31results about How to "Reduce the amount of labeling" patented technology

The invention provides a neural network model training method and device, electronic equipment and a computer readable storage medium. The neural network model training method of the neural network model comprises the following steps: executing initial training by utilizing a first training sample set to obtain an initial neural network model; performing prediction on the second training sample set by utilizing the initial neural network model to obtain a prediction result of each training sample in the second training sample set; determining a plurality of preferred samples from the second training sample set based on the prediction result; receiving a labeling result for the plurality of preferred samples, and adding the labeled plurality of preferred samples into a first training sample set to obtain an expanded first training sample set; performing update training by using the extended first training sample set to obtain an updated neural network model; under the condition that a training ending condition is met, ending the training; and repeating the prediction step, the preferred sample determination step, the sample expansion step and the training updating step under the condition that the training ending condition is not met.

The invention provides a method for training a convolution neural network through a small-scale face data set. The method is characterized by comprising the following steps: step one, using an originally marked face sample set to train a VGG face recognition model of the convolution neural network; step two, constructing a deep convolution to generate a DCGAN model of a confrontation network, andusing the originally marked face sample set to train the deep convolution so as to generate the confrontation network; step three, generating an unlabelled face sample set through DCGAN; step four, generating a face data set mark for the DCGAN; step five, using the originally marked face sample set to train a plug-and-play generated network PPGN; step six, generating a labeled face sample set through the PPGN; step seven, training the convolution neural network in combination with a DCGAN and PPGN generated sample set and the originally marked sample set; step eight, repeatedly training, namely repeating the step four, step five, step six and step seven repeatedly; and step nine, using the originally marked face sample set to finely adjust the VGG network.

The invention discloses a speech recognition method based on a domain-invariant feature, and the method applies a speech domain-invariant feature extraction model to an end-to-end speech recognition model. The feature extraction model used in the speech recognition method aims at the robustness problem, and by adding more types of speech data to train the speech feature extraction model, better parameters can be obtained and a better domain-invariant feature extraction model can be obtained. The speech recognition method based on the domain-invariant feature uses unlabeled pure speech data totrain the feature extraction model, uses a small number of speeches with text annotation to train an end-to-end acoustic model, and provides important technical support for improving the robustness ofthe end-to-end acoustic model. Compared with the prior art, the speech recognition method based on the domain-invariant feature has higher recognition accuracy in different noise environments, smaller task quantity of speech annotation tasks, and faster training and testing speed of the models.

The invention relates to a mammary gland electronic medical record entity recognition system based on multi-standard active learning, and the system is characterized in that the system comprises a preprocessing module; an entity identification module; and an active learning module. According to the invention, the active learning selection strategy for text sequence annotation is designed by considering three aspects of annotation data volume, sentence annotation cost and data sampling balance, so the total annotation workload is reduced. On the one hand, the system can be used for constructingsystems such as breast disease risk patient identification marks, disease medicine recommendation and auxiliary decision diagnosis, doctors are helped to improve the execution efficiency of breast disease standardized diagnosis and treatment, and scientific bases and suggested schemes are provided; on the other hand, doctors can be assisted in finding out potential abnormal conditions in the diagnosis and treatment process, the misdiagnosis and missed diagnosis rate is reduced, the curing probability of breast disease patients is increased, and important value is achieved for intelligent development of breast disease research.

The invention discloses an active learning method and device. The method comprises the following steps: obtaining labeled sample data; training the labeled sample data by using a machine learning model to obtain an intermediate model; inputting test data into the intermediate model for prediction to obtain a prediction result corresponding to the test data; outputting the test data corresponding to the prediction result of the fuzzy boundary as to-be-labeled sample data according to a preset evaluation index; wherein the machine learning model is trained in a multi-thread parallel mode, and the intermediate model is used for prediction. The invention provides a machine learning platform with strong universality for a user. According to the method, the labeling amount of sample data used for training of the input model can be reduced, the labor cost of manual labeling is reduced, and the processing efficiency is effectively improved. And the effect of quickly iteratively optimizing themodel at the minimum labeled data cost is achieved. And the minimum consumption of processing resources is realized.

