163 results about "Learning problem" patented technology

The invention relates to a deep learning-based question and answer matching method. The method comprises the following steps of: 1) sufficiently learning word orders and sentence local features of a question text and an answer text by utilizing two underlying deep neural networks: a long short-term memory network LSTM and a convolutional neural network CNN; and 2) selecting a keyword with best semantic matching on the basis of a pooling manner of an attention mechanism AM. Compared with existing methods, the method has the advantages of being in low in feature engineering workload, strong in cross-field performance and relatively high in correctness, and can be effectively applied to the fields of commercial intelligent customer service robots, automatic driving, internet medical treatment, online forum and community question answering.

A system and method of problem determination and resolution utilizes enhanced problem classification, and effectively categorizes any problem a user experiences by leveraging all available data to recognize the specific problem. Historical problem data is labeled with the cause of that problem and is analyzed to learn problem patterns. The historical problem data is classified into a predefined hierarchical structure of taxonomies by using an incremental online learning algorithm. The hierarchical structure and learned patterns are utilized to recognize problems and generate the root cause of the problem when given a new set of monitoring data and log data.

The trend of analyzing big data in artificial intelligence requires more scalable machine learning algorithms, among which clustering is a fundamental and arguably the most widely applied method. To extend the applications of regular vector-based clustering algorithms, the Discrete Distribution (D2) clustering algorithm has been developed for clustering bags of weighted vectors which are well adopted in many emerging machine learning applications. The high computational complexity of D2-clustering limits its impact in solving massive learning problems. Here we present a parallel D2-clustering algorithm with substantially improved scalability. We develop a hierarchical structure for parallel computing in order to achieve a balance between the individual-node computation and the integration process of the algorithm. The parallel algorithm achieves significant speed-up with minor accuracy loss.

The aesthetic quality of a picture is automatically inferred using visual content as a machine learning problem using, for example, a peer-rated, on-line photo sharing Website as data source. Certain visual features of images are extracted based on the intuition that they can discriminate between aesthetically pleasing and displeasing images. A one-dimensional support vector machine is used to identify features that have noticeable correlation with the community-based aesthetics ratings. Automated classifiers are constructed using the support vector machines and classification trees, with a simple feature selection heuristic being applied to eliminate irrelevant features. Linear regression on polynomial terms of the features is also applied to infer numerical aesthetics ratings.

The trend of analyzing big data in artificial intelligence requires more scalable machine learning algorithms, among which clustering is a fundamental and arguably the most widely applied method. To extend the applications of regular vector-based clustering algorithms, the Discrete Distribution (D2) clustering algorithm has been developed for clustering bags of weighted vectors which are well adopted in many emerging machine learning applications. The high computational complexity of D2-clustering limits its impact in solving massive learning problems. Here we present a parallel D2-clustering algorithm with substantially improved scalability. We develop a hierarchical structure for parallel computing in order to achieve a balance between the individual-node computation and the integration process of the algorithm. The parallel algorithm achieves significant speed-up with minor accuracy loss.

Outlier detection methods and apparatus have light computational resources requirement, especially on the storage requirement, and yet achieve a state-of-the-art predictive performance. The outlier detection problem is first reduced to that of a classification learning problem, and then selective sampling based on uncertainty of prediction is applied to further reduce the amount of data required for data analysis, resulting in enhanced predictive performance. The reduction to classification essentially consists in using the unlabeled normal data as positive examples, and randomly generated synthesized examples as negative examples. Application of selective sampling makes use of an underlying, arbitrary classification learning algorithm, the data labeled by the above procedure, and proceeds iteratively. Each iteration consisting of selection of a smaller sub-sample from the input data, training of the underlying classification algorithm with the selected data, and storing the classifier output by the classification algorithm. The selection is done by essentially choosing examples that are harder to classify with the classifiers obtained in the preceding iterations. The final output hypothesis is a voting function of the classifiers obtained in the iterations of the above procedure.

The invention discloses a facial recognition method based on supervised neighborhood preserving embedding (SNPE) and a support vector machine (SVM). The method comprises a training process and a test process, and includes specifically the following procedures: a, a weight matrix for a given data sample group is constructed; b, according to the weight matrix obtained in the procedure a, a generalized characteristic vector problem about the data samples is solved and an embedding matrix projecting the data samples into a low-dimensional data space is searched; c, characteristic extraction is performed on the data samples with the embedding matrix to acquire characteristic data of the low-dimensional space; d, mode classification of the acquired characteristic data in the procedure c is conducted with the SVM to realize the type recognition of the data samples. The facial recognition method provided by the invention can solve the problems in the prior that the linear dimensionality reduction method cannot maintain well inter- and intra-type sample structures; the non-linear dimensionality reduction method has large computational amount; and the common classifier has over-learning and under-learning problems.

