809 results about "Model learning" patented technology

The invention provides an unsupervised domain-adaptive brain tumor semantic segmentation method based on deep adversarial learning. The method comprises the steps of deep coding-decoding full-convolution network segmentation system model setup, domain discriminator network model setup, segmentation system pre-training and parameter optimization, adversarial training and target domain feature extractor parameter optimization and target domain MRI brain tumor automatic semantic segmentation. According to the method, high-level semantic features and low-level detailed features are utilized to jointly predict pixel tags by the adoption of a deep coding-decoding full-convolution network modeling segmentation system, a domain discriminator network is adopted to guide a segmentation model to learn domain-invariable features and a strong generalization segmentation function through adversarial learning, a data distribution difference between a source domain and a target domain is minimized indirectly, and a learned segmentation system has the same segmentation precision in the target domain as in the source domain. Therefore, the cross-domain generalization performance of the MRI brain tumor full-automatic semantic segmentation method is improved, and unsupervised cross-domain adaptive MRI brain tumor precise segmentation is realized.

A learning-based method for estimating costs or statistics of an operator in a continuous query includes a cost estimation model learning procedure and a model applying procedure. The model learning procedure builds a cost estimation model from training data, and the applying procedure uses the model to estimate the cost associated with a given query. The learning procedure uses a feature extractor and a cost estimator. The feature extractor collects relevant training data and obtains feature values. The extracted feature values are associated with costs and used to create the cost estimator. When applying the cost estimator to a continuous stream of data, the feature extractor extracts feature values from the data stream and uses the extracted feature values as inputs into the cost estimator to obtain the desired cost values.

The invention discloses a dialogue automatic reply system based on deep learning and reinforcement learning. The dialogue automatic reply system comprises a user interaction module which receives question information inputted by a user in a dialogue system interface; a session management module which records the active state of the user, wherein the active state includes historical dialogue information, user position transformation information and user emotion change information; a user analysis module which analyzes registration information and the active state of the user and portraits for the user so as to obtain user portrait information; a dialogue module which generates reply information through a language module according to the question information of the user with combination of the portrait of the user; and a model learning module which updates the language model through the reinforcement learning technology according to the reply information generated by the language model. According to the dialogue automatic reply system based on deep learning and reinforcement learning, the reply of the dialogue meeting the personality of the user can be given according to the dialogue text inputted by the user with combination of context information, the personality characteristics of the user and the intentions in the dialogue.

The invention discloses a network fault diagnosis method based on deep learning in a network virtualization environment. The network fault diagnosis method comprises the steps of: dividing a network into a physical network and a virtual network, combining the characteristics of occurrence of network faults, considering the time influencing factor, network topological connection characteristics and a mapping relation between the virtual network and the physical network, and comprehensively evaluating the network faults by means of a fault severity grading probability; regarding network characteristic parameters with influence degrees as a model learning resource, paying attention to the correspondence between variation trend of network historical data and fault tags, establishing a network fault diagnosis model with multiple fault grading probabilities in the network virtualization environment based on a viewing angle of deep learning, and training network parameters by using the network fault diagnosis model; and adjusting a fault prediction model in the training process, and utilizing an optimized and adjusted deep learning network to realize fault diagnosis in the network virtualization environment. The network fault diagnosis method can carry out deep analysis on the network parameters in the network virtualization environment, therefore the network fault diagnosis method has higher precision in predicting the network faults.

Method and apparatus for recognizing landmark buildings in an image and then locating the recognized landmark buildings onto a map together with related information wherein a first database is employed to store models formed by mathematical set descriptions of landmark buildings which are learned from a set of training images of a model-learning module captured by an imaging device for each building, and a second database is employed to store the related information of each landmark building. The model of each landmark building is represented as a set of features and the geometric relationship between them by clustering the salient features extracted from a set of training images of the landmark building.

The invention provides a knowledge graph representation learning method and system in combination with entity description. According to the method and the system, a continuous bag-of-words-based model and a convolutional neural network-based model are proposed for constructing description-based vector representation of entities; not only triple relationship information among the entities but also text information contained in entity description are utilized, and two entity vector representation modes obtained by model learning are used, so that higher accuracy can be obtained in tasks of knowledge graph complementation, entity classification and the like; and the description-based vector representation constructs entity vectors through the text information, can well represent new entities or entities inexistent in a training set, and has high practicality.

