113 results about "Cyclic network" patented technology

An improved method and system for selecting, retrieving and managing access to computer network locations. The method improves access to frequently utilized network locations by providing a recirculating list of network addresses selectable by a single user input. A user interface to the computer network can be accomplished utilizing a network browser in cooperation with a personal computer. The recirculating list of frequently utilized network locations is entered by the user into the system. The recirculating list has a selectable network location which is displayed to the user. The selectable network location can be requested and retrieved by a single user input such as selecting an icon which resides on a toolbar. The information or data associated with the selectable network location is automatically displayed in response to the user selecting the active URL of the recirculating list. The selectable network address can be accessed by a user in response to a single user input. After the single user input selects the active URL of the recirculating list the next URL becomes the active URL such that the user can efficiently select the next available selection. The list is recirculated such that a list of frequently accessed network locations can be efficiently managed and selected utilizing a single user input in cooperation with the recirculating list.

The invention discloses a video super-resolution reconstruction method based on depth learning. The technical key points are: (1) continuous images are given under the same shot, the network predictsclearer video frame images; (2) using a bi-directional circulating neural network and a depth 3D back projection network; (3) the invention combines two networks into one network, and the network is used as a network (4) for in-depth learning video super-resolution reconstruction of the invention. The training data is labeled, and the processed data video frames are passed through the network to obtain a loss function. A final object of the present invention is to input temporal and spatial information of a low-resolution video frame predicted by a bi-directional circulating network, after the3D projection network re-predicts the details of the video frame, an optimal model is obtained after repeated training. The model is applied to remove the effects of the camera shake, the blur of theobject's fast motion, out-of-focus blur, lens optical blur, depth-of-field variation, compression distortion and noise, and other degradation factors.

The invention discloses a path-based knowledge graph embedding method, which comprises the following steps of for a given single knowledge graph or a joint knowledge graph consisting of two knowledgegraphs, firstly sampling the given single knowledge graph by using a partial random walk method; next, modeling the obtained sampling path by using a new jumping cyclic network, so that the network can overcome the defect that the traditional cyclic neural network cannot identify the knowledge graph structure while effectively learning the dependency among elements in the path. For the output of the jump recurrent neural network, a type-based noise comparison estimation method is adopted to efficiently evaluate the loss of the network, and iteration is carried out in a feedback updating mode to carry out training. Compared with the prior art, the knowledge graph embedding trained by the method has the advantages of high expressibility, strong generalization ability and the like, and the accuracy and the efficiency of knowledge graph completion and entity alignment can be remarkably improved by applying the method.

A method for the supervised teaching of a recurrent neutral network (RNN) is disclosed. A typical embodiment of the method utilizes a large (50 units or more), randomly initialized RNN with a globally stable dynamics. During the training period, the output units of this RNN are teacher-forced to follow the desired output signal. During this period, activations from all hidden units are recorded. At the end of the teaching period, these recorded data are used as input for a method which computes new weights of those connections that feed into the output units. The method is distinguished from existing training methods for RNNs through the following characteristics: (1) Only the weights of connections to output units are changed by learning—existing methods for teaching recurrent networks adjust all network weights. (2) The internal dynamics of large networks are used as a “reservoir” of dynamical components which are not changed, but only newly combined by the learning procedure—existing methods use small networks, whose internal dynamics are themselves completely re-shaped through learning.

The invention discloses a traditional Chinese medicine health consultation text named entity recognition method based on transfer learning, which comprises the following steps: S1, selecting a text named entity recognition annotation data set in other fields, constructing a neural network, and performing pre-training; S2, respectively constructing a one-way recurrent neural network, and performinglanguage model training by utilizing the traditional Chinese medicine health consultation text unlabeled data set to obtain a forward language model and a reverse language model. And S3, fusing the loop network layer output characteristics of the unidirectional language model, and combining the full connection network layer and the conditional random field layer to obtain a final named entity recognition model. And S4, naming the entity by using the traditional Chinese medicine health consultation text to identify the labeled data set, and carrying out fine tuning training. According to the method, named entities in other fields can be migrated to identify labeled text knowledge and unlabeled text knowledge in the field, so that the identification accuracy and recall rate of the named entities of the traditional Chinese medicine health consultation text are effectively improved, and the convergence speed of the model is increased.

