89 results about "Sequence learning" patented technology

The invention discloses a reinforcement dual-channel sequence learning-based dialog reply generation method and system. The method comprises the following steps of: (1) modeling a context to obtain acontext semantic vector; (2) carrying out combined learning on a current dialog and the context semantic vector by utilizing an encoder so as to current a current dialog vector and an encoder vector;(3) inputting the context semantic vector and the current dialog vector into a decoder so as to obtain a first channel dialog reply draft and a decoder vector; (4) inputting the encoder vector, the decoder vector and the first cannel dialog reply draft into an embellishing device to carry out embellishing, so as to generate an embellished dialog reply of a second channel; (5) optimizing a target function by utilizing a reinforcement learning algorithm; and (6) ending model training and generating and outputting a dialog reply. By utilizing the method and system, dialog generation models can grasp global information more deeply, and replies more according with dialog scenes and having substantial contents can be generated.

The invention discloses a lip recognition method and system based on a neural network. The method comprises: a lip sequence image is acquired, and further from the acquired lip serial images, the feature of lip image is extracted, and then inputting the extracted features of the lip sequence images to the bi-directional long-short time memory network for time-space feature sequence learning, and training the features of the learned lip serial images, training features of the learned lip sequence image to a lip recognition model, and decoding and recognizing features of the extracted lip sequence image according to the trained lip sequence image features to the lip recognition model to recognize a lip result. By using the above method, it is possible to recognize the video without being influenced by the environmental noise, and recognize the lip language result, which has a high accuracy rate and a good user experience.

Disclosed are a method and a system for automatically classifying data expressed as a plurality of factors with values of a text word and a symbol sequence by using deep learning. The method comprises the steps of: inputting the data expressed by the plurality of factors so as to express a word vector including sequence information of the factors through sequence learning of words corresponding to the factors with respect to each factor constituting the data in a first model; inputting an output of the first model so as to calculate points of each category for classifying the categories of the data by using the word vector including the sequence information of the factor in a second model; and determining at least one category for the data by using the points of each category.

A computer-implemented method and system for predicting operation risks of a vehicle. The method and system obtains a training data stream of vehicular dynamic parameters and logging crash time instances; partitions the data stream into units representing dimension vectors, labels the units that overlap the crash time instances as most dangerous; labels the units, which are furthest from the units that are labeled as most dangerous, as most safe; propagates the most dangerous and the most safe labeling information of the labeled units to units which are not labeled; estimates parameters of a danger-level function using the labeled and unlabeled units; and applies the danger-level function to an actual data stream of vehicular dynamic parameters to predict the operation risks of the vehicle.

The invention discloses an image analysis method based on a recurrent neural network, and the method comprises the steps: building a plurality of first two-dimensional axial slice images based on an original three-dimensional image; Carrying out convolution operation on the plurality of two-dimensional axial slice images to obtain a high-resolution characteristic image, stacking the characteristicimage into a three-dimensional characteristic image, and then cutting the three-dimensional characteristic image into an axial view, a sagittal view and a coronal view; Processing the axial view through a sequential learning network to generate an axial sequential learning feature map; Processing the sagittal view by expanding the residual network to generate a sagittal learning feature map; Processing the coronal view by expanding the residual network to generate a coronal learning feature map; Creating a first three-dimensional body based on the sagittal learning feature map, and cutting the first three-dimensional body into a plurality of second two-dimensional axial slices; Creating a second three-dimensional body based on the coronal learning feature map, and cutting the second three-dimensional body into a plurality of third two-dimensional axial slices; Cascading the axial sequence learning feature map, the second plurality of two-dimensional axial slices and the third plurality of two-dimensional axial slices to form cascade feature mapping; Applying a convolution operation to the cascade feature mapping to obtain a fused multi-view feature; And combining the fused multi-view feature with the high-resolution feature map to carry out image segmentation.

The invention discloses a dialogue reply generation method and system based on self-commenting sequence learning. The dialogue reply generation method comprises the following steps that: (1) carryingout modeling on the context language environment of a current dialogue to obtain a context semantic vector; (2) according to the context semantic vector, establishing a dialogue language based on theself-commenting sequence learning; (3) training and testing the dialogue model to independently obtain a training award value and a testing award value; (4) calculating a difference between two awardvalues, and optimizing the dialogue model through policy gradient calculation; and (5) finishing dialogue model training, and outputting the dialogue reply. By use of the method, in a training processof a dialogue generation model, the reply with a better substantive meaning can be generated towards an evaluation index optimization direction, and dialogue generation instability can be greatly lowered.

