48 results about "Latent variable model" patented technology

An computer implemented apparatus and method are disclosed for defining areas of a field in which a crop or other vegetation is grown based on their selective ability to grow such vegetation through a growing season, or some shorter preselected time period. The method includes making a number of temporally separated measurements through air borne imaging of a field, registering the data to the geography of the field and each other, normalizing the data including converting the data to a vegetative index indicative to the presence of vegetation in the field, comparing the data to identify clusters of like value, and classifying the clusters and images to learn how the different field areas responded in growing vegetation through the season. With this method, the field may be segregated into a number of like areas called crop response zones which exhibit similar vegetative growth characteristics as an aid to a grower in his decision making regarding how to maximize yield in his field.

A computer-implemented method for dialog state tracking employs first and second latent variable models which have been learned by reconstructing a decompositional model generated from annotated training dialogues. The decompositional model includes, for each of a plurality of dialog state transitions corresponding to a respective turn of one of the training dialogues, state descriptors for initial and final states of the transition and a respective representation of the dialogue for that turn. The first latent variable model includes embeddings of the plurality of state transitions, and the second latent variable model includes embeddings of features of the state descriptors and embeddings of features of the dialogue representations. Data for a new dialog state transition is received, including a state descriptor for the initial time and a respective dialogue representation. A state descriptor for the final state of the new dialog state transition is predicted using the learned latent variable models.

A hierarchical latent structure setting unit 81 sets a hierarchical latent structure that is a structure in which latent variables are represented by a tree structure and components representing probability models are located at nodes of a lowest level of the tree structure. A variational probability computation unit 82 computes a variational probability of a path latent variable that is a latent variable included in a path linking a root node to a target node in the hierarchical latent structure. A component optimization unit 83 optimizes each of the components for the computed variational probability. A gating function optimization unit 84 optimizes a gating function model that is a model for determining a branch direction according to the multivariate data in a node of the hierarchical latent structure, on the basis of the variational probability of the latent variable in the node.

Software that performs the following steps: (i) receiving a first set of observed data pertaining to healthcare events, the first set of observed data including a subset of patient care event data pertaining to patient care events and a subset of prescription data pertaining to prescription events; (ii) generating a graphical model representing a probabilistic relationship between the patient care event data and the prescription data, the graphical model including a set of latent variable(s) estimated from the first set of observed data using an expectation maximization method; (iii) receiving a second set of observed data pertaining to healthcare events associated with a healthcare provider; and (iv) computing, using a dynamic programming approach, a first prescription score for the healthcare provider relating to a computed probability under the generated graphical model of at least one prescription event of the second set of observed data.

The invention relates to an industrial process fault detection method based on an autoregression dynamic hidden variable model. According to the method, by use of data at a normal operation state in an industrial process, a dynamical process model with universality is established, and the model predicts a model structure through a Kalman filter and a smoother and an expectation maximization algorithm, such that dynamical and static coupling relation of industrial data can be effectively extracted; and then, based on dynamic and static noise of the dynamic model, a corresponding monitoring statistical amount is constructed, and a final fault detection result is obtained. Compared with other existing methods, the method provided by the invention can greatly improve the effects of industrial process dynamic modeling and fault detection, reduces the false alarm rate and the missed alarm rate of faults, improves the monitoring performance to a quite large degree, enhances the understanding capability and the operation confidence of process operators for the process and better facilitates automation implementation of the industrial process.

The invention relates to a traffic detection method for a peer-to-peer (P2P) network. The method is used for solving the technical problem of low detection accuracy of the conventional network traffic detection method. The technical scheme is that: a classifier is trained in two stages, a value of a positive instance sample number N in a test sample is approximately estimated by using semi-supervised clustering, and a two stage variable model (TSVM) is further trained according to the value of N. Compared with a background technology, the invention makes the value of N closer to a true value, endows the trained classification TSVM with high stability and robustness and improve network traffic detection accuracy. A great amount of unmarked data takes part in the training of the classification model, and the advantages of semi-supervised learning are fully utilized; therefore, compared with a conventional supervised learning algorithm in which the model is trained only by marked data, the method is higher in accuracy and stability.