The invention relates to the field of ontology learning, in particular to an ontology concept and hierarchical relation generation method. The technical scheme includes that a PAM (probability aim model) is applied to extraction of ontology concept and hierarchical relation to improve effect of ontology concept and hierarchical relation learning so as to generate ontology concept more accurately and effectively. The method includes: establishing a PAM-based ontology concept and hierarchy generation model to effectively convert a domain ontology concept learning problem into a domain document set based statistic inference problem, and adopting a Gibbs sampling method to obtain a probability distribution feature vector; performing Wordnet-based semantic similarity calculation, generating a conception according to similarity correlation, and obtaining a set and hierarchical relation of ontology concepts. By the method, concept sets and hierarchical relation among concepts in domain ontology can be obtained more accurately and effectively.

The invention relates to a method for analyzing implicit type discourse relation based on hierarchical depth semantics, and belongs to the technical field of application of natural language processing. The method comprises the following steps of firstly, combining marked and unmarked corpuses, expanding the corpus training scale, and solving the problem of under-learning due to undersize corpus training scale; then, according to a certain rule, initializing a depth semantic vector of each corpus training hierarchy, sorting word pairs favorable for classification according to information gain value, and using the word pairs as subsequent feature selection basis; finally, designing a scoring function, combining the multiple hierarchial depth semantic information of to-be-classified discourse relation theory element pairs, utilizing the parameters of a nerve network training model, fitting a type tag of the implicit type discourse relation, and finding the model for furthest optimizing the performance, so as to complete the analysis of the implicit type discourse relation. The method has the advantages that the false judging of the traditional method based on discrete features is overcome; the analysis accuracy of the type tag of the implicit type discourse relation is improved; a user can quickly and accurately obtain the analysis result of the implicit type discourse relation.

The invention relates to a dynamic machine learning modeling method based on sample recommending and labeling and belongs to the technical field of machine learning. According to the method, a total sample set is clustered according to a CURE algorithm, the center sample and the representative sample of each cluster in results are recommended and labeled, and then, a model is established. The method has the beneficial effects that data sets of unknown categories are recommended and labeled when supervised learning problems are processed, and thus, the processing is effective, timesaving and laborsaving; and wrong samples are repeatedly trained by using machine learning, so that the model can more accurately judge the category of new data sets.

The invention discloses a robot simulation learning method based on a dynamic system model. Simulation of teaching movement of the robot is achieved through learning, specifically, the demonstration motion is modeled into a nonlinear dynamic system model through a gaussian mixture model, in addition, the stability of the motion model is guaranteed through the method with the additional stability constraint conditions, the parameter learning problem of the motion model is converted into a constraint optimization problem, so that a complete description of the motion model is obtained, and finally, the motion model obtained through learning is used as a control strategy to guide the robot to imitate the teaching motion. The method is used for teaching motion of target point fixation, the method has good stability, all the generated motion trails are converged to a target point, and has good expression capacity for simple and complex teaching movement, the generalization ability of the motion model is good, a motion track which is smooth and can be converged to a target can be generated outside the teaching movement range.

The invention discloses a mechanical arm autonomous grabbing method based on deep reinforcement learning and dynamic movement primitives. The mechanical arm autonomous grabbing method includes the following steps that firstly, a camera image assembly is installed, it is ensured that the recognition area is not shielded, grabbing target area images are preprocessed and sent to a deep reinforcementlearning intelligent agent as state information; secondly, a local strategy near-end optimization training model is established on the basis of the state and deep reinforcement leaning principle; thirdly, a new mixed movement primitive model is established by fusing the dynamic movement primitives and imitation learning; and fourthly, a mechanical arm is trained to autonomously grab objects on thebasis of the models. By means of the mechanical arm autonomous grabbing method, the problem that the mechanical arm joint movement based on traditional deep reinforcement learning is unsmooth can beeffectively solved, the learning problem of primitive parameters is converted into the reinforcement learning problem through combination with the dynamic movement primitive algorithm, and by means ofthe training method of deep reinforcement learning, the mechanical arm can complete the autonomous grabbing task.