A method for query expansion uses a representation of an input query object, such as an image, to retrieve representations of similar objects retrieved using the query object representation as a query. Given the set of image representations, a weight is predicted for each using a prediction model which assigns different weights to the image representations. An expanded query is generated as a weighted aggregation (e.g., sum) of the query object representation and at least a subset of the set of similar object representations in which each object representation is weighted with its predicted weight. A higher weight can thus be given to one of the similar object representations, in the expanded query, than to another.

VLSI layouts of generalized multi-stage and pyramid networks for broadcast, unicast and multicast connections are presented using only horizontal and vertical links with spacial locality exploitation. The VLSI layouts employ shuffle exchange links where outlet links of cross links from switches in a stage in one sub-integrated circuit block are connected to inlet links of switches in the succeeding stage in another sub-integrated circuit block so that said cross links are either vertical links or horizontal and vice versa. Furthermore the shuffle exchange links are employed between different sub-integrated circuit blocks so that spacially nearer sub-integrated circuit blocks are connected with shorter links compared to the shuffle exchange links between spacially farther sub-integrated circuit blocks. In one embodiment the sub-integrated circuit blocks are arranged in a hypercube arrangement in a two-dimensional plane. The VLSI layouts exploit the benefits of significantly lower cross points, lower signal latency, lower power and full connectivity with significantly fast compilation.The VLSI layouts with spacial locality exploitation presented are applicable to generalized multi-stage and pyramid networks, generalized folded multi-stage and pyramid networks, generalized butterfly fat tree and pyramid networks, generalized multi-link multi-stage and pyramid networks, generalized folded multi-link multi-stage and pyramid networks, generalized multi-link butterfly fat tree and pyramid networks, generalized hypercube networks, and generalized cube connected cycles networks for speedup of s≧1. The embodiments of VLSI layouts are useful in wide target applications such as FPGAs, CPLDs, pSoCs, ASIC placement and route tools, networking applications, parallel & distributed computing, and reconfigurable computing.

The invention discloses a processing method for applying a rapid spanning tree protocol (RSTP) to monocyclic network topology, and the processing method comprises following steps that two additional protocol fields are added into a BPDU (bridge protocol data unit), wherein a field A indicates the protocol type, a field B indicates topological variation information, if a port is invalid, the value of the field B is set to be true, and otherwise, the value of the field B is set to be false; after a switch receives the expanded BPDU, on one hand the field A is matched, when one or more switches do not support or are not configured into a monocyclic network working mode, all other switches on the network automatically run the standard rapid spanning tree protocol so as to cooperatively work; and on the other hand, the value of the field B is judged, if the value of the field B is true, an RSTP state machine transmits the BPDU with the true field value so as to notify a counter-side switch of link failure on the network, and each switch conducts the fault rapid restoration. Due to the adoption of the processing method, the cyclic network self-healing time in a monocyclic network topology when link failure exists can be shortened, simultaneously the standard RSTP can be automatically compatible, and the requirements of different kinds of communication services can be met.

The invention provides an image character recognition method and system based on deep learning and a medium. The method comprises the steps that the source category of an image is judged; extracting an image target area through a convolutional neural network and classifying the target area; carrying out orientation correction on the image of the target area, rotating the image to a forward orientation, and calculating an inclination angle of the image through line segment detection and a frequency domain signal analysis method; calculating a feature map of the image by using a target detectionalgorithm and a deep convolutional network, and carrying out target segmentation on the text line to carry out character recognition; according to a CRNN algorithm, combining the deep convolutional network with a bidirectional cyclic network, and carrying out end-to-end network training; and obtaining the position of the character in the picture and the model recognition content through training,and extracting character information. By adopting the computer vision and character recognition technology, the recognition problem of bill cards and table document data in the intelligent auditing process in the financial field is solved.

