49 results about "Hierarchical neural network" patented technology

In a hierarchical neural network having a module structure, learning necessary for detection of a new feature class is executed by a processing module which has not finished learning yet and includes a plurality of neurons which should learn an unlearned feature class and have an undetermined receptor field structure by presenting a predetermined pattern to a data input layer. Thus, a feature class necessary for subject recognition can be learned automatically and efficiently.

An approach for phoneme recognition is described. A sequence of intermediate output posterior vectors is generated from an input sequence of cepstral features using a first layer perceptron. The intermediate output posterior vectors are then downsampled to form a reduced input set of intermediate posterior vectors for a second layer perceptron. A sequence of final posterior vectors is generated from the reduced input set of intermediate posterior vectors using the second layer perceptron. Then the final posterior vectors are decoded to determine an output recognized phoneme sequence representative of the input sequence of cepstral features.

A parallel processing device that computes a hierarchical neural network, the parallel processing device includes: a plurality of units that are identified by characteristic unit numbers that are predetermined identification numbers, respectively; a distribution control section that, in response to input as an input value of an output value outputted from one of the plurality of units through a unit output bus, outputs control data including the input value inputted and a selection unit number that is an identification number to select one unit among the plurality of units to the plurality of units through the unit input bus; and a common storage section that stores in advance coupling weights in a plurality of layers of the hierarchical neural network, the coupling weights being shared by plural ones of the plurality of units. Each of the units includes: a data input section that receives control data as an input from the distribution control section through the unit input bus; a unit number match judgment section that judges as to whether a selection unit number included in the control data inputted in the data input section matches the characteristic unit number; a unit processing section that, based on an input value included in the control data inputted in the data input section, computes by a computing method predetermined for each of the units; and a data output section that, when the unit number match judgment section provides a judgment result indicating matching, outputs a computation result computed by the unit processing section as the output value to the distribution control section through the unit output bus, wherein, based on the coupling weights stored in the common weight storage section, the unit processing section executes computation in a forward direction that is a direction from an input layer to an output layer in the hierarchical neural network, and executes computation in a backward direction that is a direction from the output layer to the input layer, thereby updating the coupling weights.

A system and method for emotion-enhanced natural speech using dilated convolutional neural networks, wherein an audio processing server receives a raw audio waveform from a dilated convolutional artificial neural network, associates text-based emotion content markers with portions of the raw audio waveform to produce an emotion-enhanced audio waveform, and provides the emotion-enhanced audio waveform to the dilated convolutional artificial neural network for use as a new input data set.

An output value of neuron within an objective layer of a hierarchical neural network is computed. The data of the output value of neuron is stored in a memory only if the output value of neuron is greater than or equal to a predetermined value by referring to the computed output value of neuron within the objective layer. When the data of the output value of neuron on a former layer of objective layer is read from the memory, the data having a predetermined value is read, instead of the data of the output value of neuron not stored in the memory.

Generating loosely coupled parts by forming couplings between partial nodes in a hierarchical neural network in accordance with a check matrix of an error correcting code.

A connected layer feature is generated by connecting outputs of a plurality of layers of a hierarchical neural network obtained by processing an input image using the hierarchical neural network. An attribute score map representing an attribute of each region of the input image is generated for each attribute using the connected layer feature. A recognition result for a recognition target is generated and output by integrating the generated attribute score maps for respective attributes.

The invention discloses a training method of an image classification network, an image classification method and device and a server. The training method comprises the following steps: preparing a data set of pictures with labels as input in advance; Constructing corresponding hierarchical neural network structures according to different classification levels; And carrying out hierarchical training on each hierarchical neural network structure to obtain a parent class corresponding to the maximum probability value and probability values of the input pictures under the parent class belonging todifferent subclasses. According to the method and the device, the technical problem of overfitting caused by redundancy of a full connection layer due to very large data set classification data is solved. Through the training method provided by the invention, the phenomena of low network training speed and over-fitting of the network caused by too many full connection layer parameters are solved.According to the image classification method provided by the invention, hierarchical training is adopted, so that a subclass classification result can be more accurately obtained on a classificationresult of a parent class, and accurate classification is realized.