The invention discloses a variable selection method for modeling an organic pollutant quantitative structure and activity relationship. The method comprises the following steps of: calculating linear models combined with all single variables and different bivariables, and retaining a certain number of optimal models for the single variables and the bivariables; then sequentially taking out a model from the retained bivariable linear models, and combining two of the variables and each of the rest variables to form a tri-variable model until all the retained bivariable models are processed; comparing the quality of the tri-variable models, and retaining a certain number of optimal tri-variable models; and repeating, and stopping calculation until the number of variables forming the models meets the requirement, wherein the quality of the models is based on an end standard represented by q2 or a root-mean-square deviation (RMSEV) which is calculated by leave-one-out cross validation (LOOCV) or leave-multiple-out cross validation (LMOCV). The theory is simple and can be understood easily and programmed easily; and the method is quick and effective, so that the rationality of variable selection and the stability of the forecast capacity of the models are guaranteed.

The invention relates to a word definition generation method based on a recurrent neural network and a latent variable structure in the field of natural language processing. On the basis of a recurrent neural network, a variational auto-encoder (VAE) is used for modeling paraphrases, latent variable characteristics are combined. The paraphrases are extracted according to context information of defined words to generate paraphrases of the words, and the method specifically comprises the steps of establishing and arranging a basic corpus; selecting a synonym set of the defined words, and expanding the basic corpus to form a final corpus; carrying out expansion reconstruction on the word vectors of the defined words; constructing a structure model based on the recurrent neural network and thelatent variable; training a latent variable structure model based on a recurrent neural network; and inputting the to-be-paraphrased words and the context information thereof into the trained model to realize semantic paraphrasing of the to-be-paraphrased words in a specific context, thereby achieving polysemy.

The invention discloses a deep learning characteristic generalization method based on a latent variable model. The method is carried out according to the following steps that: S1: on the basis of thelatent variable model, establishing one pair of bidirectional parametric mapping between data space and latent variable space, and combining with a weighted relationship to construct a characteristicgeneralization layer; S2: embedding the characteristic generalization layer into a deep neural network, and dividing the network into three parts, including a DNN (Deep Neural Network)-1, the characteristic generalization layer and a DNN-2; S3: determining the optimization objective of the model, and defining a target function; and S4: lowering characteristic pattern data complexity, and establishing a multi-branch parallel forward propagation structure. The method is favorable for improving the generalization ability of a deep network model, and an overfitting phenomenon during small-scale data training can be lightened. Compared with other methods which use a generative model to enhance data, the method disclosed by the invention simplifies network complexity and improves training efficiency.

A computer-implemented method performed in connection with a computerized system incorporating a processing unit and a memory, the computer-implemented method involving: using the processing unit to generate a multi-modal language model for co-occurrence of spoken words and displayed text in the plurality of videos; selecting at least a portion of a first video; extracting a plurality of spoken words from the selected portion of the first video; extracting a first displayed text from the selected portion of the first video; and using the processing unit and the generated multi-modal language model to rank the extracted plurality of spoken words based on probability of occurrence conditioned on the extracted first displayed text.

A hierarchical latent structure setting unit 81 sets a hierarchical latent structure that is a structure in which latent variables are represented by a tree structure and components representing probability models are located at nodes of a lowest level of the tree structure. A variational probability computation unit 82 computes a variational probability of a path latent variable that is a latent variable included in a path linking a root node to a target node in the hierarchical latent structure. A component optimization unit 83 optimizes each of the components for the computed variational probability. A gating function optimization unit 84 optimizes a gating function model that is a model for determining a branch direction according to the multivariate data in a node of the hierarchical latent structure, on the basis of the variational probability of the latent variable in the node.

A model estimation device includes: a data input unit 101; a state number setting unit; an initialization unit; a latent variable variational probability computation unit which computes a variational probability of a latent variable so as to maximize a lower bound of a model posterior probability limited in degree of freedom; a component optimization unit which estimates an optimal type of each component and a parameter thereof so as to maximize the lower bound of the model posterior probability limited in degree of freedom and separated for each component of a latent variable model; a free parameter selection variable computation unit which computes the free parameter selection variable; an optimality determination unit which determines whether or not to continue the maximization of the lower bound of the model posterior probability; and a result output unit.