The invention discloses a continuous sign language video translation method based on the fusion of a time domain convolution network and a circulating neural network, which comprises the following steps: feature extraction of the sign language video and construction of a word list; processing of a time domain convolution network TCN; the processing of a bi-directional circulating neural network BGRU; word mapping process of feature; processing of a fusion network FL; optimization of network model parameters based on the fusion of the time domain convolution network TCN and the bi-directional circulating neural network BGRU; fusion and decoding of word coding vectors. The invention can effectively overcome the communication obstacle caused by inaccurate interpretation of sign language translation by utilizing the network structure of different perspectives, learn and analyze different manifestations of data by utilizing different networks, further improve the accuracy of sign language translation, and increase the robustness of sign language translation.

The invention relates to a circulating network man-machine conversation method based on an actor critic reinforcement learning algorithm. The system consists of two subsystems: dialogue generation system and emotion analysis system. The session generation system is based on the gate loop unit network model and uses the tagged session data set for training. Furthermore, we optimize the parameters of the trained model using the actor-critic algorithm in reinforcement learning. That is, we use the trained model to build two networks, called the 'actor' network and the 'critic' network; Further, in order to reduce training time and improve resource utilization, we have created multiple processes that assign a pair of 'actors' and 'critics' to each process.

The invention discloses a bilateral-circulation convolution network-based video super-resolution method. The method comprises the steps as follows: establishing bilateral-circulation networks, comprising a forward circulation sub-network and a backward circulation sub-network according to time sequence, wherein each circulation sub-network comprises an input sequence layer, two implication sequence layers and an output sequence layer from the bottom to the top, and each sequence layer comprises a plurality of states corresponding to video frames in different moments; connecting these states by using three convolution operations, comprising a feed-forward convolution, a cyclic convolution and a condition convolution so as to obtain a bilateral-circulation convolution network; transmitting a trained video in the established bilateral-circulation convolution network, and using a stochastic gradient descent algorithm to minimize mean square error between a predicated high resolution video and an actual high resolution video so as to iteratively optimize the weight of the network, and obtain the final bilateral-circulation convolution network; and inputting low-resolution video sequence to be processed in a final bilateral-circulation convolution network mode to obtain a corresponding super-resolution result.

VLSI layouts of generalized multi-stage and pyramid networks for broadcast, unicast and multicast connections are presented using only horizontal and vertical links with spacial locality exploitation. The VLSI layouts employ shuffle exchange links where outlet links of cross links from switches in a stage in one sub-integrated circuit block are connected to inlet links of switches in the succeeding stage in another sub-integrated circuit block so that said cross links are either vertical links or horizontal and vice versa. Furthermore the shuffle exchange links are employed between different sub-integrated circuit blocks so that spacially nearer sub-integrated circuit blocks are connected with shorter links compared to the shuffle exchange links between spacially farther sub-integrated circuit blocks. In one embodiment the sub-integrated circuit blocks are arranged in a hypercube arrangement in a two-dimensional plane. The VLSI layouts exploit the benefits of significantly lower cross points, lower signal latency, lower power and full connectivity with significantly fast compilation. The VLSI layouts with spacial locality exploitation presented are applicable to generalized multi-stage and pyramid networks, generalized folded multi-stage and pyramid networks, generalized butterfly fat tree and pyramid networks, generalized multi-link multi-stage and pyramid networks, generalized folded multi-link multi-stage and pyramid networks, generalized multi-link butterfly fat tree and pyramid networks, generalized hypercube networks, and generalized cube connected cycles networks for speedup of s≧1. The embodiments of VLSI layouts are useful in wide target applications such as FPGAs, CPLDs, pSoCs, ASIC placement and route tools, networking applications, parallel & distributed computing, and reconfigurable computing.