The invention provides a pipeline leakage detection method based on a hierarchical neural network. The neural network is layered, a rejection threshold and a rejection interval of each layer of neural network are set, a test sample is input into each layer of neural network in the leakage detection process, if one layer of neural network judges the category of the test sample, the classification process is finished, and if the layer of neural network rejects the test sample, the test sample is input to the next layer of neural network till the test sample obtains a category label. The method can effectively judge whether an existing pipeline leaks or not, accuracy of pipeline leakage detection is improved, and detection precision is good.

A domain linker region is predicted by inputting an amino-acid sequence of a protein whose structure is unknown in a hierarchical neural network having identified and learned the domain linker region. Also, the sequence characteristics of the linker domain is identified by a statistical method, and by combining the result with the secondary structure predicting method, a domain linker predicting method for an amino-acid sequence whose structure is unknown was constructed.

The invention relates to a network intrusion detection method based on a CNN-GRU hierarchical neural network, and the method comprises the steps: capturing a network flow data package, namely, a to-be-classified data package, through Wreshark software; performing data packet marking, preprocessing and data cleaning on the to-be-classified data packets, analyzing the data packets into decimal data, and converting the decimal data into a 40 * 40 single-channel grey-scale map to obtain a sample complete set; dividing the sample complete set into a training set and a test set, taking the single-channel grey-scale map matrix as an input vector, and establishing a CNN-GRU hierarchical neural network classification model through the training set; and after model training is completed, transmitting data of the test set are into the model, the model predicts the input data according to parameters obtained through training, and unknown network traffic is classified to judge whether the unknown network traffic is attack traffic. Experimental results show that the accuracy of the method for classifying the normal traffic and the attack traffic reaches 99.92%.

The invention discloses a text representation method and device based on a hierarchical neural network. The method comprises: converting each word forming a sentence into a vector; Inputting vectors corresponding to all words in the sentence into a neural network for aggregation, and outputting sentence representation corresponding to the sentence; Inputting all the sentence representations into aneural network to be aggregated, and generating document representations corresponding to all the sentence representations; And converting the document representation into a document classification vector through a full connection network, and obtaining prediction probability distribution of document classification based on the document classification vector. According to the method and a device,A hierarchical mechanism is introduced into a neural network model to solve a document representation problem for text classification; Interoperability of different tasks is better improved, a hierarchical neural system structure is fused into a neural network method, a new neural network model based on layering is caused, accuracy, performance and the like are obviously superior to those of an existing neural network model, and consumption is lower.

The invention provides an image structuring method and device. The method comprises the following steps: image characteristics are extracted based on a first neutral network, and the characteristics are represented via characteristic tensors; the following operations are performed based on a second neutral network which comprises a first layer neutral network and a second layer neutral network: based on the first layer neutral network, belong types of pixels in an image are detected according to the characteristic tensors, distance between the pixels in the image and a central point of an object to which the pixels belong is calculated, and a specific object to which the pixels in the image belong is determined according to the distance; based on the second layer neutral network, attributes of the pixels are analyzed according to the characteristic tensors and the belong types of the pixels, and attributes of the specific object are determined according to the attributes of the pixels and the specific object to which the pixels belong. Via the image structuring method and device, extra errors can be prevented from being introduced during image structuring detection object determining operation and object attribute determining operation, and accuracy of image structuring analysis can be improved in a marked manner.

The invention discloses a document-level sentiment classification method based on dynamic word vectors and a hierarchical neural network. The method comprises the following steps: obtaining a high-quality dynamic word vector by constructing and training a bidirectional language model; and inputting the obtained dynamic word vector into a hierarchical neural network to model the document, thereby obtaining a vector representation containing rich semantic information, and inputting the vector into a softmax function to classify the document. According to the sentiment classification method, thehigh-quality dynamic word vector is generated by adopting the bidirectional language model, and the hierarchical neural network is provided for modeling the document, so that the problem of insufficient semantic expression of the static word vector to the polysemy is solved, and the document modeling capability in the sentiment classification task is further improved.

A characteristic-data-extraction unit 13 extracts characteristic data containing changing information from song data, then an impression-data-conversion unit 14 uses a pre-learned hierarchical neural network to convert the characteristic data extracted by the characteristic-data-extraction unit 13 to impression data and stores it together with song data into a song database 15. A song search unit 18 searches the song database 15 based on impression data input from a PC-control unit 19, and outputs the search results to a search-results-output unit 21.

