68 results about "Local learning" patented technology

A document annotation method includes modeling data elements of an input document and dependencies between the data elements as a dependency network. Static features of at least some of the data elements are defined, each expressing a relationship between a characteristic of the data element and its label. Dynamic features are defined which define links between an element and labels of the element and of a second element. Parameters of a collective probabilistic model for the document are learned, each expressing a conditional probability that a first data element should be labeled with information derived from a label of a neighbor data element linked to the first data element by a dynamic feature. The learning includes decomposing a globally trained model into a set of local learning models. The local learning models each employ static features to generate estimations of the neighbor element labels for at least one of the data elements.

Systems and methods are disclosed for improving search results returned to a user from one or more search domains, utilizing query features learned locally on the user's device. A search engine can receive, analyze and forward query results from multiple search domains and pass the query results to a client device. A search engine can determine a feature by analyzing query results, generate a predictor for the feature, instruct a client device to use the predictor to train on the feature, and report back to the search engine on training progress. A search engine can instruct a first and second set of client devices to train on set A and B of predictors, respectively, and report back training progress to the search engine. A client device can store search session context and share the context with a search engine between sessions with one or more search engines. A synchronization system can synchronize local predictors between multiple client devices of a user.

Aspects of the present invention include object detection training systems and methods and using object detection systems and methods that have been trained. Embodiments presented herein include hybrid learning approaches that combine global classification and local adaptations, which automatically adjust model complexity according to data distribution. Embodiments of the present invention automatically determine model complexity of the local learning algorithm according to the distribution of ambiguous samples. And, embodiments of the local adaptation from global classifier avoid the common under-training problem for local classifier.

A tag recommendation for an item to be tagged is generated by: selecting a set of candidate neighboring items in an electronic social network based on context of items in the electronic social network respective to an owner of the item to be tagged; selecting a set of nearest neighboring items from the set of candidate neighboring items based on distances of the candidate neighboring items from the item to be tagged as measured by an item comparison metric; and selecting at least one tag recommendation based on tags of the items of the set of nearest neighboring items. The item comparison metric may comprise a Mahalanobis distance metric trained on the set of candidate neighboring items to correlate the trained Mahalanobis distance between pairs of items of the set of candidate neighboring items with an overlap metric indicative of overlap of the tag sets of the two items.

A tiered machine learning-based infrastructure comprises a first machine learning (ML) tier configured to execute within an enterprise network environment and that learns statistics for a set of use cases locally, and to alert deviations from the learned distributions. Use cases typically are independent from one another. A second machine learning tier executes external to the enterprise network environment and provides further learning support, e.g., by determining a correlation among multiple independent use cases that are running locally in the first tier. Preferably, the second tier executes in a cloud compute environment for scalability and performance.

The invention discloses a distributed machine learning method based on a local learning strategy, which is based on a parameter server architecture, is suitable for multi-machine cluster distributed machine learning of a data center, and is also suitable for client-cloud collaborative distributed machine learning taking a server as a cloud end, and a mobile phone or embedded equipment as a terminal. The method comprises the following steps: firstly, a server node accumulating a local gradient sum calculated by all working nodes to obtain a full gradient, and broadcasting the full gradient to all the working nodes; each working node updating the parameters for several times and then sending the local parameters to the server node; finally, the server node averaging the parameters collectedfrom the working nodes and broadcasting the average value as the latest parameter to all the working nodes; iterating the above process for multiple rounds until a convergence condition is achieved. The method is based on a local learning strategy, communication does not need to be carried out after parameters of the working nodes are updated every time, and therefore the communication expenditurein distributed machine learning is reduced.