The invention discloses a HRRP target recognition method based on attentional depth bidirectional circulating neural network, firstly, the time-domain features of the data are extracted and the extracted time-domain features are segmented and non-uniformly quantized to obtain the coding of the local structure, and then the co-occurrence matrix is obtained by using the relationship between the local structure and several surrounding local structures, and then the structural embedding characteristics of the data are obtained by the co-occurrence matrix, and then the extracted embedded features are sent to a deep neural network composed of a full connection layer and a bi-directional loop neural network based on attention LSTM for training, at the same time, according to the output of the hidden layer of the loop network, the softmax layer is used to obtain the weight parameters of the attention model. Finally, the HRRP is identified by the softmax layer and the weight of the attention model, and the recognition results are obtained.

The invention relates to an action classification method based on a connection node spatio-temporal simple circulating network and an attention mechanism. Firstly, the space dimension of SRU model isextended, and an iterative computation method of ST-SRU model. Then, in the proposed ST-SRU model and global contextual attention mechanism, GCA-ST-SRU method. Finally, the proposed method is appliedto human behavior recognition. Firstly, the features of human joints are extracted by depth network, and then the human joints are extracted by GCA-ST-SRU method recognizes the extracted features. Themethod of the invention can reduce training time consumption and improve classification accuracy, and has obvious efficiency advantages. The method of the invention can quickly deduce the speed, which is favorable for the design of the real-time motion recognition system, and is suitable for running on the platform with limited computing power, and has wide application prospect in the fields of computer vision, intelligent monitoring, human-computer interaction and the like.

Disclosed is a method based on a cyclic neural network of an attention mechanism for predicting a stock index price. The method comprises the following specific steps of 1) collecting and cleaning stock indexes and historical price data of a single stock, wherein the stock indexes include the historical price data; 2) using the attention mechanism to conduct weighting adjustment on model input values; 3) inputting the weighted model input values into the cyclic neural network for model training and prediction. By means of the method, the deep features of various influence factors of the stockindex price can be ingeniously extracted, and compared with a simple one-factor cyclic network or a traditional multi-factor cyclic network, the method can improve the accuracy of predicting the stockindex price.

Described herein are systems and methods for determining how to automatically answer questions like ''Where did Harry Potter go to school '' Carefully built knowledge graphs provide rich sources of facts. However, it still remains a challenge to answer factual questions in natural language due to the tremendous variety of ways a question can be raised. Presented herein are embodiments of systems and methods for human inspired simple question answering (HISQA), a deep-neural-network-based methodology for automatic question answering using a knowledge graph. Inspired by human's natural actions in this task, embodiments first find the correct entity via entity linking, and then seek a proper relation to answer the question--both achieved by deep gated recurrent networks and neural embedding mechanism.

The invention discloses an image super-resolution reconstruction method. The method comprises the following steps: making a learning sample set; Learning and training: a single-frame model of the convolutional neural network is adopted as a recurrent network module, a super-resolution mapping model of a recurrent neural network mode comprising N + 1 recurrent network modules is constructed, and inthe training process, a single lowest-resolution image is input into each recurrent network module in the super-resolution mapping model for single-time training; Replacing a low-resolution picture and repeatedly carrying out the single training; Carrying out single training for multiple times; And reconstruction: inputting the low-resolution image to be reconstructed into the trained super-resolution mapping model to obtain a high-resolution image. According to the method, the model is guided to learn low-frequency and high-frequency information of the image at the same time by adding the output error items related to the blurred image in the performance indexes, so that the image super-resolution reconstruction result can give consideration to reliability and details; And meanwhile, thestorage space can be greatly saved through the super-resolution mapping model in a cyclic network form.

The invention provides a skeleton action recognition method based on a multi-stream spatial attention graph convolution SRU network. The method comprises the following steps: firstly, embedding a graph convolution operator into a simple cycle unit to construct a graph convolution model to capture time-space domain information of skeleton data; meanwhile, in order to enhance the discrimination between the joint points, a spatial attention network and a multi-stream data fusion mode are designed, and a graph convolution simple cycle network model is further expanded into a multi-stream spatial attention graph convolution SRU. According to the method, the high classification precision is maintained, the complexity of the method is remarkably reduced, the reasoning speed of the model is increased, the balance between the calculation efficiency and the classification precision is achieved, and the method has a very wide application prospect.

