32 results about "Probabilistic generative model" patented technology

The invention provides a sentiment classification method for generating a model deep-convinced-degree network on the basis of the probability of depth study. According to the technical scheme of the method, a plurality of Boltzmann machine layers are stacked, namely, output of this layer is used as input of the next layer. By the adoption of the mode, input information can be expressed in a grading mode, and abstraction can be conducted layer by layer. A multi-layer sensor containing a plurality of hidden layers is the basic study structure of the method. More abstract high layers are formed through combining the characteristics of lower layers and are used for expressing attribute categories or characteristics, so that the distribution type character presentation of data can be discovered. The method belongs to monitoring-free study, and a mainly-used model is the deep-convinced-degree network. The method enables a machine to conduct characteristic abstract better so as to improve the accuracy of sentiment classifications.

A system and method facilitating object tracking is provided. The system includes an audio model that receives at least two audio input signals and a video model that receives a video input. The audio model and the video model employ probabilistic generative models which are combined to facilitate object tracking. Expectation maximization can be employed to modify trainable parameters of the audio model and the video model.

Certain aspects of the present disclosure are directed to methods and apparatus for predicting subject motion using probabilistic models. One example method generally includes receiving training data comprising a set of subject pose trees. The set of subject pose trees comprises a plurality of subsets of subject pose trees associated with an image in a sequence of images, and each subject pose tree in the subset indicates a location along an axis of the image at which each of a plurality of joints of a subject is located. The received training data may be processed in a convolutional neural network to generate a trained probabilistic model for predicting joint distribution and subject motion based on density estimation. The trained probabilistic model may be deployed to a computer vision system and configured to generate a probability distribution for the location of each joint along the axis.

One embodiment of the present invention provides a system that learns a generative model for textual documents. During operation, the system receives a current model, which contains terminal nodes representing random variables for words and cluster nodes representing clusters of conceptually related words. Within the current model, nodes are coupled together by weighted links, so that if a cluster node in the probabilistic model fires, a weighted link from the cluster node to another node causes the other node to fire with a probability proportionate to the link weight. The system also receives a set of training documents, wherein each training document contains a set of words. Next, the system applies the set of training documents to the current model to produce a new model.

One embodiment of the present invention provides a system that selectively deletes clusters of conceptually-related words from a probabilistic generative model for textual documents. During operation, the system receives a current model, which contains terminal nodes representing random variables for words and contains one or more cluster nodes representing clusters of conceptually related words. Nodes in the current model are coupled together by weighted links, so that if an incoming link from a node that has fired causes a cluster node to fire with a probability proportionate to a weight of the incoming link, an outgoing link from the cluster node to another node causes the other node to fire with a probability proportionate to the weight of the outgoing link. Next, the system processes a given cluster node in the current model for possible deletion. This involves determining a number of outgoing links from the given cluster node to terminal nodes or cluster nodes in the current model. If the determined number of outgoing links is less than a minimum value, or if the frequency with which the given cluster node fires is less than a minimum frequency, the system deletes the given cluster node from the current model.

One embodiment of the present invention provides a system that learns a generative model for textual documents. During operation, the system receives a current model, which contains terminal nodes representing random variables for words and cluster nodes representing clusters of conceptually related words. Within the current model, nodes are coupled together by weighted links, so that if a cluster node in the probabilistic model fires, a weighted link from the cluster node to another node causes the other node to fire with a probability proportionate to the link weight. The system also receives a set of training documents, wherein each training document contains a set of words. Next, the system applies the set of training documents to the current model to produce a new model.

The invention relates to a method for building a friend relationship transitive tree in a social network, which belongs to the technical field of computer data mining. The friend relationship transitive tree is built for each self node, and self nodes and friend nodes are added. A social behavior probability generating model is set up for each time period, the prior parameters of the intermedium preference probability distribution and friend relationship strength distribution of the self nodes and the candidate intermediaries of each self node and new friend nodes are initialized, the candidate intermediaries is continuously interated and updated, a candidate intermediary z with the most sampling frequency is designated as the intermediary between a self node u and a new friend node y, and z to y are added into the social network of the self node u. The method models user behaviors in the social network on the basis of friend relationship transitivity; and a flat social network layering presentation method is provided, so that the reason and an internal structure of the social network are deeply analyzed, and the future development and variation of the social network are predicted.

A system and method facilitating speech detection and / or enhancement utilizing audio / video fusion is provided. The present invention fuses audio and video in a probabilistic generative model that implements cross-model, self-supervised learning, enabling rapid adaptation to audio visual data. The system can learn to detect and enhance speech in noise given only a short (e.g., 30 second) sequence of audio-visual data. In addition, it automatically learns to track the lips as they move around in the video.