A system for generating a sequence of learning objects comprising an e-Learning course is provided. The system can include one or more processors and a memory unit for electronically storing data comprising a plurality of learning objects. The system further can include a sequence builder configured to execute on the at least one processor. The sequence builder, more particularly, can be configured to generate a sequence comprising at least a portion of the plurality of learning objects based upon a measurable user relevancy and / or a measurable query relevancy of each learning object contained in the sequence.

A video abstraction method based on attention expansion coding and decoding network comprises: regarding a video abstract as a sequence-to-sequence learning process; using time domain correlation between video frames is utilized to obtain a video frame feature sequence from an original video in the SumMe or the TVSum through a pre-training network; taking the video frame feature sequence as inputof an encoder network in an attention expansion coding and decoding network to obtain a semantic information sequence of the video frames, and then obtaining a score corresponding to each video framethrough a multiplicative attention decoding network; the scores of all the video frames forming an abstract sequence; obtaining a semantic information sequence of an abstract sequence through a retrospective encoder, constructing global semantic discrimination loss, introducing a moving average model, learning semantic correlation between the abstract sequence and a video frame feature sequence, obtaining a new abstract sequence retaining important information of an original video, and finally selecting a set final abstract through the new abstract sequence. The robustness of the model is enhanced.

A computer-implemented method and system for predicting operation risks of a vehicle. The method and system obtains a training data stream of vehicular dynamic parameters and logging crash time instances; partitions the data stream into units representing dimension vectors, labels the units that overlap the crash time instances as most dangerous; labels the units, which are furthest from the units that are labeled as most dangerous, as most safe; propagates the most dangerous and the most safe labeling information of the labeled units to units which are not labeled; estimates parameters of a danger-level function using the labeled and unlabeled units; and applies the danger-level function to an actual data stream of vehicular dynamic parameters to predict the operation risks of the vehicle.

The invention discloses an air control voice command recognition method based on deep learning. The method comprises the following steps: acquiring a voice signal to be recognized, and converting thevoice signal into 16-bit 16-kHz PCM audio data; building a deep network model; training the deep network model by using training data to obtain a voice recognition engine; performing voice segmentation on the audio data; and inputting effective audio clips obtained by voice segmentation into the voice recognition engine, and outputting a character recognition result. According to the deep networkmodel, a convolution module is used as a feature extractor; extracted feature data is processed through a reshape layer and a full-connection layer; sequence learning is carried out through a gating circulation unit; finally classification learning and decision making are carried out through the full-connection layer, so that a prediction result is obtained. According to the method, an artificialintelligence deep learning engine is adopted as a core, so that the method has the advantages of extremely high professional applicability and accent generalization ability, and lower data quantity dependence, and is obviously superior to a general voice recognition system in air control voice recognition.

The invention provides a robot 3D shape recognition method based on multi-view information fusion. The method combines the advantages of a full-view method and a single-view method, overcomes the disadvantages of the two methods, and comprises steps of: firstly, performing image similarity ordering by using an image similarity detection technique by means of the multi-view information of a 3D shape acquired by a robot in motion, so as to obtain a hierarchical depth feature by a convolutional neural network; finally, learning the visual feature with a certain time and spatial sequence by using a long short term memory model to obtain a highly abstract space-time feature. The method not only simulates the hierarchical learning mechanism of human beings, but also adds the space-time sequence learning mechanism for simulating the human learning, and realizes the high-precision classification and recognition of 3D shape by multi-view information fusion.

The invention provides a point interaction deep learning segmentation algorithm specially for solving the kidney tumor segmentation problem in a medical image. The algorithm is composed of a point interaction preprocessing module, a bidirectional ConvRNN unit and a core deep segmentation network. The algorithm starts from a tumor center position provided by an expert; in 16 directions with uniformintervals, 16 image blocks with the size of 32 * 32 are intensively collected from inside to outside according to the step length of 4 pixels to form an image block sequence, a deep segmentation network with sequence learning is used for learning the inside and outside change trend of a target, the edge of the target is determined, and segmentation of the kidney tumor is achieved. The method canovercome the influences of low contrast, variable target positions and fuzzy target edges of medical images, and is suitable for organ segmentation and tumor segmentation tasks. Compared with the prior art, the method has the following characteristics: 1) the interaction mode is simple and convenient; (2) a Sequence Patch Learning concept is provided, and a sequence image block is used for capturing a long-range semantic relationship, so that a relatively large receptive field can be obtained even in a relatively shallow network; and 3) a brand-new ConvRNN unit is provided, the inside and outside change trend of the target is learned, the interpretability is relatively high, the actual working mode of doctors is met, and the final model is high in precision and strong in applicability.