A method, a network apparatus, and a tangible machine-medium thereof for detecting a looping network packet are provided. A receiving module receives a network packet. A memory stores a routing table comprising an address related variable, a sequence number related variable and a radius related variable. A determining module determines whether the network packet is a looping network packet by comparing the parameters of the routing table with those of the network packet. With these arrangements, the routing loop caused by transmission of the network packets can be detected using current network structures.

The invention discloses a mechanical equipment health state identification method based on an embedded cycle network. The mechanical equipment health state identification method based on the embeddedcycle network comprises the following steps: firstly, obtaining a vibration signal sample of mechanical equipment and dividing the vibration signal sample into a plurality of sections of adjacent local vibration sub-signals; then extracting a time-frequency image from the sub-signals, and taking the time-frequency image as input after standardized processing; then constructing a depth residual network and training the depth residual network so as to learn various pieces of local information in the signals; and finally, embedding the trained residual network into an input layer of the cycle network by utilizing thought of migration learning, fixing a convolution part so that the convolution part does not participate in training, and re-training the embedded network by utilizing the originalsample input again so as to learn information between adjacent periods; and diagnosing the health state of the mechanical equipment by the trained embedded cycle network. According to the method, thelocal vibration characteristic and the relation between adjacent parts are modeled, the influence of noise and local interference in the signals on a classification effect is weakened, and the applicability of the network in practical engineering is improved.

The invention discloses a software defect auxiliary dispatching method based on the activeness of a text and a developer. By considering text information of a defect report and the activeness of the developer, on one hand, a two-way circulating network and pooling method is utilized to extract text characteristics of the defect report, on the other hand, a one-way circulating network is utilized to extract developer activeness characteristics at a specific moment, the two-way circulating network and pooling method and the one-way circulating network are blended, a new defect dispatching model based on the activeness of the text of the defect report and the activeness of the developer is given, and therefore the problem that the auxiliary defect dispatching accuracy is low is well solved. An experiment result on four different open-source item data sets of Eclipse and the like shows that, compared with work of the same kind, the defect dispatching prediction accuracy is obviously improved.

The invention provides an electronic medical record part-of-speech tagging method, a model training method and related device in the technical field of natural language processing. The method comprises the following steps of after acquiring a output sequence of a circulation network by using semantic information of a preliminary learning training word segmentation sequence of the circulation network, using a semantic connection network to process the output sequence of the circulation network to learn deep semantic information of a learning training segmentation sequence in the way of learningthe training segmentation sequence multiple times so that a prediction part-of-speech labeling result corresponding to a training part-of-speech sequence. Model parameters of a part-of-speech labeling model are updated based on the prediction part-of-speech labeling result and a training part-of-speech labeling result corresponding to the training segmentation sequence. Compared with the prior art, the part-of-speech labeling model can fully learn deep semantic information and long-distance feature information of the sample sequence, and the accuracy of part-of-speech labeling can be improved.

The invention discloses an unsupervised remote sensing image super-resolution reconstruction method based on a recurrent neural network. The whole network is composed of two circulation networks, thefirst circulation network takes a low-resolution training image xt as input, a high-resolution image y is generated through a first generation network, and then a low-resolution image G2 (y) is generated through a second generation network. The second circulation network takes the high-resolution training image yt as input, a low-resolution image x is generated through a second generation network,and the low-resolution image x generates a high-resolution image G1 (x) through the first generation network. And finally, reconstructing a low-resolution remote sensing image through the trained first generation network to obtain a high-resolution image. According to the reconstruction method, the non-paired high-resolution and low-resolution remote sensing image can be used for network training, and the problem that a traditional method cannot use a non-matched image pair for training is solved.