The invention discloses a multi-label learning design method based on a hashing method. Through the combination of a hashing algorithm and a multi-label learning algorithm based on Bayesian statistics, the correlation between labels is effectively utilized so as to improve the predicting performance of a multi-label learning model, labels and neighbors of the labels are introduced to computation of the posterior probability through the characteristics of the neighbors, the correlation between the labels is fully considered, and the accuracy of the algorithms is improved; the problem that the label space in multi-label learning of large-scale data is higher in dimension and sparse is solved through an MinHash algorithm; the purpose of learning large-scale data is achieved by finding the neighbors through locality sensitive hashing (LSH), the neighbors can be rapidly and efficiently found, and the expandability of the multi-label learning algorithm is improved.

The invention discloses a knowledge learning and privacy protection based big-data user purchase intention predicating method which comprises following steps of: (1) performing normalization processing on a large number of historical data and a small number of current data; (2) grouping the data and establishing a training sample set; (3) counting user purchase intention probability of each group; (4) calculating group labels; (5) training the training set by using an improved support vector machine; (6) constructing a prediction function; (7) inputting to-be-predicted data into the predication function to obtain a prediction result. As the improved support vector machine is used in the method, the small number of current data set probability information and the large number of historical data set probability information are blended into a structural risk minimization learning framework, learning of knowledge in different periods is realized by virtue of constructing similar distance items among data, and accordingly, the knowledge learning and privacy protection based big-data user purchase intention predicating method which is applicable for learning problems of big samples is constructed.

The invention discloses a deep learning problem classification method combining a multi-level attention mechanism, and relates to the field of information processing technology. The deep learning problem classification method comprises the following steps: constructing an interrogative word vector set, wherein the interrogative word vector set includes interrogative word vectors and common word vectors including interrogative word information; extracting window mapping of interrogative sentences by convolution operation according to the interrogative word vector set; extracting sequential characteristics of the interrogative sentences according to the window mapping; and classifying the interrogative sentences according to the sequential characteristics. Through adoption of the method, semantic information of interrogative words in the interrogative sentences is enhanced. Moreover, a convolutional neural network and a long-term and short-term memory model are fused through the attention mechanism in deep learning, thereby effectively increasing the accuracy of problem classification.

The invention provides a method and a device for determining sample labels. The method for determining the sample labels comprises the following steps of extracting on-line feedback data to acquire initial model training data; extracting scene features from the initial model training data; adding the scene features into the initial model training data to acquire sample data; calculating the sample quality score of each group of sample data; labeling each group of sample data according to the corresponding sample quality score of the corresponding group of sample data; learning the labeled sample data to generate a model; and predicting labels of the sample data by using the model. In a multi-example learning problem, labor is reduced when the sample is labeled, a large amount of labor is saved, and the labeling efficiency can be improved.

Disclosed is a general learning framework for computer implementation that induces sparsity on the undirected graphical model imposed on the vector of latent factors. A latent factor model SLFA is disclosed as a matrix factorization problem with a special regularization term that encourages collaborative reconstruction. Advantageously, the model may simultaneously learn the lower-dimensional representation for data and model the pairwise relationships between latent factors explicitly. An on-line learning algorithm is disclosed to make the model amenable to large-scale learning problems. Experimental results on two synthetic data and two real-world data sets demonstrate that pairwise relationships and latent factors learned by the model provide a more structured way of exploring high-dimensional data, and the learned representations achieve the state-of-the-art classification performance

Existing methods for analyzing person-specific ‘connectomes’ are not computationally equipped for scalable, flexible, and integrated processing across multiple network resolutions and drawing from disparate data modalities—a major roadblock to utilizing ensembles and hierarchies of connectomes to solve person-specific machine-learning problems. The processes implemented in software described herein consists of an end-to-end pipeline for deploying ensembles and hierarchies of network-generating workflows that can utilize multimodal, person-specific data to sample networks, extracted from that data, across a grid of network-defining hyperparameters. In essence, this pipeline enables users to perform ensemble sampling of connectomes for given individual(s) based on any input phenotypic datatype, constructed from any data modality or hierarchy of modalities at any scale, and based on any set of network-defining hyperparameters.

The aesthetic quality of a picture is automatically inferred using visual content as a machine learning problem using, for example, a peer-rated, on-line photo sharing Website as data source. Certain visual features of images are extracted based on the intuition that they can discriminate between aesthetically pleasing and displeasing images. A one-dimensional support vector machine is used to identify features that have noticeable correlation with the community-based aesthetics ratings. Automated classifiers are constructed using the support vector machines and classification trees, with a simple feature selection heuristic being applied to eliminate irrelevant features. Linear regression on polynomial terms of the features is also applied to infer numerical aesthetics ratings.