The invention relates to the field of computer vision and deep learning, in particular to a hippocampus segmentation method based on sequence learning. The method comprises the following steps: step 1, preprocessing an original image set A; 2, building a network model, wherein the hippocampus segmentation network model comprises an encoding part, a bidirectional convolution long and short memory network and a decoding part; Step 3, a model is trained; And performing forward propagation on the anatomical plane image set D, E and F to obtain a single iteration result, and calculating a loss function to obtain weight models J, K and L through backward propagation. According to the method, the hippocampus structure in the human brain nuclear magnetic resonance image is efficiently, automatically and accurately segmented by utilizing the method based on the deep learning network, and the operation speed is relatively high while the high segmentation precision is ensured. Besides detection of hippocampus, the network provided by the invention can be retrained, so that the network can be applied to detection and segmentation of other organs or tissues.

The invention discloses a multi-modal fusion sign language recognition system and method based on graph convolution. The system comprises a feature extraction module, a feature fusion module, a sequence learning module and an alignment translation module. The method comprises the following steps: 1, respectively extracting color, depth and skeleton features of video frames from a sign language video database by using a convolutional neural network and a graph neural network; 2, combining the multi-modal features, and fusing the network fusion features through a multi-modal sequence; 3, constructing a bidirectional recurrent neural network to perform sequence learning on a series of fused fragment-level features; and 4, aligning the feature sequence through the connected subjective timing sequence classification model, and translating a complete sign language sentence. According to the invention, translation of continuous sign language sentences can be realized, and the accuracy of continuous sign language translation is improved.

The invention relates to the field of natural languages, in particular to a natural language generation method based on SQL syntax tree node types. The method does not need a large amount of manual operation and does not require natural language to support various sentence patterns. Comparing with a natural language generation method based on sequence-to-sequence learning. According to the invention, the text information of the SQL language can be obtained; tree-shaped structured data of an SQL syntax tree and a tree-shaped long short-term memory network are combined for use; the grammatical structure information of the SQL statement is obtained more fully, the method has practical application significance, the defect that a development document and online data are consulted manually is overcome, the time cost and the labor cost are greatly reduced, and the working efficiency is improved.

The invention provides a portable electroencephalogram depression detection system in combination with a demographic attention mechanism. On one hand, the accuracy of electroencephalogram signal sequence learning and modeling is improved by using a convolutional neural network, and on the other hand, demographic information of individuals is introduced in combination with the attention mechanism,and more effective depressive disorder detection is realized. The system comprises an electroencephalogram data acquisition module, a data preprocessing module and a depressive disorder detection module, wherein the electroencephalogram data acquisition module is used for acquiring resting state electroencephalogram original data of a subject; the data preprocessing module is used for carrying outdata preprocessing on the collected original data; and the depressive disorder detection module is used for completing depressive disorder detection based on the electroencephalogram data after datapreprocessing, constructing and training a model by adopting an artificial neural network to classify electroencephalogram signals, and fusing the demographic information into a modeling process of the electroencephalogram signals by jointly using convolution operation and the attention mechanism.

Embodiments of the invention disclose an artificial intelligence-based retrosynthesis prediction method and device, equipment and a storage medium. The method comprises the following steps of: acquiring a graph structure of a product molecule and attribute features of atoms in the product molecule; predicting a fracture chemical bond in the product molecule at least according to the graph structure of the product molecule and the attribute features of atoms in the product molecule through a graph neural network model; carrying out bond breaking treatment on the product molecule based on the fracture chemical bond to obtain at least one synthon; and predicting a reactant molecule at least according to a character string corresponding to a retrosynthesis reaction type, a character string corresponding to the product molecule and a character string corresponding to the at least one synthon through a sequence learning model. According to the method, the prediction precision of the retrosynthesis reaction of an organic compound can be effectively improved, so the prediction process of the retrosynthesis reaction of the organic compound is easier to visualize and is more interpretable.

A method includes enabling a robot to learn a mapping between English language commands and Linear Temporal Logic (LTL) expressions, wherein neural sequence-to-sequence learning models are employed to infer a LTL sequence corresponding to a given natural language command.

