100 results about "Simultaneous learning" patented technology

The invention discloses a gesture recognition method based on 3D-CNN and convolution LSTM. The method comprises the steps that the length of a video input into 3D-CNN is normalized through a time jitter policy; the normalized video is used as input to be fed to 3D-CNN to study the short-term temporal-spatial features of a gesture; based on the short-term temporal-spatial features extracted by 3D-CNN, the long-term temporal-spatial features of the gesture are studied through a two-layer convolutional LSTM network to eliminate the influence of complex backgrounds on gesture recognition; the dimension of the extracted long-term temporal-spatial features are reduced through a spatial pyramid pooling layer (SPP layer), and at the same time the extracted multi-scale features are fed into the full-connection layer of the network; and finally, after a latter multi-modal fusion method, forecast results without the network are averaged and fused to acquire a final forecast score. According to the invention, by learning the temporal-spatial features of the gesture simultaneously, the short-term temporal-spatial features and the long-term temporal-spatial features are combined through different networks; the network is trained through a batch normalization method; and the efficiency and accuracy of gesture recognition are improved.

The invention provides a face attribute recognition method of a deep neural network based on cascaded multi-task learning and relates to the computer vision technology. Firstly, a cascaded deep convolutional neural network is designed, then multi-task learning is used for each cascaded sub-network in the cascaded deep convolutional neural network, four tasks of face classification, border regression, face key point detection and face attribute analysis are learned simultaneously, then a dynamic loss weighting mechanism is used in the deep convolutional neural network based on the cascaded multi-task learning to calculate the loss weights of face attributes, finally a face attribute recognition result of a last cascaded sub-network is used as the final face attribute recognition result based on a trained network model. A cascading method is used to jointly train three different sub-networks, end-to-end training is achieved, the result of face attribute recognition is optimized, different from the use of fixed loss weights in a loss function, a difference between the face attributes present is considered in the invention.

The invention discloses a multimodal brain network feature fusion method based on multi-task learning, and the multimodal brain network feature fusion method based on the multi-task learning includes the steps of preprocessing the obtained functional magnetic resonance imaging (fMRI) images and diffusion tensor imaging (DTI) images, registrating the preprocessed fMRI image to the standard AAL template, carrying out a fiber tracking for preprocessed DTI images, calculating fiber anisotropy (FA) value, and constructing structure connection matrix through the AAL template. Clustering coefficient of each brain area in a function connection matrix and the structure connection matrix is calculated to be regarded as function features and structure features. As two different tasks, the function features and the structure features assess an optimal feature set by solving the problem of multi-task learning optimization. The method uses information with multiple modalities complementing each other to learn simultaneously and to classify, improves the classification accuracy, solves the problems that a single task feature does not consider the correlation between features, and the fact that only one modality feature is used for pattern classification can bring to insufficient amount of information.

A visual question and answer method based on a multi-modal decomposition model is provided. The method is characterized in that the image is trained in an ImageNet dataset and image features are extracted, the question is marked as texts into feature vectors; a collaborative attention model is introduced into the basic network architecture, image and question related features are learned, and fine-grained correlation among multi-modal features are characterized; and multi-modal features enter a multi-modal decomposition bilinear pool (MFB) or a multi-modal decomposition high order cell (MFH) module image to generate a fusion image question feature z, and the z is put into the classifier to predict the best matching answer. According to the method provided by the present invention, the collaborative attention model is used to predict the correlation between each spatial grid in the image and the question, and the best matching answer can be accurately predicted in a facilitated manner;and in combination with the image attention mechanism, the model can effectively understand which image region is important to the question, so that the performance of the model and the accuracy of the question and answer can be significantly improved.

The present invention discloses an on-line real-time multi-person synchronous interdynamic learning virtual classroom system and its method. Based on convenient property of internetwork network, real-time feedback mechanism and enhanced interdynamic mechanism said invention provides a hommization platform so as to make user freely study the information desired to be studied by user. Said system includes: curricula-variable hall system module, virtual white board system module, curriculum cartoon library module, evaluation item pool module and study result analysis module. Said invention also provides the method for utilizing said system to obtain good study effect.