Artificial neural networks (ANNs) are a distributed computing model in which computation is accomplished using many simple processing units (called neurons) and the data embodied by the connections between neurons (called synapses) and the strength of these connections (called synaptic weights). An attractive implementation of ANNs uses the conductance of non-volatile memory (NVM) elements to code the synaptic weight. In this application, the non-idealities in the response of the NVM (such as nonlinearity, saturation, stochasticity and asymmetry in response to programming pulses) lead to reduced network performance compared to an ideal network implementation. Disclosed is a method that improves performance by implementing a learning rate parameter that is local to each synaptic connection, a method for tuning this local learning rate, and an implementation that does not compromise the ability to train many synaptic weights in parallel during learning.

The invention discloses a satellite cluster configuration adjustment planning method for space target collaborative observation. According to the method, each satellite is regarded as an intelligent Agent with autonomous planning and decision-making capability, a coordination graph based on a cluster communication structure is established, and tasks optimized by a global optimal configuration adjustment strategy are decomposed to local learning of each satellite based on the communication coordination graph. According to the method, the top-layer planning of configuration adjustment and bottom-layer control are comprehensively considered; solving is carried out from the perspective of cluster cooperation planning; the global planning learning task is divided into the paired Agents according to the communication relationship, so that the autonomy and intelligence of the satellite cluster are enhanced, the on-orbit viability and the task execution level of the satellite cluster are improved, and a new idea and method are provided for on-orbit autonomous operation and management of the satellite cluster.

The invention provides a load distribution method for a power system based on multi-target distribution estimation, which mainly solves the problem of load distribution of the power system. The implementation steps are as follows: initializing a parent population; repairing infeasible individuals in the parent population; calculating the fitness value of the parent population; obtaining a first new population by local learning; obtaining a second new population by modeling and sampling; combining the two new populations to obtain a filial population; repairing infeasible individuals in the filial population; calculating the fitness value of the filial population; and combining the filial population with the parent population, obtaining a next iterated parent population by means of quick non-domination sequencing selection, and taking non-domination individuals in the next iterated parent population as a final load distribution scheme if reaching the maximum iterated times, otherwise, continuing iterating. The load distribution scheme obtained by the invention has the advantages that the amount of the discharge of pollution gas and the amount of the consumption of fuels of the power system are reduced, and more optional load distribution schemes are provided.

The invention provides a method for establishing a student growth portrait based on group sparse fusion hospital big data, and the method comprises: obtaining multi-source high-dimensional data of students, including student basic information data, school learning data, daily life data and hospital regular training practice data; performing data preprocessing on the multi-source high-dimensional data; compressing and storing the preprocessed sample data set based on a triple representation method in a sparse algorithm; grouping the feature tags with strong correlation by using a binary K-meansclustering algorithm, and constructing a student growth portrait tag system which is more suitable for actual conditions of students; combining the labels into different feature label groups, and acquiring clustering distribution information of the sample data feature label groups based on local learning; and under the guidance of clustering distribution information, obtaining feature weights bymeans of group sparse regression, wherein the feature weights are used for evaluating the importance of features, and selecting corresponding important feature tags. A student growth portrait label system is established according to the high-dimensional data, and a student growth portrait fused with hospital big data stereoscopicity is constructed. By implementing the method, the collected data can be more complete, the constructed student portrait index system is more stereoscopic, complete and accurate, and subjective randomness is avoided to a certain extent.

Systems and methods are disclosed for improving search results returned to a user from one or more domains, utilizing query features learned locally on the user's device. One or more domains can inform a computing device of one or more features related to a search query upon which to the computing device can apply local learning. A local search system can include a local database, a local search history and feedback history database, and a local learning system to identify features about query terms. The features can be learned from the user's interaction with both local search results and remote search results, without sending the user interaction information or other user identification information to a remote search engine. A locally learned feature can be used to extend a query, bias a query term, or filter query results.

A learning system and method applies computer network and achieves an online language learning through integration of local learning and remote companion oral practice. The local user selects a curriculum topic from a database for learning. A user-mating module establishes network connections according to the online user status. A curriculum-coding module encodes the characteristics of the curriculum contents and transfers the characteristic codes to the remote user site. A curriculum-decoding module at the remote user site receives and decodes the characteristic codes to generate correspondent curriculum pages. A page synchronization module returns and updates the curriculum pages for the users. Therefore, the local and remote users take oral practices on the curriculum pages through audio / video communication.