The utility model relates to a DBN (Deep Belief Network) based ADHD (Attention Deficit Hyperactivity Disorder) discriminatory analysis method. The ADHD discriminatory analysis method comprises the following steps: step 1, pre-processing; step 2, characteristic extracting and classifying: depending on the DBN that is formed by stacking RBMs (Restricted Boltzmann Machines), classified and reversely adjusted in a layer-by-layer manner through softmax finally. The targets of the RBMs in layer-by-layer training are to maximize the likelihood function of the probability function, to introduce in the comparison divergence, and to update the weight function, so that the hidden layer becomes the approximate representation of the visible layer, the hidden layer of the first layer serves as the visible layer of the second layer, by parity of reasoning, the RBM layers of the DBN are obtained, and the last hidden layer is adopted as the input of the softmax to obtain the corresponding output, namely, the classification. The adopted DBN is a probability generative model, is formed by stacking the multiple RBMs with the hidden layers and the visible layers, simulates the layer-by-layer abstract characteristic process when the human brain processes signals, and abstracts the equivalent characteristic expression of the original signals to apply in the field of ADHD classification.

The invention relates to a friend recommendation method based on friend relationship spread in a social network, and belongs to the technical field of computer data mining. The method includes the steps that a potential friend relationship spread network is created for each self node; in each time period, a social behavior probability generative model is built, wherein social behaviors are perceived by interests, candidate middlemen and candidate interest fields are updated in an iterative mode, and the side-to-side weight in the potential friend relationship spread network is set according to the iterative result; as for each self node, the probability of choosing every friend to be the middleman is calculated on the partial layer of the potential friend relationship spread network in a random walking mode, the probability of choosing every node to be new friends is calculated on the whole layer of the potential friend relationship spread network in a random walking mode, and then the new friends with the largest probability are added in a recommended friend list. By constructing the potential friend relationship spread network, reasons and the structure of the social network are deeply analyzed from the angle of user behaviors, and then accurate friend recommendation services are provided.

The invention relates to a wind power climbing event probability scene prediction method. According to the method, distribution characteristic consistency and time sequence characteristic consistency are achieved through building of a multi-objective fitness function of a cumulative density function and a high-order moment autocorrelation function. On the basis of a genetic algorithm of the multi-objective fitness function, iterative optimization is conducted on probability generation model parameters to obtain a large number of prediction scenes, and climbing event probability characteristics excavated from scene capture zones are used for evaluating the prediction method. Actual data of some aboard wind field are selected for conducting example calculation and statistical analysis, the result shows that the statistical result of the multi-objective function is more accurate than that of a single-objective function, the probability scene prediction method can accurately estimate the characteristic quantity of a climbing event, the correctness of the method is proved, and the method can provide guidance for a probability scene generation method and a climbing event prediction model.

The invention discloses a depth generative adversarial method for underwater acoustic signal denoising, and belongs to the technical field of underwater acoustic signal denoising. The method comprisesthe following steps: firstly, carrying out sampling and feature extraction on an original underwater acoustic signal, then sending the extracted signal into a Gaussian restricted Boltzmann machine, and carrying out semi-supervised pre-training on a probability generation model; and finally, constructing a deep generative adversarial model, sending data generated in the probability generation model and a real label data stream into a Bernoulli restricted Boltzmann machine adversarial model, and performing supervised training. According to the method, a generative adversarial model is introduced into a restricted Boltzmann probability model according to the feature extraction characteristics of underwater acoustic signals, so that the problems of strong dependence and over-fitting of a restricted Boltzmann machine in the training process caused by complex signals carried by underwater sound are effectively eliminated, and the self-applicability of the training model is higher.

The present invention discloses a hyperspectral and multispectral image rapid fusion method. The method comprises the implementation steps of: (1) establishing a training set and a test set; (2) inputting a spectrum image; (3) establishing a probability generation model; (4) training a probability generation model; and (5) obtaining a fusion spectrum image. The hyperspectral and multispectral image rapid fusion method overcomes the problems that a representational capacity is weak, the model is not stable and the operation time is long in the prior art, generated hyperspectral image with highresolution comprises more information, a space transfer matrix without priori conditions is employed to allow the model to be more stable and allow the fusion speed to be faster, and therefore the hyperspectral and multispectral image rapid fusion method is a high-speed and efficient hyperspectral image fusion processing method.

The invention discloses a recommendation algorithm combining user comments and scoring information. The recommendation algorithm is implemented by the following specific steps: constructing a probability generation model for discovering potential topic dimensions in a user comment text; constructing a recommendation objective function based on the combination of a user score matrix decomposition model and a topic discovery model; and realizing product recommendation prediction based on the user comment text and the scoring data through iterative computation of the target function. According tothe algorithm, the comment information of the user is fully considered, the potential theme distribution in the comment text is utilized, the user scoring data and the user comment text are combined,and the cold start problem in a recommendation system is effectively solved; and meanwhile, compared with a method of separately considering two data sources, the method can more accurately perform score prediction, and is particularly suitable for score prediction of new products and new users.