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

The invention discloses a skin disease detecting method based on deep learning and a system thereof. The method comprises the steps of preprocessing a picture of a to-be-detected skin, inputting the picture into an improved Faster RCNN neural network model, wherein the neural network model outputs the disease area and the disease type of the to-be-detected skin; wherein the neural network model utilizes a deep width residual error network for performing characteristic extraction on the input picture, and using the extracted characteristic picture as the shared characteristic of the disease area and the disease type for realizing simultaneous learning of two optimization objects. The skin disease detecting method can more effectively improve model performance and has improved detecting effect than basic Faster RCNN. The method and the system are suitable for face skin disease detection and disease detection at other parts or non-healthy area detection. The method and the system can be used for detection of various diseases in a medical industry and a medical beautifying industry.

Provided are sentence embedding method and apparatus based on subword embedding and skip-thoughts. To integrate skip-thought sentence embedding learning methodology with a subword embedding technique, a skip-thought sentence embedding learning method based on subword embedding and methodology for simultaneously learning subword embedding learning and skip-thought sentence embedding learning, that is, multitask learning methodology, are provided as methodology for applying intra-sentence contextual information to subword embedding in the case of subword embedding learning. This makes it possible to apply a sentence embedding approach to agglutinative languages such as Korean in a bag-of-words form. Also, skip-thought sentence embedding learning methodology is integrated with a subword embedding technique such that intra-sentence contextual information can be used in the case of subword embedding learning. A proposed model minimizes additional training parameters based on sentence embedding such that most training results may be accumulated in a subword embedding parameter.

The invention discloses a cross-media sparse Hash indexing method. The method comprises the steps of (1) performing unified modeling on incidence relations among data of a plurality of modes through hyper-graphs; (2) learning dictionaries of the plurality of modes simultaneously through a dictionary learning frame, applying regular constraint to sparse and hyper-graph incidence relations and learning data of each mode to obtain corresponding dictionaries; (3) using the learned dictionaries as Hash functions, and performing Hash encoding on new data through corresponding mode dictionaries; (4) converting the Hash codes into sparse code sets through corresponding Hash strategies to change sparse code similarity calculation problems into the set similarity calculation problems, and performing similarity calculation through a similar jaccard distance measurement mode.

The invention discloses a knowledge graph completion method and device, a storage medium and electronic equipment, and belongs to the technical field of computers. The knowledge graph completion method comprises the steps of obtaining a to-be-verified target knowledge text, generating a plurality of triples according to the target knowledge text and a preset knowledge graph, calculating each triple to obtain a corresponding confidence coefficient, verifying a target triple based on the corresponding confidence coefficients, and complementing the knowledge graph according to the verification result. Therefore, according to the method, the mixed model combining the text coding technology and the graph embedding technology is provided to learn the context and the structured knowledge at the same time, the reliable triple confidence score is obtained, advantage complementation of the two methods is achieved, the calculation overhead is remarkably reduced, and the complementation accuracy is improved. The invention further provides an adaptive integration scheme, scores of the coding method and the graph embedding method are fused in an adaptive mode, and the knowledge graph completion accuracy is further improved.

The invention provides a fast image retrieval method based on asymmetric depth discrete hash, a method for constructing a retrieval model and a retrieval model, wherein that scheme of the invention considers the compactness within a class and the separability between the hash codes of a query image and a database image, constructs a measure learning model, and simultaneously learn an approximate binary code of the query image and a discrete hash code of the database image through depth learning and discrete optimization; By constructing an asymmetric hash code learning framework, combining measure learning, depth learning and discrete optimization, a depth convolutional neural network is trained to learn discrete discriminant hash codes for database images and query images. So that the minimum Hamming distance between classes of hash codes is greater than the maximum Hamming distance within the class.

Disclosed is an intelligent presentation method. The intelligent presentation method of the present disclosure may support a presentation to be smoothly performed by learning content of the presentation while a presenter is presenting and performing a function required for the presentation in response to a command voice. The intelligent presentation-assisting device of the present disclosure may be associated with an artificial intelligence module, a robot, an augmented reality (AR) device, a virtual reality (VR) device, a device related to a 5G service, and the like.