The invention relates to an industrial control intrusion detection method for hierarchical dependency relationship modeling. The communication traffic is monitored in a data packet window mode, features are extracted from three aspects of a traffic layer, a data packet layer and a content layer, and dependency relationships of the three dimensions are respectively constructed through a parallel LSTM neural network in order to convert the relationship features of each single aspect into a multilayer dependency relationship model; and the outputs of different LSTM neural networks are integratedthrough a depth auto-encoder to form a final abnormal value output result. The attack traffic with different dependency relationships can be detected by analyzing the network flow sequence of the given data packet windows, and the implicit relationship of the traffic is analyzed from multiple dimensions and side surfaces, so that the range of traditional traffic detection is expanded, and the accuracy of attack traffic detection is improved; and the long-sequence learning capability of the LSTM neural networks is utilized to the maximum extent, and the change of the data flow is explained fromthe perspective of time dimension.

The invention relates to a device and system for determining a sequence of blood vessel condition parameters of a blood vessel by utilizing a computer. The device includes a memory, a processor and acomputer executable instruction stored in the memory and capable of running on the processor. The device is used for acquiring an image block sequence of a blood vessel path; and based on the acquiredimage block sequence of the blood vessel path, predicting a sequence of blood vessel condition parameters of the blood vessel path by utilizing a trained deep learning model formed by series connection of a data stream neural network, a recurrent neural network and a conditional random field model. By utilizing advantages of the recurrent neural network in sequence learning and advantages of a convolutional neural network in image learning, and through end-to-end model training, the device overcomes several critical problems in rapid computing of a blood vessel condition parameter sequence, such as computing speed increasing, manual intervention reducing for feature extraction and accuracy ensuring.

The invention relates to a character recognition method and device, computer equipment and a storage medium. The character recognition method includes the steps: firstly, a feature extraction neural network performs feature extraction to obtain a shared feature map, and then a character center sequence table is obtained according to the shared feature map and a character center detection neural network, and features are extracted based on the character center sequence table to obtain a feature sequence, and finally character recognition is performed based on the obtained feature sequence. According to the character recognition method, feature sampling is carried out based on a text center point, so that features of irrelevant positions can be effectively filtered; and meanwhile, the character recognition method performs feature collection based on the character central point sequence table and maintains the overall vision of features. The method successfully combines the idea of two-dimensional spatial distribution with the idea of sequence learning, and is excellent in recognition of irregular characters of the image.

The invention discloses a short-term wind power prediction method based on double-time sequence feature learning, and the method comprises the following steps: building a training set and a test set,and converting original data into labeled data at the same time; Adopting a singular spectrum analysis method to perform de-noising and principal component selection on the original wind power data; Constructing a double-time-sequence feature learning neural network model composed of a local time sequence learning module and two long-short-term memory networks, and obtaining local wind power dataat different moments according to the input of the neural network model; And the neural network model outputs the double-time sequence characteristics processed by one local time sequence learning module and two long and short term memory networks through a full connection layer, and performs final regression analysis to obtain a to-be-predicted wind power value at the t + 1 moment at the t moment. According to the method, through principal component selection and multi-scale time sequence characteristic learning of original data, accurate prediction of the power generation power of the singlefan of the wind power plant is finally realized.

The invention, which belongs to the field of intelligent traffic systems, provides a traffic state estimation method based on kmeans clustering and deep sequence learning, thereby solving the problemthat the traffic state of a whole expressway cannot be estimated under the condition that traffic flow data of part of road sections in the urban expressway cannot be acquired in real time. The methodis characterized by comprising the following steps: (1), dividing an expressway network; (2), carrying out modeling and data acquisition of an expressway; (3), preprocessing and normalizing the data;(4), calculating the Euclidean distance between the traffic flow data through a kmeans clustering algorithm, and determining the traffic state grade of each data point; and (5), designing a deep sequence learning Seq2Seq model, and carrying out traffic state identification on the whole road network through model iterative learning. The method gives full consideration to the relation of traffic flows between road segments and gives play to the advantages of a machine learning algorithm in the traffic field; the traffic state of the whole road network can be obtained in time; and reliable traffic information can be provided for a driving main body.