The invention discloses an image classification method based on a multi-task multi-instance support vector machine. The method comprises the following steps: establishing T learning tasks for T groups of images; performing multi-instance processing on images of the T learning tasks; constructing one class package for each category of images in the T tasks; establishing an Euclidean distance formula from instances in the class packages to multi-instance packages; constructing instance distance vectors from the class packages to the multi-instance packages; establishing a weight Euclidean distance formula from the class packages to the multi-instance packages; performing constraining to enable distances from the multi-instance packages to their corresponding categories to be smaller than distances to other categories; establishing an optimization problem of the multi-task multi-instance support vector machine; converting the optimization problem into a problem of a conventional single-task single-instance support vector machine problem; and solving the optimization problem of the support vector machine. According to the method related to by the invention, the weight Euclidean distance formula is optimized, through performing instance processing on the images, a learning problem of the multi-task multi-instance support vector machine is established, an ideal weight is optimized, and thus performance of an image classifier is improved.

The invention provides a method and system for determining the characteristic significance of a machine learning sample. The method comprises that (A) a historical data record which comprises a mark and at least one piece of attributed information related to a machine learning problem is obtained; (B) the obtained historical data record is utilized to train at least one characteristic pool model which is used to provide a prediction result related to the machine learning problem on the basis at least part of characteristics; and (C) an effect of the at least one characteristic pool model is obtained, and significance of the different characteristics is determined according to the obtained effect of the at least one characteristic pool model. In the step (B), at least one continuous characteristic among at least part of the characteristics is used to implement a discrete operation and further to train the characteristic pool model. Via the method and system, significance of the characteristics in the machine learning sample can be determined effectively.

The invention provides an image characteristic extraction method and a pedestrian detection method. The image characteristic extraction method comprises steps that, middle-level image characteristics of a target image are acquired, an initial value of a hidden meaning characteristic is then acquired, a reconstruction error constraint term, a sparsity constraint term and a discriminability constraint term are respectively determined according to the middle-level image characteristics and the initial value of the hidden meaning characteristic; and the hidden meaning characteristic is determined according to the reconstruction error constraint term, the sparsity constraint term and the discriminability constraint term. According to the image characteristic extraction method, higher-level meaning information and more powerful discriminability information are both considered as crucial factors for determining characteristic performance, a characteristic containing richer information and more powerful discriminability is acquired through respectively optimizing a proposed hidden meaning characteristic learning problem and the largest discriminability constraint, and thereby higher pedestrian detection accuracy is realized.

The present invention relates to an individually customized online learning system and, more particularly, to an individually customized online learning system which, when respective learners make online connections using terminals belonging to the learners, automatically discloses learning problems on the terminals of respective learners, which instructs the learners to solve the learning problems using the terminals, which receives the learner's answers through the terminals and evaluates the students' grades, which additionally provides information regarding time taken to solve each problem, information regarding answers to respective problems, and information regarding solving of respective problems, and which can automatically adjust the level of difficulty of the automatically disclosed problems using information regarding time taken to solve each problem and to submit the answer and information regarding erroneous answers to corresponding problems, thereby disclosing problems according to the learning level of learners. Accordingly, an individually customized online learning system according to the present invention comprises: an examination server comprising a problem DB, which comprises problem information and example information subordinate to the problem information, and a learner DB, which has already registered and record-managed learner information, such that, when learner information within the learner DB is authenticated, the problem information and example information within the problem DB are output while being subordinate to the authenticated learner information; and a learner terminal, which interworks with the examination server, which connects to the examination server, inputs learner information, and performs authentication from the examination server, and which, when the authentication is successful, receives the problem information and the example information from the examination server, wherein the learner terminal outputs the problem information and the example information on the screen and, when one of pieces of the example information that are subordinate to the problem information is selected by an input device, generates selection information; the learner terminal matches the selection information so as to be subordinate to the problem information, the example information, and the learner information and transmits the same to the examination server, thereby generating examination record information; and the learner terminal connects to the examination server and enables the examination record information to be watched.

The invention provides a multi-task learning based method for recognizing race and gender through a human face image, relates to the technical fields such as the digital image processing field, the mode recognizing field, the computer vision field and the physiology field, and aims to solve the problems of recognizing of the race and the gender from a static human face or a video human face under a plurality of occasions. The multi-task learning method is a learning method for improving the learning performance through related task learning and has the advantages of recognizing the learning difference between the tasks, sharing related features of the tasks, improving the learning performance through the relevance, and reducing the high-dimension small sample over-learning problem. The multi-task learning method can be applied to the human face based race and gender recognizing; different semantics are treated as different tasks, and on that basis, the semantics-based multi-task feature selection is proposed. With the adoption of the method for recognizing the race and the gender, the generalization capacity of a learning system and the recognizing effect can be obviously improved.