The invention relates to a method for user portrait extraction based on a multilayer latent variable model and relates to the field of data mining and recommendation systems. A user portrait is extracted according to a social curation network, and the method for user portrait extraction based on the multilayer latent variable model is provided according to data of two modes including text description information of collected entries and user behaviors on a forward chain. A latent Dirichlet allocation (LDA) model is introduced to the text description information to obtain user's latent subject distribution, and subject interest distribution is obtained based on the user's latent subject distribution; and users' interest distribution is obtained in combination with the user's latent subject distribution and the subject interest distribution. A users' social community is found based on the multilayer latent variable model, and user recommendation results are obtained in combination with Jensen-Shannon divergence ascending sort. According to the method, the users' social community is found by utilization of information of the two different modes including the user text description information and the user behaviors on the forward chain, and user recommendation is achieved.

The invention discloses a latent variable model-based user preference extraction method. The method comprises the steps of firstly selecting N commodity relative attributes to form a commodity property set, building according to historical data to obtain a bayesian network, searching in the bayesian network to obtain a maximal semi-clique, and then inserting a latent variable L, showing user preference, into the maximal semi-clique, so as to obtain a latent variable model, wherein L being equal to 1 shows that a user prefers, and L being equal to 0 shows that the user does not prefer; performing parameter learning on the latent variable model to obtain a conditional probability table of various nodes in the latent variable mode; then according to the conditional probability table of the latent variable L, performing user preference extraction: searching to obtain an attribute combination item corresponding to a conditional probability maximum when L is equal to 1, wherein the attribute combination item corresponds to commodity types most preferred by the user; searching to obtain an attribute combination item corresponding to the conditional probability maximum when L is equal to 0, wherein the attribute combination item corresponds to commodity types least preferred by the user. By aiming at the user preference hidden in commodity evaluation data, the more objective and realistic user preference results are extracted by the structure of the bayesian network.

The invention relates to a method for tracking the human body posture based on visual information. According to the technical scheme, the method comprises the steps that in the training stage, according to real three-dimensional human body posture time series data in training data and with the adoption of the Gauss latent variable model algorithm, the human body posture is studied, so that a human body posture motion prior model is obtained; in the tracking test stage, video data to be tested and a first frame of the three-dimensional human body posture in the video data are input, a latent variable motion model is established with the use of the human body posture motion prior model studied in the training stage, and the particle updating step in an immune particle filtering algorithm is achieved through the latent variable motion model. The method for tracking the human body posture based on the visual information has the advantages of being high in practicality, accuracy and efficiency.

A hierarchical latent structure setting unit 81 sets a hierarchical latent structure that is a structure in which latent variables are represented by a tree structure and components representing probability models are located at nodes of a lowest level of the tree structure. A variational probability computation unit 82 computes a variational probability of a path latent variable that is a latent variable included in a path linking a root node to a target node in the hierarchical latent structure. A component optimization unit 83 optimizes each of the components for the computed variational probability. A gating function optimization unit 84 optimizes a gating function model that is a model for determining a branch direction according to the multivariate data in a node of the hierarchical latent structure, based on the variational probability of the latent variable in the node.

The invention belongs to the technical field of children personalized behavior analysis, and discloses a children personalized behavior statistical analysis system and method based on a latent variable model. The children personalized behavior statistical analysis method includes the steps: according to an established latent variable model, applying a latent variable to analysis of a personalizedbehavior problem through a mathematical model; making a main tool-scale for latent variable measurement, the scales including an evaluation scale and an attitude scale, and the scales being divided into a three-point scale, a five-point scale and a seven-point scale from the perspective of questionnaire question options; and analyzing the internal relation between the difference influencing the personalized behaviors of the children and the latent variable factors, discovering the potential ability of the children from the behavior performance of the children, and providing scientific suggestions for the personalized development of the children. According to the children personalized behavior statistical analysis method, the special ability of some children who do not reach the standard can be discovered, or the potential ability of personalized children can be explored, so that scientific suggestions and guidance directions suitable for personalized development of children can be given; and in cooperation with an enterprise, an education product is developed according to the predicted personalized behavior preference result of the children.