The invention provides a semi-automatic labeling method for image data. The semi-automatic labeling method comprises the following steps: carrying out partial image-level labeling on an unlabeled image; putting the annotation data into a collaborative weak supervision recognition model for training; reconstructing the collaborative weak supervision recognition model to obtain a strong supervisionreconstruction model; detecting an unlabeled image by using the strong supervision reconstruction model to obtain a detection result; and training a strong supervision reconstruction model by using the manually labeled image data. A certain amount of weak label data is used in the early stage, and then the model is gradually improved in the later stage in an active learning mode, so that the precision is ensured while the labeling amount is small.

The invention discloses a data annotation processing method, device and system. The method comprises the following steps: acquiring a blood sample image, and preprocessing the blood sample image to obtain a leukocyte image; classifying the leukocyte images by using a neural network model to obtain a leukocyte classification result; obtaining a judgment result of whether the classification result is correct, and if the judgment result indicates that the classification result is correct, storing the leukocyte image and the classification result; and if the judgment result indicates that the classification result is incorrect, obtaining a correct annotation result of the leukocyte image, the correct annotation result being obtained by manually classifying and annotating the leukocyte image bythe target user at the client. The technical problems of high labor cost and time cost, low labeling efficiency and high error rate due to the fact that current blood cell image labeling needs professional medical knowledge and experience of labeling personnel are solved.

The invention provides a text sentiment classification method and device. A text coding model belonging to the same field as a to-be-classified text is used for analyzing context semantic information in the to-be-classified text in an omni-directional mode to obtain a corresponding text vector; meanwhile, a position vector of the sentiment classification attribute word is acquired, and then the text vector and the position vector are spliced to obtain a text and a position vector. The text and the position vector contain context semantic information of to-be-classified text and position information of the emotion classification attribute words, and the target emotion classification model can define an emotion analysis object according to the position information of the emotion classification attribute words, so that the accuracy of an emotion analysis result is improved. Moreover, the text vector can more accurately represent the semantic information of the text to be classified, so that the emotion classification model can better understand the context information of the text, and the classification accuracy is further improved. In addition, the data annotation quantity can be greatly reduced by adopting the pre-trained text coding model.

The invention provides an image annotation method and device, an image semantic segmentation method and device, and a model training method and device, relates to the technical field of artificial intelligence, and is used for improving the sample image annotation efficiency. According to the image annotation method, the edge pixel points in the sample image are detected, the target image blocks in the image blocks in the sample image are screened according to the edge pixel points, and the target image blocks are annotated, so that the annotation result of the sample image is obtained, and due to the fact that all the pixel points in the sample image do not need to be annotated, the annotation quantity in the sample annotation process can be relatively reduced, and the efficiency of annotating the sample image is improved; and as the image has certain redundant information, the accuracy of the image semantic segmentation model is not influenced even if all pixel points in the sample image are not annotated and the image semantic segmentation model is trained.

The invention relates to an image processing method, an image processing device and a storage medium. The method comprises the steps of obtaining a to-be-processed image; selecting at least one category based on category information output by the to-be-processed image through the image classifier, and determining a thermodynamic diagram of each category in the at least one category based on the category information; for each category of thermodynamic diagrams in the at least one category of thermodynamic diagrams, determining a first positioning frame set corresponding to the target object inthe to-be-processed image; determining a second positioning frame set of the to-be-processed image according to an unsupervised target detection algorithm; and determining a target positioning frame set in the to-be-processed image according to the first positioning frame set and the second positioning frame set, the target positioning frame being used for representing the position of the target object in the to-be-processed image. Through the technical scheme of the invention, the target detection algorithm and the deep learning algorithm are combined, the data acquisition difficulty is low,the data annotation amount is small, and the position of the target object in the to-be-processed image can be quickly and accurately determined.

The embodiment of the invention provides a multi-task attribute image recognition method and device, electronic equipment and a storage medium. The method comprises: obtaining a to-be-identified image needing to be input into the target identification network, wherein the target recognition network is obtained by training a multi-task attribute recognition network through sample data obtained through an active learning method, and the target recognition network comprises a shared network used for extracting public image features and a plurality of task networks used for extracting task imagefeatures; inputting the to-be-identified image into the shared network for image feature extraction to obtain public image features of the to-be-identified image; inputting the public image features into the task network for task feature extraction to obtain task image features of the to-be-identified image; and performing task result classification based on the task image features to obtain an attribute identification result. Time and calculation expenditure are saved, the model operation speed is increased, and the cost-to-effect ratio of multi-attribute identification is reduced.