A document annotation method includes modeling data elements of an input document and dependencies between the data elements as a dependency network. Static features of at least some of the data elements are defined, each expressing a relationship between a characteristic of the data element and its label. Dynamic features are defined which define links between an element and labels of the element and of a second element. Parameters of a collective probabilistic model for the document are learned, each expressing a conditional probability that a first data element should be labeled with information derived from a label of a neighbor data element linked to the first data element by a dynamic feature. The learning includes decomposing a globally trained model into a set of local learning models. The local learning models each employ static features to generate estimations of the neighbor element labels for at least one of the data elements.

The invention discloses as ample classification method based on a weighted PTSVM (projection twin support vector machine), and the method comprises the following steps: respectively constructing in-class neighbor graphs Gs and an inter-class neighbor graph Gd in all sample classes and among different sample classes; calculating sample weights according to the in-class neighbor graphs Gs of all sample classes, and calculating the weighted mean center of each sample class; determining a reverse sample class, which is nearer to a specific sample class, according to the inter-class neighbor graph Gd, and constructing an optimization problem in a linear mode; solving a dual problem of the optimization problem, obtaining the decision hyperplanes of two classes of samples: xTw1+b1=0 and xTw2+b2=0, and carrying out the classification of unknown samples according to the decision hyperplanes, wherein w1 and w2 are respectively the projection axes of the first and second classes of samples, x represents samples in an n-dimensional vector space, and b1 and b2 respectively represent the biases of the decision hyperplanes of two classes of samples. The method improves the local learning capability of an algorithm to a certain extent, and greatly reduces the solving complexity of the algorithm.

The invention discloses a medical image registration method based on cross hill-climbing memetic quantum evolutionary computation, and mainly aims to solve the problem that in the prior art, the registration effect is poor. The method includes the steps that images are read in, and relevant parameters are initialized; an initial population is generated through a chaotic method; registration parameters corresponding to individuals in the population are computed; image conversion is conducted on the floating images according to the registration parameters to acquire the converted images; similarity between the converted images and a reference image is calculated, a local generation optimal registration parameter is looked for, and quantum revolution door update is conducted on all the individuals in the population according to the optimal registration parameter; cross hill-climbing memetic local learning is conducted on the optimal refrigeration parameter; whether an optimal individual meets a forgetting condition or not is judged, and forgetting operators are executed on the optimal individual; circulation conditions are controlled, if circulation is ended, the registration result is output. By the adoption of the method, the better registration result can be acquired, and the method can be used for registering the medical images.

The invention provides a robust acoustic scene recognition method based on local learning, and belongs to the technical field of sound signal processing. The robust acoustic scene recognition method comprises the steps: firstly, sound signals of different acoustic scenes are collected, and frequency domain feature extraction is conducted; extracted feature data are pre-processed; then the normalized data are subjected to mean value translation, and data augmentation is conducted through a mixup method; then a convolution neural network model is established according to the local learning thought, a training sample set after data augmentation is input into the model to be trained, and the trained model is obtained; and finally, a to-be-recognized sample is sequentially subjected to frequency domain feature extraction data pre-processing, and input into the trained model to be recognized, and the acoustic scene recognition result is obtained. The problem that the acoustic scene recognition accuracy is low under the conditions of audio channel mismatch and the unbalanced number of different channel samples is solved; and the robust acoustic scene recognition method can be suitable foracoustic scene recognition with various channels and the unbalanced number of the different channel samples.

