134 results about "Latent variable" patented technology

A mechanism for compiling a generative description of an inference task into a neural network. First, an arbitrary generative probabilistic model from the exponential family is specified (or received). The model characterizes a conditional probability distribution for measurement data given a set of latent variables. A factor graph is generated for the generative probabilistic model. Each factor node of the factor graph is expanded into a corresponding sequence of arithmetic operations, based on a specified inference task and a kind of message passing algorithm. The factor graph and the sequences of arithmetic operations specify the structure of a neural network for performance of the inference task. A learning algorithm is executed, to determine values of parameters of the neural network. The neural network is then ready for performing inference on operational measurements.

Embodiments of the present invention include methods and systems for predicting the likelihood of topics appearing in a set of data such as text. A number of latent variable methods are used to convert the data into a set of topics, topic values and topic profiles. A number of time-course methods are used to model how topic values change given previous topic profiles, or to find historical times with similar topic values and then projecting the topic profile forward from that historical time to predict the likelihood of the topics appearing. Embodiments include utilizing focus topics, such as valence topics, and data representing financial measures to predict the likelihood of topics. Methods and systems for modeling data and predicting the likelihood of topics over other dimensions are also contemplated.

The disclosed terrain model is a generative, probabilistic approach to modeling terrain that exploits the 3D spatial structure inherent in outdoor domains and an array of noisy but abundant sensor data to simultaneously estimate ground height, vegetation height and classify obstacles and other areas of interest, even in dense non-penetrable vegetation. Joint inference of ground height, class height and class identity over the whole model results in more accurate estimation of each quantity. Vertical spatial constraints are imposed on voxels within a column via a hidden semi-Markov model. Horizontal spatial constraints are enforced on neighboring columns of voxels via two interacting Markov random fields and a latent variable. Because of the rules governing abstracts, this abstract should not be used to construe the claims.

A method for human face detection that detects faces independently of their particular poses and simultaneously estimates those poses. Our method exhibits an immunity to variations in skin color, eyeglasses, facial hair, lighting, scale and facial expressions, and others. In operation, we train a convolutional neural network to map face images to points on a face manifold, and non-face images to points far away from that manifold, wherein that manifold is parameterized by facial pose. Conceptually, we view a pose parameter as a latent variable, which may be inferred through an energy-minimization process. To train systems based upon our inventive method, we derive a new type of discriminative loss function that is tailored to such detection tasks. Our method enables a multi-view detector that can detect faces in a variety of poses, for example, looking left or right (yaw axis), up or down (pitch axis), or tilting left or right (roll axis). Systems employing our method are highly-reliable, run at near real time (5 frames per second on conventional hardware), and is robust against variations in yaw (±90°), roll (±45°), and pitch (±60°).

The invention discloses a method for 3D reconstruction and texture generation from single-view face based on multi-task learning, belonging to the computer vision field. The method comprises the following steps of: selecting a special viewpoint for rendering a three-dimensional human face model; generating depth map and texture map as truth value data from special viewpoint; designing an integrated learning and coding network based on feature sharing of depth information and texture information; designing the branch decoding network to recover the depth map from the shared features, and recovering the depth map, designing the mutual information with shared features as latent variables to maximize the generation of antagonistic network, and restoring the texture unwrapping map, adjusting the proportion of each task loss function and training the model; interpolating the depth map and using texture map to reconstruct the 3D face mesh model with texture details. The invention utilizes multi-task learning to carry out single-view face three-dimensional reconstruction, texture generation and style migration, and has the advantages of high speed, low cost and the like.

The invention relates to a power system measurement missing value reconstruction method based on depth learning. The reconstruction method comprises the following steps: cleaning and selecting the measured historical data from SCADA to obtain a training set of a neural network; an improved generative countermeasure network based on Wasserstein distance being constructed and trained; Adam being chosen as the optimizer to optimize the latent variable of WGAN network, and the improved GAN based on Wasserstein distance being obtained; the measurement data with missing values and corresponding binary mask matrix being inputted into the WGAN network model, and the reconstructed measurement data being finally obtained. The method can be applied to data cleaning and correction or to reconstructionof lost measurements when the system is attacked by communication. As the neural network of the invention automatically learn the complex spatio-temporal relationship which is difficult to be explicitly modeled such as the correlation between measurement and the load fluctuation rule, the feasibility of taking the reconstructed data as a pseudo measurement is ensured.