It is to learn an object identification parameter while suppressing an influence of the background area. The object identification parameter learning system includes: a feature extracting device for obtaining a feature of an object from the image; a background specifying device for specifying a background area of the image; a background replacing device which replaces feature components corresponding to the background area of the feature with other values; and an identification parameter update device for updating the identification parameter based on the feature components replaced by the background replacing device. The identification parameter can be learnt by generating a plurality of pieces of feature data of the object with different backgrounds from a single object image through replacing the background area of the feature of the object.

The invention discloses an intelligent teaching-use piano and a piano teaching method. The intelligent teaching-use piano comprises a central processing unit, a record unit, a power circuit, a keyboard input unit, a conversion unit, a judging unit and a real-time prompt unit, the record unit, the power circuit, the keyboard input unit and the conversion unit are connected with the central processing unit respectively, the judging unit is connected with the central processing unit and the conversion unit, the real-time prompt unit is connected with the judging unit, the real-time prompt unit is capable of prompting a learner timely when the learner plays wrongly, so that a better learning effect is achieved, the learner can learn step by step according to the teaching information (images, videos, voices and the like) stored in the record unit, the learner can control the study rhythm independently, after one lesson is mastered, the learner can continue with the next lesson, meanwhile, the learner can look over study records through the record unit, the accuracy can be looked over timely, and the learning enthusiasm and learning efficiency can be improved.

The invention discloses a DDNN training method and a DDNN-based multi-view target recognition method and system, and belongs to the field of cloud computing. The method comprises the steps of obtaining an information entropy of a distributed deep neural network cloud side model to a sample image; constructing a DDNN target function based on the information entropy of the sample image; and according to the DDNN target function, jointly training an edge side model and a cloud side model of the DDNN. In the knowledge migration method focused on a teacher-student network, under the background of multiple outlets of a DDNN level, the adaptive training method based on sample weighting is provided, scores of samples are obtained from a deep outlet of a DDNN, the samples are weighted through the scores to distinguish simple and complex samples, the weighted samples are used for training a cloud side model and an edge side model at the same time, and the communication traffic is minimum while good classification precision is guaranteed. The cloud side model guides the whole training process of the edge model, and the edge model can learn real labels and cloud side migration knowledge at thesame time.

The invention discloses an unsupervised point cloud feature learning method and device based on local global bidirectional reasoning, and the method comprises the following steps: obtaining three-dimensional point cloud data, carrying out the feature extraction of the three-dimensional point cloud data through a preset neural extraction network, and obtaining a local point cloud feature and a global point cloud feature corresponding to the three-dimensional point cloud data; processing the global point cloud features through a decoder network to obtain an original point cloud and an estimatedpoint cloud normal vector; and performing unsupervised point cloud feature learning according to the original point cloud, the estimated point cloud normal vector and the local point cloud features toobtain network parameters, and generating a target classifier according to the network parameters. Therefore, structural information and semantic knowledge can be learned at the same time, the quality of unsupervised learning features is improved, and therefore the performance of unsupervised point cloud recognition is improved.

A method and apparatus are provided for learning a model for the appearance of an object while tracking the position of the object in three dimensions. Under embodiments of the present invention, this is achieved by combining a particle filtering technique for tracking the object's position with an expectation-maximization technique for learning the appearance of the object. Two stereo cameras are used to generate data for the learning and tracking.

The invention discloses a method for constructing a universal embedding framework of a multi-semantic heterogeneous graph, which comprises the following steps of: 1, constructing a neighborhood exploration strategy alpha-exploration, and smoothly splicing two exploration strategies, namely, DFS and BFS, so as to adapt to different heterogeneous network structures; 2, based on alpha-exploration, constructing an HNSE model, wherein the HNSE model comprises an alpha-exploration neighborhood exploration layer, a multi-semantic learning layer and a node classification layer; and learning low-dimensional embedding of the nodes while heterogeneous information and semantic information of the nodes are reserved; 3, realizing a multi-layer HNSE model in a residual form, and connecting a full-connection output layer behind the multi-layer HNSE model; and 4, constructing three expansion strategies of the HNSE. According to the method, each vertex of the multi-semantic heterogeneous graph is embedded by aggregating adjacent / meta-path neighbor nodes of different types, and a node aggregation sampling strategy combining meta-path neighbors and direct neighbors is designed for the HNSE, so that a multi-head attention mechanism in the HNSE is guided, and capture of node multi-semantic information is improved by utilizing meta-paths.