The invention discloses a multi-sampling rate soft sensing method based on a dynamic hidden variable model. By taking a large number of process variables with different sampling rates in chemical processes and a small number of key quality variables as the modeling samples, dynamic latent variables which can contain the characteristics of multi-rate data are extracted while fully considering the autocorrelation and cross-correlation of data, and the estimation of model parameters is realized by an expectation maximization algorithm and a Kalman filter algorithm. Based on this model, the corresponding soft-sensing method is established to solve the estimation problem of multi-rate dynamic key quality variables. The method not only realizes the multi- sampling rate information processing, but also can make full use of the data information. Moreover, the dynamic characteristics of the data can be fully considered by a Kalman filter, and the dynamic latent variables can be accurately estimated, so that the estimation and description of the key quality variables which are difficult to be directly measured can be better realized for a few of the dynamic latent variables after dimension reduction, and the soft sensing accuracy and application range can be improved.

The invention discloses a dynamic non-gaussian process monitoring method based on a dynamic latent independent variable. The dynamic non-gaussian process monitoring method aims to combine the advantages of a dynamic latent variable model capable of processing dynamic data and an independent component analysis model capable of processing non-gaussian data. Specifically, the dynamic non-gaussian process monitoring method comprises the steps: firstly, a dynamic latent variable algorithm is utilized to extract auto-correlation dynamic characteristic components and cross-correlation static characteristic components; secondly, after the characteristic components are whitened, the combined whitening characteristic components are utilized as initial independent components to obtain a dynamic latent independent variable model in an iteration mode; and finally, dynamic non-gaussian process monitoring is carried out based on the dynamic latent independent variable. It can be said that the methodutilizes the capacity, for separately extracting the dynamic components and the static components, of the dynamic latent variable algorithm, then an independent component analysis algorithm which canextract the non-gaussian characteristic components is further combined, and thus the dynamic non-gaussian process monitoring method is feasible.

The invention provides a thermal power plant boiler flue gas oxygen content soft measurement method, and relates to the technical field of thermal power plant power generation. The method comprises the steps of obtaining historical working data of a thermal power plant boiler; Normalization processing is carried out and then is divided into a control variable and a state variable. carrying out feature selection on the state variable data set by adopting a Ridge Regression method; and dividing the control variable data set and the state variable data set into a training set and a test set, training the DBN model by using the training set of the control variable and the state variable respectively to obtain a control variable model and a state variable model, and performing nonlinear combination on the control variable model and the state variable model to obtain a final flue gas oxygen content combined prediction model. The method can overcome the limitation of a traditional algorithm network structure, extracts the deep features of the data, has the advantages of high prediction precision, high convergence speed and the like, provides a basis for the application of an advanced control algorithm, and is beneficial to improving the boiler efficiency and reducing the boiler emission.

The invention relates to a roadside pedestrian trajectory generation algorithm based on an adversarial generative network. A multi-mode prediction trajectory is generated by using a social attention mechanism and a pedestrian trajectory latent variable; by adversarial generation training of a trajectory generator and a discriminator, capabilities of a generator and the discriminator are continuously optimized, and the trajectory generation accuracy of the generator is improved; a social attention mechanism based on the head orientation is provided, the head orientation of the pedestrian is obtained through the final speed direction of the pedestrian, the cosine value of the included angle between pedestrians is calculated according to the head orientation information, and the soft attention and hard attention mechanism optimizes the output of the social attention mechanism through the calculated angle information. Converging and outputting operation are carried out by a maximum poolinglayer. a new latent variable generation method is provided, two feedforward neural networks are used for learning latent variables from a pedestrian historical track and an observation track respectively, inputs of a latent variable generator comprise the position, the speed and the acceleration, and distribution of three kinds of latent variables is generated from the three inputs respectively.