Methods and systems for modeling user arrival and choice in the context of off-street parking solutions. A first component models the arrival and duration of stay of users as a function of time, taking into account different user profiles (or “clusters”), captured by a latent variable. A second component provides a ranking function (for each user cluster), wherein the input features describing the “choice” constitute status variables associated different car park(s), and the output constitutes a preferred car park and a pricing scheme. The system simulates different user behaviors by assuming some standard groups of users will behave similarly. Groups of users or user profiles are learned automatically. The profiles are then employed as a key element for automatically learning a decision function of parking users, and automatically learning one decision function per profile.

A method and system are disclosed for generating a list of trips on an associated transportation network, the list ranked in accordance with time-dependent modeling of passenger preferences. User preferences of choosing a specific public transportation service or change point are modeled by a set of latent variables. Any actual trip on the network is converted into a set of pairwise preferences implicitly made by the passenger during the trip. Sequences of services matrices and change points matrices from the retrieved set of trips and non-negative factorization of the services and change points matrices is performed to smooth the matrices. The set of pairwise preferences are used to learn a ranking function and the output of a journey planner is re-ranked using the ranking function.

A data source may provide a plurality of time-series measurements that represent normal operation of a cyber-physical system (e.g., in substantially real-time during online operation of the cyber-physical system). A stateful, nonlinear embedding computer may receive the plurality of time-series measurements and execute stateful, nonlinear embedding to project the plurality of time-series measurements to a lower-dimensional latent variable space. In this way, redundant and irrelevant information may be reduced, and temporal and spatial dependence among the measurements may be captured. The output of the stateful, nonlinear embedding may be utilized to automatically identify underlying system characteristics of the cyber-physical system. In some embodiments, a stateful generative adversarial network may be used to achieve stateful embedding.

An adaptive computer-implemented e-learning system and method able to deliver a learning approach that customizes and adapts to each learners' skills and knowhow. A computer-implemented cognitive diagnosis process for predictions on skills and knowledge of a learner comprising the steps of: building a domain ontology; establishing a items bank; referencing the items bank using elements from the domain ontology elements (semantic-grounding of the items); establishing a training database of learners answers; estimating Cognitive Diagnostic Model (CDM) parameters and determining an appropriate model to be used; running the model to estimate a learner level on non-abstract latent variables; applying a Mastery of domain Knowledge and competency Skill (MKS) approach to estimate the learner level on abstract latent variables, or extracting relevant hierarchies from the ontology and use the CDM appropriate model to estimate those attributes; and building a diagnosis report from the results of steps 6) and 7).

A system for predicting variability of travel time for a trip at a particular time may utilize a machine learning model including latent variables that are associated with the trip. The machine learning model may be trained from historical trip data that is based on location-based measurements reported from mobile devices. Once trained, the machine learning model may be utilized for predicting variability of travel time. A process may include receiving an origin, a destination, and a start time associated with a trip, obtaining candidate routes that run from the origin to the destination, and predicting, based at least in part on the machine learning model, a probability distribution of travel time for individual ones of the candidate routes. One or more routes may be recommended based on the predicted probability distribution, and a measure of travel time for the recommended route(s) may be provided.

The present disclosure is directed towards methods and systems for determining multimodal image edits for a digital image. The systems and methods receive a digital image and analyze the digital image. The systems and methods further generate a feature vector of the digital image, wherein each value of the feature vector represents a respective feature of the digital image. Additionally, based on the feature vector and determined latent variables, the systems and methods generate a plurality of determined image edits for the digital image, which includes determining a plurality of set of potential image attribute values and selecting a plurality of sets of determined image attribute values from the plurality of sets of potential image attribute values wherein each set of determined image attribute values comprises a determined image edit of the plurality of image edits.