The invention provides a video hash retrieval method of an attention mechanism. The method comprises the following steps: (1) video preprocessing: sampling video frames, and constructing video pairs;(2) video frame feature extraction: performing feature extraction on each frame by using a convolutional neural network; (3) video feature learning: learning the video by using a twin network, a long-term and short-term memory (LSTM) neural network and an attention mechanism; (4) dimension reduction and training: performing dimension reduction on the video features by using a full connection layerto obtain a hash code with a desired length, and learning network parameters by using a gradient descent algorithm; and (5) retrieval: obtaining the hash code of each video by using one path of network of the twin network, calculating the Hamming distance with other videos, and performing sorting to obtain the nearest video. Compared with the prior art, the method has the advantages that the spatial information and the time information of the video are learned at the same time, the calculation cost is greatly reduced through sampling and other technologies, and the accuracy of video retrievalis also improved.

Methods of statistical learning for Bayesian inference in the context of efficient optimization schemes for image restoration are presented. Second and third order priors that may be learned while maintaining graph representability are identified. A framework to learn and impose prior knowledge on the distribution of pairs and triplets of labels via graph cuts is presented. The disclosed methods optimally restore binary textures from very noisy images with runtimes in the order of seconds while imposing hundreds of statistically learned constraints per node.

The invention discloses a video feature learning method based on video and text description pair discriminant. A video-text pair is formed by the video and the text description matched with the video;a three-dimensional convolutional network is adopted to extract video features, a DistilBERT network is adopted to extract text description features, the video and the corresponding text descriptionhave similar semantic features through training, the text description automatically becomes a label of the corresponding video, and a deep learning network is trained and constructed and used for learning the video features. According to the method for performing video feature learning by using text description information as auxiliary information, efficient video feature representation can be learned while the labor cost is effectively reduced, and all data in a data set can be more effectively utilized to obtain video representation with higher discrimination capability by using a video feature learning method by using the video and the text description for discriminant.

The invention discloses a semi-supervised classification method capable of simultaneously learning an affinity matrix and a Laplacian regularized least square, which mainly comprises the following steps: firstly, a joint model capable of simultaneously learning the affinity matrix and the Laplacian regularized least square is established according to a training sample; secondly, the block coordinate descent method is used to optimize all kinds of variables in the model; and finally, the soft label of the sample is obtained by a Laplacian regularized least square classifier, and the dimension with the largest element in a label vector is selected as the category of the sample. The invention effectively fuses the sparse self-representation problem of samples and the Laplacian regularized least square classifier, and realizes the simultaneous optimization and mutual improvement of the sample affinity matrix and the Laplacian regularized least square classifier in the learning process. Theinvention has an explicit classifier function, so that the problem of an external sample can be effectively handled. Compared with other semi-supervised classification methods, the method has more accurate classification accuracy and good application prospects.

The invention discloses a rolling bearing fault diagnosis method and system based on relational knowledge distillation, and belongs to the technical field of fault diagnosis. After the original vibration signals of the bearing are collected, a time-frequency diagram is constructed for each processing sample to serve as a fault sample, the fault sample serves as input of a fault diagnosis system, and due to the fact that the time-frequency diagram contains complete time-frequency information of the vibration signals, the real-time response efficiency and accuracy of fault diagnosis are improved. A student model is adopted to simultaneously learn a multivariate relationship between the output soft label of Softmax of a teacher model and output of a plurality of samples in the last pooling layer, namely, a student network learns from two aspects of a teacher structure and output of a single sample in the teacher network; and the classification performance of the fault diagnosis system is effectively improved under the condition that the memory and the training time are not increased. According to the invention, bearing fault diagnosis is realized by using a relational knowledge distillation transfer learning method, and the calculation complexity is effectively reduced through the idea of replacing a large model with a small model.