The invention relates to an item semantic annotation method based on a use event of a heterogeneous item. The method comprises the following steps: step 1) of establishing an item potential relationship intensity mining model; step 2) of extracting an item implicit feature; step 3) of extracting an item explicit feature; step 4) of training an item semantic annotation model; after the above four features are extracted, combining (FLatent+Ftext+FS+FT) together as training data of a binary classification SVM classifier, and gathering labels predicted by all independent binary classification device as the semantic annotation result of the item. The item semantic annotation method has the advantages that based on a latent variable model establishing item implicit relationship strength based onthe use event of the heterogeneous item, the item implicit feature is extracted therefrom, the category semantic label of the given item is predicted by comprehensively utilizing the explicit and implicit features, and the method is significantly superior to an existing item labeling method based on explicit features in terms of efficiency and performance and the like.

A computer-implemented method for dialog state tracking employs first and second latent variable models which have been learned by reconstructing a decompositional model generated from annotated training dialogs. The decompositional model includes, for each of a plurality of dialog state transitions corresponding to a respective turn of one of the training dialogs, state descriptors for initial and final states of the transition and a respective representation of the dialog for that turn. The first latent variable model includes embeddings of the plurality of state transitions, and the second latent variable model includes embeddings of features of the state descriptors and embeddings of features of the dialog representations. Data for a new dialog state transition is received, including a state descriptor for the initial time and a respective dialog representation. A state descriptor for the final state of the new dialog state transition is predicted using the learned latent variable models.

The invention provides a story generation method and device based on a hidden variable model, and the method comprises the steps of generating the outline of a story according to a title of the story based on a trained planning model in the hidden variable model, wherein the hidden variable in the hidden variable model is the outline; and generating the story according to the outline and the title based on a generation model in the trained hidden variable model. According to the present invention, on one hand, the story outline is used as the auxiliary guidance information to enable the generated story to be more smooth and to stick to the topic, the information difference between the title and the story in the story generation can be reduced, and the story generation difficulty is reduced; on the other hand, the story outline is modeled as a hidden variable, the form of the hidden variable is not limited, the hidden variable can contain more information, and the hidden variable is completely learnable and has very high learning capability; in addition, the method can be expanded into the text generation in other forms, and has very good practicability and universality.

The invention discloses a hierarchical latent variable model-based news recommendation method. The method comprises the steps of firstly extracting different types of news from major portal websites; then according to tags of the news, extracting news features; for registered users of a system, extracting explicit features and implicit features of the users to form user features; according to the user features and the news features, clustering the users and the news; adopting different latent variable models for the clustered news and users; predicting scores of the users to the news; selecting out multiple pieces of the news with the highest scores; and recommending personalized news to the users.

The invention discloses a variable model, which comprises a vertical pillar, a sliding plate, a transverse motion motor, a grabbing manipulator, a vertical motion motor, a base, ball head telescopic rods, a supporting sleeve, a vertical shaft, radial ball bearings, thrust ball bearings, rotary brackets, a rotary motion motor, a linear guide rail pair and a supporting flat plate, wherein the linear guide rail pair on the vertical pillar is connected to the sliding plate; the vertical motion motor is capable of transmitting the sliding plate to move up and down by virtue of a screw lead; the grabbing manipulator, under the control of an air cylinder, can complete a motion of grabbing the ball head telescopic rods and can achieve precise positioning; the ball head telescopic rods are supported in inner holes of the supporting flat plate and are equally divided into multiple parts along the circumference, and the ball head telescopic rods are designed in multiple layers in a height direction; the rotary motion motor can drive the supporting sleeve to do a rotary motion by virtue of a gear transmission pair; a control system, in accordance with corresponding data measured from a human body, can conduct entry programming, and can drive the vertical motion motor, the transverse motion motor and the rotary motion motor; and an enveloping surface, in a human body form, is formed by the ball head bodies of all ball head telescopic rods. With the application of the variable model, it can cut dresses according to figures and can complete trying verification in a process of making clothing.