The embodiment of the invention discloses a method and device for identifying data exception, a server and a medium. A specific embodiment of the method comprises: obtaining a target data sequence in a preset time period; determining a predicted value corresponding to the target data sequence; extracting a data feature index based on the target data sequence and the predicted value; and inputting the data feature indexes into a pre-trained autonomous learning model, and generating prompt information used for representing whether data exception exists or not. According to the embodiment, the workload of personnel is reduced, and the accuracy of data exception identification is improved.

The embodiment of the invention provides a method and system for detecting the abnormal state of a railway overhead line system bolt. The method comprises: acquiring an image of a to-be-detected railway contact network bolt; and based on the trained image generation model and the multiple spatial mapping model, carrying out anomaly detection analysis on the to-be-detected railway contact network bolt image to obtain a railway contact network bolt abnormal state detection result. According to the embodiment of the invention, based on an unsupervised learning algorithm, the problems of poor generalization, difficulty in defining abnormal categories and great reduction of accuracy of unseen abnormal samples of an existing detection algorithm are solved, and the unsupervised learning algorithmdoes not need to define the abnormal categories in detail, so that the data annotation amount is greatly reduced, and the working intensity of detection personnel is reduced.

The invention discloses a model building method and system, a paragraph label obtaining method and a medium, and relates to the field of natural language processing transfer learning, and the method comprises the steps: collecting all judgment document data from a database, and obtaining pre-training data; defining paragraph labels of different types of judgment documents; marking paragraph labels of different types of judgment documents to obtain training data; constructing a judgment document structured model; pre-training the model; training a pre-trained judgment document structured model by utilizing the training data; and debugging the trained judgment document structured model to obtain a final judgment document structured model, wherein the input of the judgment document structured model is judgment document text data, a task prefix is added to a paragraph of the input judgment document, and the output of the judgment document structured model is paragraph label text data of the judgment document. The model established by adopting the method can predict any type of judgment document paragraph label after being trained.

The present invention discloses a self-learning face verification method. The method comprises the steps of: the step 101: employing a marked face sample set L to train a convolutional neural network;the step 102: employing the convolutional neural network to perform marking of an unmarked sample set U; the step 103: employing the sample set U to perform fine tuning of the convolutional neural network; the step 104: employing the sample set L to perform fine tuning of the convolutional neural network; and the step 105: repeatedly performing the steps 102, 103 and 104 for many times.

The invention discloses a repeated incoming call preprocessing method, device and equipment based on 95598 and a storage medium. The processing method comprises the following steps: S101, data acquisition; s102, removing business data from the obtained data to obtain screened data; s103, labeling and primary screening are carried out; s104, analyzing the primarily screened data by using text similarity, judging whether the primarily screened data are the same appeal event or not, and if so, inputting the text content of the repeated incoming call corresponding to the primarily selected data into an algorithm unit; if not, sending the primarily selected data to a processing end; s105, the algorithm unit judges whether the received text content is similar or not, if yes, it is considered that the corresponding appeal event is a repeated call, and the repeated call is output; and if not, sending the corresponding text content to the processing end in the step S104. According to the method, the neural network data annotation amount is reduced, the training period is shortened, the text analysis accuracy can be improved, manpower is reduced, and the early-stage preparation time is shortened.

The invention discloses a labeled sample determination method. The labeled sample determination method comprises the steps of obtaining a pre-trained classification model and a classification target;repeating the following steps to iteratively update the classification model until a preset stop condition is met, and taking the corresponding sample set when the preset stop condition is met as a to-be-labeled sample set; predicting samples in a sample set by utilizing the classification model to obtain a classification score of each sample belonging to each classification target; performing fusion sorting on the classification score of each sample belonging to each classification target to obtain a plurality of fusion sorting results; determining a to-be-labeled sample set from the plurality of fusion sorting results; and updating the classification model by utilizing the to-be-labeled sample set. According to the method, the expert annotation amount required by model training can be remarkably reduced, the labor cost is saved, the benefit of unit annotation is improved, the model is quickly iterated, and the method is different from a single-strategy active learning scheme, so thatthe problem that high-weight samples generated by fusion sorting of a single strategy are omitted is effectively solved.

The embodiment of the invention provides a neural language network model training method and device, equipment and a storage medium, which are used for reducing the labeling amount of training sample data and improving the training efficiency of a language model. The method comprises the following steps: acquiring training sample data; cyclically executing the following steps until the neural language network model obtained by training meets a preset requirement: predicting the unlabeled training sample data by using the neural language network model obtained by previous training, and determining an identification probability for representing that each training sample data is identified; according to a preset selection strategy, on the basis of the recognition probability of each training sample data, selecting a part of training sample data requests from the training sample data which are not labeled for manual labeling; and obtaining the manually labeled training sample data, and training the neural language network model obtained by the previous training based on the manually labeled training sample data to obtain a new neural language network model.