A learning-based method for estimating costs or statistics of an operator in a continuous query includes a cost estimation model learning procedure and a model applying procedure. The model learning procedure builds a cost estimation model from training data, and the applying procedure uses the model to estimate the cost associated with a given query. The learning procedure uses a feature extractor, a confidence adjustor and a cost estimator. The feature extractor collects relevant training data and obtains feature values. The extracted feature values are associated with costs and used to create the cost estimator. The extracted feature values, the associated costs, the cost estimator, and a user interface are used to create a confidence adjuster. When applying the confidence adjuster and the cost estimator to a continuous stream of data, the feature extractor extracts feature values from the data stream, uses the extracted feature values as input into the confidence adjuster to determine whether or not the cost estimator should be used, and if so, uses the extracted feature values as inputs into the cost estimator to obtain the desired cost values.

The invention discloses a printing image defect detection method, which comprises the following steps: carrying out real-time image model learning in a gray scale region and a gradient region by aiming at specific images on a large image; comparing the large image to be detected to the a gray scale region model and a gradient region model established during learning; realizing small-dimension strong-contrast defect detection through Blob cluster analysis; when no small-dimension strong-contrast defect is detected, dividing the large image into sub regions, and respectively calculating image integrated features; adopting a variable threshold method for carrying out threshold division on each sub region; carrying out Blob cluster analysis on divided images; and realizing the large-area weak-contrast defect detection. Compared with the prior art, the invention does not rely on a reference template, is not sensitive on the real-time imaging brightness change, can overcome the defects of missing detection, error detection and poor self adaptation of the template by using a reference template comparison detection method, and can simultaneously solve the problems of eliminating tiny wrinkles and blackspots in printing and papermaking industries.

The disclosed embodiments provide a system that detects anomalous events. During operation, the system obtains machine-generated time-series performance data collected during execution of a software program in a computer system. Next, the system removes a subset of the machine-generated time-series performance data within an interval around one or more known anomalous events of the software program to generate filtered time-series performance data. The system uses the filtered time-series performance data to build a statistical model of normal behavior in the software program and obtains a number of unique patterns learned by the statistical model. When the number of unique patterns satisfies a complexity threshold, the system applies the statistical model to subsequent machine-generated time-series performance data from the software program to identify an anomaly in an activity of the software program and stores an indication of the anomaly for the software program upon identifying the anomaly.

A face recognition system based on adaptive learning includes a specific person detection and tracking unit for detecting and tracking a specific person from a moving image. A facial feature extraction unit extracts a plurality of facial feature vectors from the detected and tracked specific person. A face recognition unit searches for a given registration model by comparing the extracted facial feature vectors with facial feature vectors of the registration models previously stored in a user registration model database. A learning target selection unit selects a facial feature vector to be added to a record of the given registration model from among the extracted facial feature vectors. A registration model learning unit adds and updates the selected facial feature vector to the record of the given registration model.

A vehicular control device includes (i) situation detection units that detect situation of the vehicle, occupants, or a surrounding area; (ii) a storage unit for storing a model representing association between variables indicating the situation and variables indicating presence or absence of operation needs for each in-vehicle unit, (iii) an inference execution unit that determines presence or absence of the operation needs for each in-vehicle unit using the stored model; (iv) a control unit that controls each in-vehicle unit based on a result of determination by the inference execution unit; (v) an operation detection unit that detects the operation of each in-vehicle unit; and (vi) a model learning unit that learns the stored model using the detected situation when the operation of each in-vehicle unit is detected by the operation detection unit, and an event involving the presence of operation needs for the in-vehicle unit.

The embodiment of the invention provides an interest point recommendation method on user dynamic preference and an attention mechanism. The interest point recommendation method comprises the followingsteps: S1, obtaining a historical behavior record of a user, constructing a historical behavior sequence of the user, and dividing the historical behavior sequence of the user into a long-term historical behavior sequence and a short-term historical behavior sequence; S2, respectively inputting the long-term historical behavior sequence and the short-term historical behavior sequence into a long-term preference model and a short-term preference model to learn the long-term preference and the short-term preference of the user; S3, integrating the long-term preferences and the short-term preferences of the users to obtain final preferences of the users; and S4, calculating a score of the user to the place according to the final preference of the user, and recommending interest points to theuser according to the score of the user to the place. According to the interest point recommendation method, dynamic modeling of user preferences is realized, and accurate representation of the userpreferences can be obtained, and the interest point recommendation effect is improved.