The invention provides an electrocardiosignal classifying method and system based on LRF-ELM and BLSTM. The electrocardiosignal classifying method includes the steps: acquiring electrocardiosignal data, preprocessing the electrocardiosignal data to obtain a dataset, and using the electrocardiosignal data in the dataset as input data of a neural network; using a LRF-ELM network as a feature extractor, learning spatial information in the electrocardiosignal data, and through three stacked random convolution and pooling processes, extracting feature data of different dimensions in the electrocardiosignal data; and after fusion, using the extracted feature data as input of a sequence learning stage, adopting a deep BLSTM network to carry out sequence learning, and finally outputting electrocardiosignal classifying results. According to the electrocardiosignal classifying method and system based on the LRF-ELM and the BLSTM, time information and spatial information of electrocardiosignals are taken into account at the same time, and therefore, not only can electrocardiosignal features be extracted efficiently and rapidly, but also good classification and identification properties are ensured.

The invention discloses a traffic index prediction method and device based on a sequence-to-sequence learning model. The method comprises the steps of obtaining a traffic index input time sequence; taking the traffic index input time sequence as an input object from the sequence to a sequence learning model, wherein the sequence-to-sequence learning model is composed of an encoder network and a decoder network, the encoder network is an LSTM network taking an LSTM unit as a basic cycle unit, and the decoder network is an LSTM network taking an LSTM unit as a basic cycle unit; performing feature extraction on the traffic index input time sequence by using an encoder network to obtain a time change feature vector; and processing the time change feature vector by using a decoder network, andtaking a traffic index output time sequence obtained by processing as a prediction result. According to the method, the prediction precision of the traffic index is improved, and the nonlinear changeof the traffic index can be better described.

The invention discloses a music-driven dance generation method comprising the following aspects of 1) collecting a dance data set; 2) performing advanced feature extraction on music and dance sequences; and 3) providing a brand-new cross-domain sequence pair sequence learning framework for real dance generation. According to the method in the invention, the efficiency is high, the accuracy is alsoimproved, the model is more complex, the model size is small, the calculation speed is high, the prediction process is simpler, and the music can be quickly combined with the dance sequences.

Managing passwords is provided. A machine training process is performed using a set of existing passwords to train a machine learning component. Members of a set of semantic categories are used to categorize respective passwords in the set of existing passwords. Password strengths corresponding to a set of candidate passwords are evaluated using the machine learning component. A resource is secured with a candidate password having a password strength greater than or equal to a defined password strength threshold level.

Embodiments relate to a first processing node that processes an input data having a temporal sequence of spatial patterns by retaining a higher-level context of the temporal sequence. The first processing node performs temporal processing based at least on feedback inputs received from a second processing node. The first processing node determines whether learned temporal sequences are included in the input data based on sequence inputs transmitted within the same level of a hierarchy of processing nodes and the feedback inputs received from an upper level of the hierarchy of processing nodes.

The invention discloses an optical character recognition method of a spiking neuron supervised learning model based on membrane voltage driving. The method comprises the steps of obtaining an opticalcharacter sample set, a phase encoding method is adopted, a pulse sequence of the phase encoding method is adopted, a spiking neuron supervised learning model based on membrane voltage driving is established and trained, and the trained spiking neuron supervised learning model is used for recognizing optical characters to be detected. According to the invention, a phase encoding method is adoptedto encode a pulse sequence; establishing a spiking neuron supervised learning model based on membrane voltage driving and carrying out training; and the trained spiking neuron supervised learning model is used to identify the optical character to be detected, so that the spiking neuron sequence learning efficiency is significantly improved, and the optical character recognition efficiency and accuracy are further improved.

The invention provides a depth model for arrhythmia classification and a method and device using the model, and the depth model comprises a representation learning part and a sequence learning part. The representation learning part is used for receiving an equal-length sequence analyzed by the original electrocardiosignal; the representation learning part is constructed based on an MSCNN structure and is composed of two convolution block branches stacked in different scales; the convolution kernel of the first branch is large in scale and used for capturing low-frequency information of electrocardiosignals and outputting the low-frequency information in a multi-scale feature mode; the convolution kernel of the second branch is small in scale and used for capturing high-frequency information of the electrocardiosignal and outputting the high-frequency information in a multi-scale feature mode; the multi-scale feature output by the first branch and the multi-scale feature output by the second branch are spliced to form a multi-scale depth feature which is input to a sequence learning part; the sequence learning part is constructed on the basis of a Seq-Seq network taking LSTM as a basic unit, and an attention mechanism layer is arranged between an encoder and a decoder of the Seq-Seq network; the output is a time sequence depth feature.