The invention discloses an IPTV user behavior mode mining method based on a TDC-LDA model, and the method comprises the steps: firstly, assuming the generation process of a program watched in a specific behavior mode of a user, a watching time point and a watching time length by using the idea of a probability generation model, and building the TDC-LDA model; secondly, carrying out model fitting in a Gibbs sampling mode, and deducing a sampling formula of hidden variable condition probability and a calculation formula of a user-behavior mode matrix, an interest theme-program matrix and a timeperiod-time point matrix; then, according to a Gibbs sampling process of the TDC-LDA model, obtaining a behavior mode matrix, an interest theme-program matrix and a time period-time point matrix of each user; and finally, making recommendation to the user through the viewing behavior pattern distribution of the user. The invention provides a novel TDC-LDA user viewing behavior pattern mining model, and viewing duration information is added on the basis of an existing cLDA model.

One embodiment of the present invention provides a system that merges similar clusters of conceptually-related words in a probabilistic generative model for textual documents. During operation, the system receives a current model, which contains terminal nodes representing random variables for words and contains cluster nodes representing clusters of conceptually related words. Nodes in the current model are coupled together by weighted links, wherein if a node fires, a link from the node to another node causes the other node to fire with a probability proportionate to the weight of the link. Next, the system determines whether cluster nodes in the current model explain other cluster nodes in the current model. If two cluster nodes explain each other, the system merges the two cluster nodes to form a combined cluster node.

The invention discloses a POI recommendation method based on heterogeneous network, as to the technical field recommend by the POI, a new probability generation model of comment sentences is established, this model can deal with heterogeneous data across websites, and consider the influence of geographic region division, user community division, review content and user interaction behavior on POIprediction results, and solve the problem of data sparseness and low user trust in the recommendation process, which can significantly improve the accuracy of POI recommendation.

The invention discloses a topic model based hierarchical classification method and system for microblog user emotions. The classification method comprises the following steps: S1, microblog content is acquired and preprocessed, and to-be-classified words are obtained, wherein the to-be-classified words are one or more kinds of adverbs, verbs and adjectives; S2, the to-be-classified words are subjected to feature dimension reduction; S3, the to-be-classified words after feature dimension reduction are subjected to emotion classification of the microblog content according to a hierarchical classification model, and nodes on all layers of the hierarchical classification model are words representing the certain emotions. The topic model based hierarchical classification method has the following advantages: a hierarchical classification framework is designed, LDA (latent Dirichlet allocation) as a probability generation model is adopted to describe data according to the structural characteristic of short text of the microblog content, and further, feature dimension reduction and extraction are performed. By means of introduction of modules meeting text characteristics, the classification accuracy is improved, and better classification result is obtained.

Certain aspects of the present disclosure are directed to methods and apparatus for predicting subject motion using probabilistic models. One example method generally includes receiving training data comprising a set of subject pose trees. The set of subject pose trees comprises a plurality of subsets of subject pose trees associated with an image in a sequence of images, and each subject pose tree in the subset indicates a location along an axis of the image at which each of a plurality of joints of a subject is located. The received training data may be processed in a convolutional neural network to generate a trained probabilistic model for predicting joint distribution and subject motion based on density estimation. The trained probabilistic model may be deployed to a computer vision system and configured to generate a probability distribution for the location of each joint along the axis.

The invention discloses a super-resolution image processing method based on a probability generation model. The method comprises: (1), inputting an image; (2), obtaining a training sample; (3), obtaining a testing sample; (4), training a probability generation model; and (5), testing the probability generation model. According to the invention, compared with the common method, the super-resolutionimage processing method enables the generated high-resolution image to have higher resolution than that generated by the common method and to have more information; and with the prior information ofthe probability generation model and information implied by a hidden variable, the image super-resolution speed is increased. The method is a high-efficiency super-resolution image processing method.

The invention discloses a multispectral and high spectral image super-resolution method based on a probability generation model. The multispectral and high spectral image super-resolution method basedon a probability generation model includes the steps: 1) establishing a data set; 2) establishing a probability generation model; 3) inputting a spectral image; 4) maximizing the joint probability distribution logarithm; and 5) obtaining a spectral image after super-resolution. The multispectral and high spectral image super-resolution method based on a probability generation model overcomes theproblem that the prior art uses the prior information of the estimated spatial transformation matrix to make the model parameters inaccurate and the shallow model have poor representation ability, sothat the model parameters of the multispectral and high spectral image super-resolution method and the super-resolution high spectral image can be more accurate, and is an efficient high spectral image super-resolution method.