A data processing apparatus includes an obtaining unit configured to obtain time-series data from a wearable sensor, an activity model learning unit configured to learn an activity model representing a user activity state as a stochastic state transition model from the obtained time-series data, a recognition unit configured to recognize a current user activity state by using the activity model of the user obtained by the activity model learning unit, and a prediction unit configured to predict a user activity state after a predetermined time elapses from a current time from the current user activity state recognized by the recognition unit.

A structured prediction model learning apparatus, method, program, and recording medium maintain prediction performance with a smaller amount of memory. An auxiliary model is introduced by defining the auxiliary model parameter set θ^(k) with a log-linear model. A set Θ of auxiliary model parameter sets which minimizes the Bregman divergence between the auxiliary model and a reference function indicating the degree of pseudo accuracy is estimated by using unsupervised data. A base-model parameter set λ which minimizes an empirical risk function defined beforehand is estimated by using supervised data and the set Θ of auxiliary model parameter sets.

The invention discloses a deep learning-based electric power suspected complaint work order recognition method, and relates to complaint recognition method for electric power customer services. At present, electric power suspected complaint work orders are artificially recognized, so that the efficiency is low, the timeliness is bad and human resources are wasted. The method comprises the following steps of: 1) deep learning model configuration; 2) complaint feature label extraction; 3) complaint sample formatting; 4) model learning and training; 5) suspected complaint recognition; and 6) suspected complaint classification. According to the method, a series of work such as data cleaning and sorting, complaint tendency word extraction, data modeling and optimization, sample iterative learning and training, machine learning and prediction and the like can be carried out on complaint lists through a deep learning technology, so that deep learning-based intelligent recognition and classification of suspected complaint work orders are realized, the intelligent work experience is enhanced and the service quality control work efficiency is improved.

A statistical model learning device is provided to efficiently select data effective in improving the quality of statistical models. A data classification means 601 refers to structural information 611 generally possessed by a data which is a learning object, and extracts a plurality of subsets 613 from the training data 612. A statistical model learning means 602 utilizes the plurality of subsets 613 to create statistical models 614 respectively. A data recognition means 603 utilizes the respective statistical models 614 to recognize other data 615 different from the training data 612 and acquires each recognition result 616. An information amount calculation means 604 calculates information amounts of the other data 615 from a degree of discrepancy among the statistical models of the recognition results. A data selection means 605 selects the data with a large information amount and adds the same to the training data 612.

The invention discloses a news event evolution analysis method based on time sequence distribution information and a topic model and relates to the field of text analysis. The method comprises the following steps: firstly, dividing a corpus into a plurality of sub-corpuses according to time by analyzing distribution characteristics, presented on a time sequence, of a news report, and by using a K-Means clustering algorithm; secondly, sequentially performing topic modeling on each sub-corpus by using the topic model, and learning the model through a Gibbs sampling method to obtain topic distribution information of each sub-corpus; finally, connecting topics between which the distance is minimum in series by calculating a Jensen-Shannon distance between each two topics in the adjacent sub-corpuses, wherein the topics are connected in series to obtain a main topic of an event, and auxiliary topics except the main topic in each sub-corpus are concerns and new developments of the event in each stage. According to the method, the mainline of event development in a news prediction and new concerns burst in each stage can be better described.