According to the task type dialogue system based on deep network learning provided by the invention, two new processing flows are put forward: selection of a word-character-level cyclic network and aneural network. The hierarchical hybrid code network exceeds other models and achieves the most advanced performance regardless of the correct rate of each sentence of conversation or the correct rateof each complete conversation. The hierarchical hybrid code network can surpass the ability of the hybrid code network to mainly return to its coding sentence vector and identify the superiority of unregistered words. The word bag and average word vector of the mixed code network lack word order information, and this is exactly what the word-character-level loop network can do..

The invention discloses a transformer complex working condition identification method based on deep learning. The method comprises the following steps that 1, original sample data is acquired; 2, theoriginal sample data is built into a data set with type identification, a data set without type identification and a data set of test data; 3, according to a certain time window, window obtaining andgrouping are conducted on the built data sets; 4, voltage and current sequence signals in the window are processed to obtain frequency spectrum data; 5, the frequency spectrum data is subjected to recurrent neural network training; 6, a trained recurrent neural network is tested and optimized; 7, field data is input to the optimized recurrent neural network for conducting identification and precise positioning of transformer complex working conditions. According to the method, by adopting the recurrent network, complex mixed faults can be accurately judged and precisely positioned, and the robustness and a practical level of transformer complex working condition identification are improved.

VLSI layouts of generalized multi-stage and pyramid networks for broadcast, unicast and multicast connections are presented using only horizontal and vertical links with spacial locality exploitation. The VLSI layouts employ shuffle exchange links where outlet links of cross links from switches in a stage in one sub-integrated circuit block are connected to inlet links of switches in the succeeding stage in another sub-integrated circuit block so that said cross links are either vertical links or horizontal and vice versa. Furthermore the shuffle exchange links are employed between different sub-integrated circuit blocks so that spacially nearer sub-integrated circuit blocks are connected with shorter links compared to the shuffle exchange links between spacially farther sub-integrated circuit blocks. In one embodiment the sub-integrated circuit blocks are arranged in a hypercube arrangement in a two-dimensional plane. The VLSI layouts exploit the benefits of significantly lower cross points, lower signal latency, lower power and full connectivity with significantly fast compilation.The VLSI layouts with spacial locality exploitation presented are applicable to generalized multi-stage and pyramid networks, generalized folded multi-stage and pyramid networks, generalized butterfly fat tree and pyramid networks, generalized multi-link multi-stage and pyramid networks, generalized folded multi-link multi-stage and pyramid networks, generalized multi-link butterfly fat tree and pyramid networks, generalized hypercube networks, and generalized cube connected cycles networks for speedup of s≧1. The embodiments of VLSI layouts are useful in wide target applications such as FPGAs, CPLDs, pSoCs, ASIC placement and route tools, networking applications, parallel & distributed computing, and reconfigurable computing.

The invention discloses a facial feature point positioning method based on an end-to-end circulation network. The method includes the following steps: (1) a data preparation stage; (2) an end-to-end model designing stage; (3) a model training stage; and (4) a model testing stage to obtain a facial feature point positioning result of a face. Depth features of embedded shape information and a circulation neural network are cooperated, so the facial feature point positioning precision is greatly improved. In addition, through utilization of inherited neural network features, the model computational complexity will be greatly reduced, and the ultra-real-time facial feature point positioning speed is achieved while the good facial feature positioning precision is maintained.

The invention discloses a 3D medical image recognition and segmentation method based on Unet and LSTM. The method comprises the steps: a medical image preprocessing stage: reading a medical image in a3D format, enabling the medical image to be decomposed into a 2D image sequence in a z-axis direction, and carrying out the z-score normalization processing of image data in a 2D level; a segmentation network training stage: using the normalized 2D image sequence samples to divide a training set for training, independently taking out the output of the intermediate layer of the U-net as an intermediate variable sequence, and using the intermediate variable sequence to train an LSTM network; and a segmentation network identification and inference stage: inputting the sample into the network toobtain pixel-level segmentation output, and combining output sequences into a 3D matrix to obtain a final result. In the aspect of 3D medical image recognition, the mode that the 3D medical image is segmented into the 2D sequence and the intermediate variable is processed in combination with the cyclic network is adopted. The calculated amount is reduced, and the efficiency of recognition work isimproved.