The invention discloses a text multi-view emotion classification method and system, a terminal and a storage medium. The method includes: obtaining view word- emotional phrases and text expressions ina text; inputting the document representation into an encoder established by a variational auto-encoder VAE method to generate a discrete probability distribution of the emotional polarity of the view angle; inputting the visual angle word and the emotional polarity discrete probability distribution into a decoder established by a variational auto-encoder VAE method to generate the emotional wordprobability corresponding to the visual angle word; generating emotional polarities corresponding to the viewing angle by minimizing variational boundaries while simultaneously learning the encoder and the decoder. The invention can solve the problems that manual labeling is time-consuming and laborious, and it is difficult to obtain manual labeling in multi-view emotion classification method.

The invention discloses a meta-learning-based domain increment method. The method comprises the following steps: S1, constructing a pre-training model; s2, training an old model by using the pre-training model; s3, training a new model. According to the invention, randomly reserved 5% memory data and new data are mixed and finely tuned to train a new model, and a cross entropy loss function and aknowledge distillation loss function are combined to guide the learning of the new model, so that the new model learns the classification knowledge of the data in the new field while memorizing the classification knowledge in the old field, and the expenditure of data storage and training time is greatly reduced.

The invention provides a hybrid expert reinforcement learning method and system, which utilize a multi-head agent with shared network parameters as a plurality of experts and train the agent through adeep deterministic strategy gradient algorithm so as to learn a series of similar tasks at the same time. Each expert is introduced into an uncertainty estimate of performing an action in a state toenhance the anti-overfit Q-value assessment capability and the overall performance of the model. These enable agents to extract, migrate, and share learned knowledge (feature expressions) among different tasks, thereby improving learning efficiency of a single task and effectiveness of scheduling experts among a plurality of tasks. Different from the traditional data-driven design of a hybrid expert system, the hybrid expert system adopts a self-supervised gating network to determine an expert with the highest potential to process each interaction of an unknown task, and can completely calibrate the scheduling accuracy through uncertainty estimation fed back by the expert system under the condition of no man-made external supervision information.

The invention discloses a continuous learning method for federated learning. After determining a learning task, a federal learning server and each client independently construct an auxiliary data set for a subsequent training process by collecting a plurality of public data sets which are matched with a task type and have no privacy problem; a client side enables a local model to learn an auxiliary data set and an auxiliary label while learning a new task by means of knowledge distillation loss, so that forgetting of old knowledge is reduced; the server side enables the aggregation model to learn the auxiliary data set and the auxiliary label at the same time by means of knowledge distillation loss, so that forgetting of the model in the aggregation process is reduced. According to the invention, on the basis of privacy security and low communication cost, the continuous learning ability of the federal model is improved.

The invention discloses a training method and device of a knowledge pre-training model and electronic equipment, and relates to the technical field of voice, natural language processing and deep learning. According to the specific implementation scheme, a training text is acquired, the training text comprises a structured knowledge text and a corresponding article, and the structured knowledge text comprises a head node, a tail node and a relationship between the head node and the tail node; and the knowledge pre-training model is trained to be trained according to the training text. Accordingto the method, the knowledge pre-training model to be trained can learn common knowledge and rich semantic knowledge at the same time, joint training of the common knowledge and the semantic knowledge can be achieved, and a training entity does not need to be embedded into the knowledge pre-training model to be trained; the performance gain of the knowledge pre-training model is not limited by the embedding quality of the training entity, the knowledge pre-training model can obtain rich context information from articles in the training text, dynamic adjustment can be carried out, and the flexibility is high.

The invention discloses a knowledge graph joint representation learning method fusing graph convolution and a translation model. The knowledge graph joint representation learning method comprises the following steps of 1) constructing a direct adjacency matrix and an indirect adjacency matrix corresponding to a knowledge graph according to the knowledge graph; 2) designing a drawing convolutional network which comprises an input layer, two hidden layers and an output layer, optimizing the attention coefficients of the adjacent point nodes to a central node, and obtaining the vector representation of the nodes by learning the structure information of the direct adjacent nodes and the indirect adjacent nodes; 3) learning the semantic information of the relationship by adopting a translation model to obtain the vector representation of an entity and the relationship; and 4) fusing the graph convolutional network and the translation model, and obtaining the final vector representation of the knowledge graph through the continuous iterative learning. According to the present invention, the structure information and the relation semantics of the knowledge graph can be learned at the same time, and the vector representation precision of the knowledge graph is improved.