improved.

The invention discloses a planar inverted F-shaped antenna resonant frequency prediction method based on semi-supervised learning. The method comprises the following steps: establishing a mapping relationship between four related parameters of the width of a short-circuit metal sheet, the length of a radiation metal sheet, the width of the radiation metal sheet and the height of the radiation metal sheet of the planar inverted F-shaped antenna and an actually measured resonant frequency by using a Gaussian process and a support vector machine; carrying out iterative training by utilizing a cooperative training method of a Gaussian process and a support vector machine in combination with unmarked data, wherein the trained semi-supervised cooperative training model can be used for predictingresonant frequencies of other planar inverted F-shaped antennas. According to the method, the problems that in existing electromagnetic optimization design, more marking samples are needed during model training, electromagnetic simulation software HFSS needs to be called for multiple times, the calculation cost is high, and consumed time is long can be solved; compared with a modeling mode basedon traditional supervised learning, the resonant frequency prediction capability of the method has certain advantages.

The invention discloses a track-variable weight losing rehabilitation robot. The rack-variable weight losing rehabilitation robot comprises a rack, an upper limb training mechanism and a lower limb training mechanism; in the upper limb training mechanism, the upper ends of second connecting rods and support rods form revolute pairs, the lower ends of the second connecting rods and the head ends of third connecting rods form revolute pairs, and first connecting rods of which the top ends are provided with handles are inserted into the second connecting rods to form moving pairs and revolute pairs; in the lower limb training mechanism, pedals are installed on pedal plates, the front ends of pedal plates are connected with the front ends of cranks to form revolute pairs, the back ends of the cranks are vertically and fixedly connected with an output shaft of a power mechanism, the middles of the pedal plates are connected with the tail ends of the third connecting rods to form revolute pairs, the back ends of the pedal plates are connected with a pedal plate support frame to form moving pairs and revolute pairs, the head end of the pedal plate support frame is connected with the upper end of a linear actuator to form a revolute pair, the tail end of the pedal plate support frame is connected to the rack to form a revolute pair, and the lower end of the linear actuator is hinged to the rack. According to the track-variable weight losing rehabilitation robot, co-training of the upper limbs and the lower limbs can be achieved, and tracks during training can be adjusted.

The invention discloses a federal transfer learning-based model training method and a computational node, which can be applied to the field of artificial intelligence, and the method comprises the following steps: training a model parameter G of a feature extractor and a model parameter T of a subtask model (such as a classifier) on each source domain through local labeled data, sending all G to a target domain, training a model parameter D1 of a domain discriminator on each source domain, training a model parameter D2 of the domain discriminator on the target domain, aggregating all D1 and D2 at a server side or a target domain side to obtain an aggregation parameter value D, and sending D to each source domain, and performing repeated iterative adversarial training on each source domain through the respective feature extractor and discriminator. According to the method, domain alignment is achieved through the adversarial training process, only model parameter values are transmitted among the domains, data or data features are not transmitted, data privacy is protected, and collaborative training of the models is achieved under the condition that domain alignment and data privacy are considered.

The invention provides a shoe sole pattern image clustering method based on a collaborative network structure. The system comprises a shoe sample or suspect sole pattern learning network connected with category information based on characteristic layer difference loss and an on-site sole pattern unsupervised clustering network. Pre-training is carried out through data sets with different attributes, feature subspace provided by supervised learning of shoe pattern or suspect sole pattern mass data can be utilized to restrict clustering learning of on-site sole patterns or unmarked suspect pattern features, and therefore the clustering process is based on the evidence. Besides, a pre-training strategy of supervised and unsupervised network cooperative training is provided based on the sequence of the training models so that the cooperative effect between the networks can be more effectively reflected and the clustering precision of the sole pattern image can be enhanced.

The invention relates to an integrated collaborative training method and device for zero sample classification and terminal equipment, and the method comprises the steps: dividing an obtained data setinto a training set and a test set, calling the training set and the test set as a visible class and an invisible class, training attribute prediction networks of different structures, selecting twonetworks as a main network and an auxiliary network, calculating attribute mapping parameters, synthesizing virtual features of invisible classes according to the attribute mapping parameters, combining the virtual features with a plurality of classifiers to complete training of the classifiers, extracting the features of the invisible classes by using a main network and an auxiliary network, predicting the features of the invisible classes by using the classifiers, and endowing the invisible classes meeting conditions with pseudo tags according to a classifier voting mechanism; adding invisible classes endowed with pseudo tags into a training set to train the attribute prediction network again so that prediction precision of a network model is improved when different ZSL embedding methodscan be used for training to select a main network and an auxiliary network. The invention easily expands to other zero sample learning methods and performance is improved.

The invention provides a text information extraction method and related equipment. The method comprises the steps of obtaining a first language marking corpus set, a first language unmarked corpus set, a second language marking corpus set and a second language unmarked corpus set according to a first language corpus text and a second language corpus text; cooperatively training a first language classifier and a second language classifier by utilizing each corpus set; utilizing a first language classifier to classify a first language target entity pair obtained according to the mixed statement,and classifying the second language target entity pair obtained according to the mixed statement by utilizing a second language classifier, and obtaining an entity relationship of the mixed entity pair of the mixed statement according to a classification result of the first language target entity pair and the second language entity pair. According to the method, the entity relationship can be accurately extracted from the texts using two different languages.

The invention discloses a continuous sign language recognition system based on multi-language collaboration. The system uses a public visual feature encoder to extract feature expressions, and uses different time sequence modeling networks (i.e., target sequence models) to learn corresponding sign language features for sign languages of different languages. A shared time sequence encoder (namely ashared sequence model) is used for expressing the same visual mode among different sign languages, initialization is carried out through language embedding vectors, multilingual sign language recognition under a single framework is achieved through multilingual collaborative training, the visual commonality among different sign languages is fully mined, and the sign language recognition performance is improved.