The invention discloses a disease lung CT segmentation and quantitative analysis method and system based on deep learning, and belongs to the technical field of pathological diagnosis, and the method comprises the following steps: segmenting an effective region of a CT lung scanning image through a deep segmentation model, and taking the effective region as a segmentation mask; carrying out statistics on Hu value distribution of the segmentation mask, and calculating the effective inflation volume of the lung according to the gas percentage of the inflated tissue under different Hu values; and extracting a segmented mask image, and calculating the lung density by using a deep adversarial network model. By means of the deep segmentation model, the labeling amount and labeling difficulty of data are greatly reduced, the lung inflation amount and the lung weight are calculated on the basis after the effective area of the lung is segmented, the general situation and development of the illness state of a patient can be observed in a multi-dimensional mode through the indexes, and a doctor can be assisted in early screening and diagnosis of the illness.

The invention provides a method for training a convolutional neural network through a small-scale human face data set, which is characterized in that it comprises steps: Step 1: using the original marked human face sample set to train a convolutional neural network VGG face recognition model; Step 2: Construct a deep convolutional generation confrontation network DCGAN model, and use the original labeled face sample set to train a deep convolutional generation confrontation network; Step 3: Generate an unlabeled face sample set through DCGAN; Step 4: Generate a human face with DCGAN Face dataset annotation; Step 5: Use the original labeled face sample set to train the plug-and-play generation network PPGN; Step 6: Generate a labeled face sample set through PPGN; Step 7: Combine the samples generated by DCGAN and PPGN Set and the original labeled sample set to train the convolutional neural network; Step 8: Repeat the training, that is, repeat steps 4, 5, 6, and 7 multiple times; Step 9, use the original labeled face sample set to fine-tune the VGG network.

The invention provides a text entity relationship recognition method, a text entity relationship recognition system and a medium. The text entity relationship recognition method comprises the steps of: generating a support data set based on a reference text of a target domain, wherein the support data set comprises positive example entity pairs belonging to X preset specified relationship types and negative example entity pairs not belonging to the X preset specified relationship types; generating a query data set comprising at least one entity pair in a to-be-tested text of the target domain;and based on the support data set and the query data set, determining a relationship type to which each entity pair in the query data set belongs by using a machine learning model, wherein the machine learning model is pre-trained based on data of other domains except the target domain.

The invention proposes a speech recognition method based on domain invariant features, and applies the speech domain invariant feature extraction model to the end-to-end speech recognition model. The feature extraction model used in the present invention aims at the problem of robustness. By adding more types of voice data to train the voice feature extraction model, better parameters can be obtained, and a better domain-invariant feature extraction model can be obtained. The speech recognition method based on domain-invariant features, using unlabeled pure speech data to train the feature extraction model, and using a small amount of speech with text annotations to train the end-to-end acoustic model, provides an important way to improve the robustness of the end-to-end acoustic model. technical support. Compared with the prior art, the present invention has higher recognition accuracy in different noise environments, a smaller task load of voice tagging tasks, and faster model training and testing speeds.

The invention discloses a video behavior recognition method and device, and relates to the technical field of video behavior recognition. A specific embodiment of the method comprises the following steps: inputting video data collected by a target camera into a feature model to output a feature vector of the target camera; respectively calculating the distance between the feature vector of the target camera and the feature vector corresponding to each camera group, and screening the camera group most similar to the target camera; and taking the video behavior recognition model corresponding to the camera group as a video behavior recognition model of the target camera, and performing behavior recognition on video data collected by the target camera. According to the embodiment, the technical problem that a large amount of data acquisition and labeling work consumes a large amount of time and manpower can be solved.