The invention relates to a method for building a friend relationship transitive tree in a social network, which belongs to the technical field of computer data mining. The friend relationship transitive tree is built for each self node, and self nodes and friend nodes are added. A social behavior probability generating model is set up for each time period, the prior parameters of the intermedium preference probability distribution and friend relationship strength distribution of the self nodes and the candidate intermediaries of each self node and new friend nodes are initialized, the candidate intermediaries is continuously interated and updated, a candidate intermediary z with the most sampling frequency is designated as the intermediary between a self node u and a new friend node y, and z to y are added into the social network of the self node u. The method models user behaviors in the social network on the basis of friend relationship transitivity; and a flat social network layering presentation method is provided, so that the reason and an internal structure of the social network are deeply analyzed, and the future development and variation of the social network are predicted.

The invention discloses a classified identification algorithm of ship targets based on the manifold distance characteristic. According to the algorithm, the similarity measuring standard among target objects is provided on the basis of the distance weight among samples represented by weighted corrected local anti-entropy operators. Compared a traditional distance measuring standard, the similarity measuring standard can reflect the practical space distribution of the objects more effectively; and neighbor samples are preprocessed and screened based on the standard, the classification efficiency of a probability generation model is improved, and the real-time processing requirement for ship target classified identification is met.

The invention discloses a POI recommendation method based on a heterogeneous network, relates to the technical field of POI recommendation, and establishes a new comment statement probability generation model. The model can handle cross-site heterogeneous data, and simultaneously considers geographical area division, The impact of user community division, comment content, and user interaction behavior on POI prediction results solves the problem of data sparseness and low user trust in the recommendation process, and can significantly improve the accuracy of POI recommendation.

The invention provides a DBN model optimization method based on a PSO algorithm. The DBNs is a probability generation model. The method is opposite to a neural network of a traditional discriminationmodel. The generation model is used for establishing joint distribution between observation data and labels; According to the invention, both P (Object | Label) and P (Label Object) are evaluated, andthe discrimination model only evaluates the P (Label Object). Wherein the DBNs is composed of a plurality of Restricted Boltzmann Machanes layers, and the type of a typical neural network is shown inthe figure. These networks are restricted into a visual layer and a hidden layer, there being connections between the layers, but there is no connections between the units within the layers. Hidden layer units are trained to capture correlation of high-order data represented at a visual layer. In order to make prediction of the DBN model more accurate, optimization needs to be conducted on the DBN, and the optimization comprises the number of layers, the number of units of each layer, the learning rate and the like. Therefore, the invention provides a DBN model optimization method based on aPSO algorithm. the optimal unit number and the RBM optimal learning rate of the input layer and the hidden layer are searcjed by adopting a parallel PSO (Particle Swarm Optimization) algorithm, and parallelization on a Spark platform is carried out based on a MapReduce principle.

The utility model relates to a DBN (Deep Belief Network) based ADHD (Attention Deficit Hyperactivity Disorder) discriminatory analysis method. The ADHD discriminatory analysis method comprises the following steps: step 1, pre-processing; step 2, characteristic extracting and classifying: depending on the DBN that is formed by stacking RBMs (Restricted Boltzmann Machines), classified and reversely adjusted in a layer-by-layer manner through softmax finally. The targets of the RBMs in layer-by-layer training are to maximize the likelihood function of the probability function, to introduce in the comparison divergence, and to update the weight function, so that the hidden layer becomes the approximate representation of the visible layer, the hidden layer of the first layer serves as the visible layer of the second layer, by parity of reasoning, the RBM layers of the DBN are obtained, and the last hidden layer is adopted as the input of the softmax to obtain the corresponding output, namely, the classification. The adopted DBN is a probability generative model, is formed by stacking the multiple RBMs with the hidden layers and the visible layers, simulates the layer-by-layer abstract characteristic process when the human brain processes signals, and abstracts the equivalent characteristic expression of the original signals to apply in the field of ADHD classification.

The invention relates to a friend recommendation method based on friend relationship spread in a social network, and belongs to the technical field of computer data mining. The method includes the steps that a potential friend relationship spread network is created for each self node; in each time period, a social behavior probability generative model is built, wherein social behaviors are perceived by interests, candidate middlemen and candidate interest fields are updated in an iterative mode, and the side-to-side weight in the potential friend relationship spread network is set according to the iterative result; as for each self node, the probability of choosing every friend to be the middleman is calculated on the partial layer of the potential friend relationship spread network in a random walking mode, the probability of choosing every node to be new friends is calculated on the whole layer of the potential friend relationship spread network in a random walking mode, and then the new friends with the largest probability are added in a recommended friend list. By constructing the potential friend relationship spread network, reasons and the structure of the social network are deeply analyzed from the angle of user behaviors, and then accurate friend recommendation services are provided.