The invention relates to the field of production and control of a finish-rolled band steel, in particular to a method for controlling a thickness of a finish-rolled band steel. An adaptive control method for increasing the thickness precision of a finish-rolled steel band by utilizing a roll gap is characterized in that the method comprises the following steps of establishing a roll gap modification quantity static table which performs indexing according to the layers of band steels; before rolling the band steels, identifying the layer of the current rolling band steel, determining the roll gap modification current value, summing the roll gap modification current value with the zero point modification current values of racks, obtaining a roll gap model learning coefficients of the racks used for rolling setting pre-calculation; after rolling the band steels, detecting the obtained band steel final rolling thickness value and the actual roll gap value of the rack at the finish-rollingend, calculating the values, obtaining the band steel final rolling thickness deviation, the initial roll gap deviation value and the roll gap modification update value, for example, when the layer of the same coil of band steel is different with that of the former coil of band steel, the roll gap modification update value is saved in the roll gap modification quantity static table; and collecting the roller speed, roll gap, rolling force and final rolling thickness actually measured data of each rack, and then performing zero point modification of the roll gap. According to the method provided by the invention, the roll gap modification value and zero point modification are used jointly, so that the adaptive capability of the roll gap model under the condition that the rolling planning specification is variable can be improved, thereby improving the setting precision of the model and realizing accurate control to the thickness of the band steel.

The invention discloses an event joint extraction method fusing local features and deep learning. The event joint extraction method fusing local features and deep learning comprises the steps of performing entity extraction, part-of-speech and dependency analysis and PV-utilization; training document vectors by using a PV-DM model learning hidden features, and entering the GCN; thirdly, identifying a memory unit between the event element and the event trigger, and helping to identify an event type and an event argument; and finally, marking the final event type by using the CRF, and identifying event elements by using a Softmax function in a full connection layer. The method has the beneficial effects that the subject information in the article is learned by means of the document vector, then the relation between events is mined by means of the graph convolutional network, and finally the information between the event type and the event element is learned by means of the memory unit and the local feature, so that the event joint extraction is completed, and the recognition performance is improved.

An information processing device includes a model learning unit that carries out learning for self-organization of internal states of a state transition prediction model which is a learning model having internal states, a transition model of the internal states, and an observation model where observed values are generated from the internal states, by using first time series data, wherein the model learning unit learns the observation model of the state transition prediction model after the learning using the first time series data, by fixing the transition model and using second time series data different from the first time series data, thereby obtaining the state transition prediction model having a first observation model where each sample value of the first time series data is observed and a second observation model where each sample value of the second time series data is observed.

The invention discloses a spam mail judging system based on behavior analysis. The system comprises a system interface for human-machine interface operation and acquiring user behavior data in real time from an online mail system, a user behavior collection module responsible for acquiring various types of user behavior information from the system interface, performing preliminary analysis and statistics, and storing relevant results into a user behavior database through a database interface, a user creditworthiness model learning module for receiving various model parameters set by a user and the user behavior information from the user behavior database, learning user creditworthiness models and storing the learned user creditworthiness models into a user creditworthiness model base, and a user creditworthiness estimation module for estimating the creditworthiness of each user and updating an estimation result into a user creditworthiness database. In the spam mail judging system, the current mail sending behavior and historic mail sending behaviors are considered at the same time; and the manual examining quantity is reduced greatly.

The invention brings forward a Chinese word segmentation method by using character embedding based on word context and a neural network. The character embedding is learnt on large-scale automatic segmentation of data. The learnt character embedding is used as inputting of a segmentation model of the neural network so as to effectively help model learning. The method comprises the concrete steps of learning character embedding on large-scale automatic segmentation of data according to the word context and lexeme marking; and utilizing character-embedding as inputting of the segmentation model of the neural network so that segmentation performance is improved. Compared with other Chinese word segmentation method technologies based on the neural network, the method adopts the character-embedding technology based on the word context so that word information is effectively integrated into the segmentation model. Therefore, the accuracy of a word segmentation task is improved.

The invention discloses a wireless local area network intrusion detection system based on deep learning, and relates to the field of network space security. The system comprises model learning and intrusion detection, the model learning is responsible for pre-constructing an RNN neural network, and intrusion detection uses the constructed RNN neural network to complete an intrusion detection taskin real time. An LSTM recurrent neural network is adopted to carry out classification prediction on data traffic of a wireless local area network with a time sequence characteristic; the type of the target network traffic sequence is judged according to the output predicted by the identification module; The intrusion behavior in the target network is identified, different granularity processing isperformed according to the set priorities and processing methods, early warning and prevention are performed on possible information security problems, the confidentiality, the usability and the integrity of the wireless local area network are guaranteed, and the security level of the wireless local area network is improved.