The invention discloses a message forwarding method and device in a VXLAN (virtual extensible local area network). The method includes: when local learning of the MAC address of a source mainframe fails, a first VTEP transmits the information of the source mainframe to a second VTEP to allow the second VTEP to save the correspondence among the MAC address of the source mainframe, the information of the source mainframe and a VXLAN tunnel interface, wherein the information of the source mainframe is used for identifying the local outlet interface of the MAC address of the source mainframe; the first VTEP receives the data message of the second VTEP, wherein the data message carries the information of the source mainframe; the first VTEP forwards the data message to the source mainframe through the local outlet interface according to the target MAC of the data message and the source mainframe MAC address in the information of the source mainframe. By the method, broadcast flooding caused by source MAC address learning failure can be reduced.

The invention discloses a local learning feature weight selection-based medical data classification method. The method comprises the following steps of: firstly obtaining attributes of samples according to a training sample set, and calculating weight vectors corresponding to attributes according to the attribute values by utilizing a gradient descent weight updating manner so as to ensure the astringency, achieve a stopping criterion of an algorithm in a relatively high speed, shorten the calculation time and reduce the calculation complexity; and carrying out feature selection according to the calculated weight vectors so as to obtain an optimum feature set, standardizing to-be-assessed data samples and carrying out feature selection in an optimum feature subset, and classifying the to-be-assessed data sample after the feature selection so as to realize dimensionality reduction of the data samples. According to the method provided by the invention, the calculation complexity is reduced and the calculation time is shortened while the dimensionality reduction is realized. The invention furthermore provides a local learning feature weight selection-based medical data classification device which also can realize above technical effects.

The present invention relates to an on line colloquial language learning system of real person for integrating computer local learning and far-end training works. Said system includes lesson data base, on-line paring module, lesson coding module, page synchronization module and voice / video equipment, etc. Said invention also provides the steps for implementing said invented colloquial language learning system.

The invention discloses a sparse coding method based on local and global regularization, pertaining to the technical field of sparse coding (SC). The method utilizes local regression regularization tofind out a potential geometric structure of data. Concretely speaking, the overall data space is divided into a large number of partial regions. Each sample is linearly represented by a sample of a partial region to which each sample belongs, which is called as a local learning conception. On the basis of the local learning conception, a kernel global regression method is utilized for grabbing aglobal geometric structure of data. Therefore, partial and global regularization are utilized for obtaining an inherent geometric structure of data.

The invention relates to the technical field of multi-party data cooperation, and provides a model training method and device based on a privacy data set. The method comprises the following steps: training a server-side model based on a public data set and a real label corresponding to the public data set; obtaining a first model output sent by each client, wherein the first model output is obtained by inputting the public data set into the local learning model, and the local learning model is obtained based on the privacy data set and corresponding label training; training a server side model based on the corresponding public data output by each first model; inputting the public data set into the server-side model to obtain a second model output; and sending the second model output to each client, so that each client performs retraining of the local learning model based on the second model output and the public data set. Therefore, on the premise of avoiding leakage of the privacy data set, model training is carried out by taking the privacy data set as a part of training samples based on knowledge distillation and knowledge fusion.

The invention discloses a distributed learning privacy protection method based on differential privacy, the method is applied to n user nodes in a network, each user has a group of independently distributed data samples, and the method comprises the following steps: S1, an initialization stage; S2, a user node local learning stage; S3, the user node acquires neighbor node information and updates the neighbor node information; S4, a noise disturbance adding stage; S5, broadcasting stage. According to the method, the privacy protection problem in current distributed learning can be solved so that the user node updates own local parameters through the neighbor node and sends the parameters subjected to noise processing to the neighbor node, and therefore, personal sensitive data of the user can be protected from being leaked in a decentralized network environment.

The invention discloses a steganalysis method based on local learning, comprising: constructing a training sample database including positive and negative samples; specific to any sample to be detected, searching for most similar K positive and negative sample couples in the sample database to form a local training set; performing classifier training learning based on the local training set, and adding the constraint of the positive and negative sample couples in a learning process to obtain an optimum classifier by using an optimization algorithm; and discriminating and classifying samples to be detected by the clarifier to obtain detection results of whether the samples to be detected performs steganography. The steganalysis method takes full advantage of local learning which has the advantages of better overcoming the problem of great intra-class change, reducing noise influence and using less priori knowledge, improves steganalysis results, and can be applied to a steganalysis algorithm analysis system based on pattern recognition.