The invention discloses a target detection model training method for visual detection and a target detection method. The training method comprises the following steps of: 1) selecting a plurality of annotated data and non-annotated data during each iterative training; 2) inputting the labeled image sample data into a target detection model for training, and obtaining a prediction result of each labeled image sample data; 3) loss calculation is carried out according to the prediction result and the corresponding label, and the loss Ls of supervision training is obtained; 4) performing weak enhancement and strong enhancement on each piece of unlabeled image sample data; 5) inputting the weakly enhanced sample data into the target detection model for prediction, and taking an obtained prediction result as a pseudo label corresponding to the strongly enhanced sample data; 6) inputting the strongly enhanced sample data into the target detection model for prediction, and performing loss calculation according to an obtained prediction result and a corresponding pseudo label to obtain a loss Lu of unsupervised training; and 7) adjusting parameters of the target detection model according to Ls and Lu.

The invention provides an automatic industry classification method and system. The method comprises the step of determining a target patent range. The method also comprises the following steps of: defining a target industry tree; generating marks on the target industry tree; performing target patent rough classification by using the marks; and performing target patent fine classification accordingto a rough classification result. According to the automatic industry classification method and system provided by the invention, a direct-push learning method is used to realize the full mining of small-annotation-amount information; the information of the IPC is used, and therefore, so that information dimensions is enriched, and a calculation amount is reduced; hierarchical vectors generated by abstracts, claims and specifications are used; information in the aspect of word order relations is reserved; and patent texts are mined more deeply.

The invention relates to a breast electronic medical record entity recognition system based on multi-standard active learning, which is characterized in that it comprises: a preprocessing module; an entity recognition module; and an active learning module. The present invention considers three aspects of labeling data volume, sentence labeling cost, and data sampling balance, and designs an active learning selection strategy for text sequence labeling to reduce the total labeling workload. On the one hand, the present invention can be used to build systems such as identification marks for breast disease risk patients, disease drug recommendation, decision-making assistance, etc., to help doctors improve the implementation efficiency of standardized diagnosis and treatment of breast diseases, and to provide scientific basis and suggestions; on the other hand, it can also assist Doctors discover potential abnormalities in the diagnosis and treatment process, reduce misdiagnosis and missed diagnosis rates, and improve the cure rate of patients with breast diseases, which is of great value to the intelligent development of breast disease research.

The invention discloses a financial text classification method and device based on a prompt template and electronic equipment, and the method comprises the steps: obtaining an original financial corpus, and converting the financial corpus into an MLM tag based on the prompt template; converting the MLM tag to generate an MLM tag training sample; constructing a pre-training model, inputting the MLM label training sample into the pre-training model, training the pre-training model, and generating a financial text classification model; and inputting financial text data to be identified into the financial text classification model to obtain a financial text classification result. According to the embodiment of the invention, the classification model can still achieve the generalization ability similar to that of a pre-training model under the condition that only a small amount of labeled data exists; the labeling amount is greatly reduced, and large-scale data dependence is reduced; and the financial text classification efficiency is improved.

The invention provides a keyword generation method, which comprises the following steps of: extracting a first important word in a to-be-mined text based on the to-be-mined text in a first extraction mode; based on a second extraction mode, extracting a second important word in the to-be-mined text; based on the first important word and the second important word, unifying word forms and forming an initial important word set which comprises keywords and non-keywords; based on the initial important word set, non-keywords are removed, and a keyword set is generated. According to the keyword generation method, the important words are extracted in two modes, the annotation amount of the to-be-mined text is reduced, the labor cost is saved, and the keyword extraction accuracy is improved. The embodiment of the invention provides a storage medium comprising the keyword generation method and a keyword generation device. The embodiment of the invention provides a server comprising the keyword set.

The present invention provides an industry automatic classification method and system, wherein the method includes determining the target patent scope, and further includes the following steps: defining a target industry tree; generating marks on the target industry tree; using the marks to perform rough classification of target patents; Carry out target patent sub-classification according to the rough classification results. The industry automatic classification method and system proposed by the present invention use the direct push learning method to realize the full mining of small label information; use the information of IPC to enrich the information dimension and reduce the amount of calculation; use abstract, The hierarchical vectors generated by claims and descriptions retain information on word order relationships and dig deeper into patent texts.

The invention provides a semi-automatic labeling method for image data, which includes: performing partial image-level labeling on unlabeled images; putting the labeled data into a collaborative weakly supervised recognition model for training; transforming the collaborative weakly supervised recognition model to obtain strong Supervised transformation model; use the strong supervision transformation model to detect unlabeled images and obtain the detection results; use artificially labeled image data to train the strong supervision transformation model. The method of the present invention uses a certain amount of weakly labeled data in the early stage, and then uses active learning to gradually improve the model in the later stage, thereby ensuring accuracy while reducing the amount of labeling.