Outlier detection methods and apparatus have light computational resources requirement, especially on the storage requirement, and yet achieve a state-of-the-art predictive performance. The outlier detection problem is first reduced to that of a classification learning problem, and then selective sampling based on uncertainty of prediction is applied to further reduce the amount of data required for data analysis, resulting in enhanced predictive performance. The reduction to classification essentially consists in using the unlabeled normal data as positive examples, and randomly generated synthesized examples as negative examples. Application of selective sampling makes use of an underlying, arbitrary classification learning algorithm, the data labeled by the above procedure, and proceeds iteratively. Each iteration consisting of selection of a smaller sub-sample from the input data, training of the underlying classification algorithm with the selected data, and storing the classifier output by the classification algorithm. The selection is done by essentially choosing examples that are harder to classify with the classifiers obtained in the preceding iterations. The final output hypothesis is a voting function of the classifiers obtained in the iterations of the above procedure.

The invention relates to a micro-expression recognition method based on active migration learning. The method comprises the following steps: (1) extracting micro-expression and micro-expression features; (2) establishing and solving a micro-expression active migration learning problem; (3) recognizing a microexpression. Based on the internal relation between macro expressions and micro expressions, a bridge between macro expressions and micro expressions is constructedthrough an asymmetric linear translator. Less labeled samples are adopted during the initial stage of the active learning. By means of the above translator, the supervision information of the macro-expression domain can be utilized by micro expressions in thetransformation domain. As a result, a high-quality sample can be selected for the active learning in the micro-expression domain. After the sample is manually marked, the sample is added to an existing training set. Therefore, a more effective classifier is obtained through training.

The invention aims at providing a method and device for sending and receiving a message. The method for sending and receiving the message comprises firstly, receiving an internet protocol message to be sent from a source virtual machine, then removing a target media access control address in the internet protocol message in order to generate a virtual extensible LAN message, and finally sending the virtual extensible LAN message to a switch corresponding to a target virtual machine according to a preset address list recorded with the mapping relation between the internet protocol address of the target virtual machine and the internet protocol address of a host machine. The media access control address learning problem is solved, at the same time, the virtual extensible LAN message gettingrid of the redundant media access control address is employed, the message is forwarded completely based on the internet protocol address of the target virtual machine, the bandwidth utilization rateis effectively increased, and good scalability is achieved at the same time.

The invention provides a robustness dynamic motion method based on reinforced learning and an all-body controller. The method comprises the steps of designing a reinforced learning process surroundinga phase space planning frame (PSP); utilizing a directional walking simplified restriction model inherent in PSP for carrying out phase space planning, reinforced learning problem and learning strategy evaluation; meanwhile, adopting the all-body dynamic controller as an accelerated speed command for calculating in an operation space; utilizing differential positive motion to convert into joint accelerated speed; and optimizing a counter-acting force of a non-drive robot according to the accelerated speed, so that two parameters adopting position or time as outputs can be calculated at the same time, multiple walking modes can be produced, and the process speed is applicable to real-time control. The invention provides a robustness dynamic walking controller formed by a dynamic motion plan program, a robustness reinforced learning process and a novel all-body motion controller, so that higher calculation efficiency is realized, and excellent robustness is obtained.

The invention discloses an encryption method for an error learning problem in ring domain and a circuit. The method comprises the following steps: sampling a polynomial and a noise polynomial and performing the number-theory transformation; operating the result after the number-theory transformation, obtaining a public key and a ciphertext, and completing the encryption of the to-be-encrypted information. The invention also discloses a circuit to realize the method and the circuit comprises: an encryption controller, a to-be-encrypted information storage device, a Gaussian sampling module, a read-only storage device, a Gaussian data storage module, a number-theory conversion processor, an iterative modular multiplication module and a ciphertext storage module. The Gaussian sampling module samples and generates a polynomial and a noise polynomial; the number-theory conversion processor is used to perform the number-theory transformation to the polynomial, the noise polynomial and the constant polynomial and to generate the ciphertext after operations on the to-be-encrypted information, the noise polynomial and the public key. The method and the circuit of the invention greatly increase the operational efficiency of the circuit, reduce the loss of the circuit, and ease the realization cost of an encryption circuit for the error learning problem in ring domain.