The invention discloses a federated learning scheduling method based on calculation unloading in a heterogeneous network, which comprises the following steps: firstly, a macro base station trains a machine learning model by using local data of a user, and the user can unload the data to an edge server near a small base station for calculation; then, the precision of federal learning and energy consumption and time delay generated in operation and communication are comprehensively considered, an optimization problem is established, and the optimization problem is divided into three sub-problems of optimizing local learning precision and unloading decision, user calculation frequency and user resource block allocation. Through simulation, the effects obtained under different parameter conditions are analyzed. Simulation results show that the method provided by the invention can effectively reduce the overhead in the training process.

The invention discloses a feature selection method based on a group sparse norm and local learning. The method is used for selecting key features representing dataset structures in data to relieve a ''dimension disaster''. By the adoption of an unsupervised learning mode, the method can be applicable to a dataset lacking part of tags. According to the method, first, tags of data lacking tags are inferred through local learning; and then some features most capable of distinguishing different tag data globally are found out on a regressive architecture plus the group sparse norm. The method hasthe advantages that a manifold structure of the data is saved through local learning, a group structure of the data is saved through the group sparse norm, therefore, compared with a traditional feature selection method, the features most capable of maintaining a local structure of the original dataset can be selected through combination of local learning and the group sparse norm.

The invention provides a transverse federal learning method and device and a storage medium. The method mainly comprises the following steps that: a central server selects an initial federal learning model and parameters and issues the initial federal learning model and parameters to each client; each client simultaneously starts model training based on respective local training data so as to obtain a local learning model; sending a local learning model to a next client for training by adopting a cyclic communication mode among the clients, updating a cyclic communication sequence after all the clients finish training, and continuing to transmit the model and training until a set training frequency is reached; and each client sends the finally trained model to the central server for aggregation, and then the federal learning model is updated by using an aggregation result until the loss function is converged, so that the training is completed. According to the method, when a transmission model or gradient parameters are trained, annular model propagation is firstly carried out, and then star propagation is carried out to a server, so that a new transverse federated learning model training architecture is constructed.

A method for gaze estimation. The method includes processing training data and determining one or more local-learning base gaze estimation model based on the training data. The local-learning base gaze estimation model(s) can be used for determining one or both of: 2D gaze points in a scene image and 3D gaze points in scene camera coordinates.

The invention provides a non-negative matrix factorization clustering method based on dual local learning. The non-negative matrix factorization clustering method based on dual local learning comprises the steps of S10, selecting a to-be-classified data matrix V, a cluster number a1 and a cluster number a2 according to a to-be-clustered image; S20, constructing an objective function O according tothe data matrix V; S30, outputting a class result by using an iteration method according to the objective function O; and S40, clustering the to-be-clustered images according to the class result. According to the clustering method, double-structure learning is combined, a collaborative clustering problem is converted into a non-negative matrix v problem with orthogonal constraint, the complexityof the problem is simplified, the method has representativeness and universality, the complexity is low, the operation speed in the clustering process is greatly increased, and the clustering efficiency is improved.

The embodiment of the invention provides a model learning method, device and system based on private data protection. The method comprises the step that non-centralized computing platform member objects in the alliance send model parameters of a learning model which is trained locally to centralized computing platform member objects in the alliance; model parameters of learning models of at leasttwo member objects are centrally calculated in the alliance obtained by the platform member objects through multi-party security calculation, and determining target model parameters; the centralized computing platform member objects send the target model parameters to non-centralized computing platform member objects in the alliance; and the member objects of the non-centralized computing platformin the alliance adjust the local learning model based on the target model parameters.