The invention discloses a universal blood glucose prediction method based on frequency band separation and data modeling, which comprises the steps that a human body subcutaneous blood glucose measurement signal is analyzed; the latent timing sequence dynamic characteristic of the human body subcutaneous blood glucose measurement signal is extracted; a frequency band separation threshold is defined; the subcutaneous blood glucose measurement signal is divided into a high frequency band and a low frequency band; timing sequence autocorrelation of a low frequency blood glucose signal is analyzed; and an autoegression blood glucose prediction model is established. According to the universal blood glucose prediction method, re-modeling for a new object is not required after the sufficient blood glucose measurement signals are acquired, and real-time blood glucose prediction can be performed by directly calling a prediction model of other individual, so that the modeling working capacity and the complexity are simplified greatly; the modeling cost can be lowered greatly; the prediction precision is improved due to the fact that a universal model adopts a method based on frequency band separation and latent variable modeling; and the universal blood glucose prediction method is easy to implement, and indicates a new direction for research of a blood glucose prediction modeling method.

The invention discloses a ring main unit cable core temperature soft measurement method based on a neighborhood preserving embedded regression algorithm. The ring main unit cable core temperature soft measurement method comprises that firstly, based on the local feature extraction strategy of the neighborhood preserving embedded algorithm, a regression optimization function which takes the internal temperature and the internal humidity of a ring main unit, the cable core current and the cable surface temperature as input, and takes the cable core temperature of a cable in the ring main unit as output is established, local features of input data and output data are reserved, and the maximum relationship between data is obtained; then based on lower-dimension latent variables of data, input and output features which construct data regression are obtained; and a cable core temperature soft measurement model is established. The ring main unit cable core temperature soft measurement method is advantaged in that by means of a data local feature extraction method, a traditional neighborhood preserving embedded algorithm is modified to be a regression model, and key variable information, of the ring main unit, which cannot be measured easily is obtained. According to the invention, the problem that the temperature of the cable core in the ring main unit cannot be measured easily is solved, and the accuracy and operability of on-line monitoring and fault locating of the ring main unit are improved.

In training data, a similarity matrix is generated for each of types of data corresponding to different kernels, and graph Laplacians are formed individually from the similarity matrices. An entire graph Laplacian is defined as linear combination of the individual graph Laplacians with coupling constants. Observation variables and latent variables associated therewith are assumed to form normal distributions, and the coupling constants are assumed to form a gamma distribution. Then, on the basis of a variational Bayesian method, a variance of the observation variables and the coupling constants can be figured out with a reasonable computational cost. Once the variance of the observation variables and the coupling constants are figured out, a predictive distribution for any input data can be figured out by means of a Laplace approximation.

To detect a statistical change-point that appears in time-series data with a high accuracy. A first model learning section 102 learns the occurrence probability distribution of time-series data 111 as a first statistical model (for example, a latent Markov model) defined by a finite number of variables including a latent variable. In the subsequent processing, the degree of a temporal change in the probability distribution is computed for each of the probability distribution of the entire first statistical model, its partial probability distribution (the probability distribution of the latent variable and conditional probability distribution contingent on the value of the latent variable), and the probability distribution in which the above plural probability distributions are linearly-combined with a weight. The change-point of the time-series data 111 is detected on the basis of the computed degree of the change.

The present invention is directed to a new joint estimation framework employing MAP estimation based on pixel-based latent variables for tissue types. The method combines the geometrical information described by latent MRF, statistical relation between tissue types and P-C coefficients, and Poisson noise models of PCD data, and makes possible the continuous Baysian estimation from detected photon counts. The proposed method has better accuracy and RMSE than the method using FBP and thresholding. The joint estimation framework has the potential to further improve the accuracy by introducing more information about tissues in human body, e.g., the location, size, and number of tissues, or limited variation of neighboring tissues, which will be easily formulated by pixel-based latent variables.

A method of detecting variance by regression model has the following steps. Step 1 is preparing the FDC data and WAT data for analysis. Step 2 is figuring out what latent variable effect of WAT data by Factor Analysis Step 3 is utilizing Principal Component Analysis to reduce the number of FDC variables to a few independent principal components. Step 4 is demonstrating how the tools and FDC data affect WAT data by Analysis of covariance model, and constructing interrelationship among FDC, WAT and tools. The interrelationship can point out which parameter effect WAT significantly. By the method, when WAT abnormal situation happened, it is easier for engineers to trace where the problem is.