An artificial neural network for detecting and identifying errors in digital circuits is provided. Data from digital circuits are received and organized into current data set patterns by a supervisory control and data acquisition system. The supervisory control and data acquisition system transmits the current data set patterns to an artificial neural network error detection module. The artificial neural network error detection module compares the actual output of each current data set pattern to a calculated output for a corresponding stored data set pattern. The artificial neural network error detection module determines whether a match condition exists for the comparison. The artificial neural network error detection module outputs the results of the determination to a user interface.

The present invention provides a construction method of a large-scale hierarchical neural network. The construction method of the large-scale hierarchical neural network comprises: designing a neural unit; designing a connection among a neuronal population; designing an information transmission mechanism among the neuronal population; and obtaining the large-scale hierarchical neural network according to the connection of the neuronal population and the information transmission mechanism. The construction method of a large-scale hierarchical neural network provided by the embodiment of the invention may perform repetition of human brain working process through the construction of the large-scale hierarchical neural network, therefore the problems that images cannot be identified or cannot be clearly identified because of the rotation and the scaling of objects may be solved.

In a hierarchical neural network having a module structure, learning necessary for detection of a new feature class is executed by a processing module which has not finished learning yet and includes a plurality of neurons which should learn an unlearned feature class and have an undetermined receptor field structure by presenting a predetermined pattern to a data input layer. Thus, a feature class necessary for subject recognition can be learned automatically and efficiently.

An output value of neuron within an objective layer of a hierarchical neural network is computed. The data of the output value of neuron is stored in a memory only if the output value of neuron is greater than or equal to a predetermined value by referring to the computed output value of neuron within the objective layer. When the data of the output value of neuron on a former layer of objective layer is read from the memory, the data having a predetermined value is read, instead of the data of the output value of neuron not stored in the memory.

A parallel processing device that computes a hierarchical neural network includes: a plurality of units identified by a characteristic unit numbers; a control section that outputs control data, including an input value and a selection unit number, to the plurality of units; and a storage section that stores a plurality of coupling weights, each of the coupling weights being associated with layer information. Each of the units includes: a data input section that receives control data from the control section; a unit number match judgment section that judges whether the selection unit number matches the characteristic unit number; a unit processing section that computes the output value; and a data output section that outputs the output value to the control section when the unit number judgement section judges that the selection unit number matches the characteristic unit number.

A system and method for emotion-enhanced natural speech using dilated convolutional neural networks, wherein an audio processing server receives a raw audio waveform from a dilated convolutional artificial neural network, associates text-based emotion content markers with portions of the raw audio waveform to produce an emotion-enhanced audio waveform, and provides the emotion-enhanced audio waveform to the dilated convolutional artificial neural network for use as a new input data set.

The invention discloses a sea surface target one-dimensional range profile noise reduction convolutional neural network identification method, and belongs to the field of radar signal processing. Aiming at a low signal-to-noise ratio condition, original HRRP data is reasonably pre-processed to construct multiple types of sea surface target data sets under different signal-to-noise ratio conditions, a one-dimensional noise reduction convolutional neural network is constructed by using a deep learning technology, the signal-to-noise ratio of the low signal-to-noise ratio data is improved on the basis of keeping the high signal-to-noise ratio data free of fluctuation, and the residual structure of a convolutional neural network is utilized to reduce the learning burden of the deep neural network, so that an intelligent sea surface target classification and recognition model integrating noise reduction and classification is constructed, the recognition accuracy of the sea surface target is improved, the sea surface target recognition performance under the condition of low signal-to-noise ratio is improved, the classification and identification capability of the sea radar in a complex sea surface environment is enhanced, and the method has popularization and application values.

The invention requests to protect a text abstract generation system and method based on adversarial learning and a hierarchical neural network, and belongs to the field of text abstracts of natural language processing. The system comprises a discriminator module, a preprocessing module, a word embedding module, a sentence embedding module, a generation module and an adversarial learning module. According to the invention, on the basis of an encoder decoder model (Seq2Seq), a new hierarchical division model is provided. An encoder part of the Seq2Seq is divided into a word embedding layer and asentence embedding layer, and an enhanced memory mechanism is introduced into each layer, so that the model can better understand text meanings, adversarial learning is introduced during decoding, arecognizer is arranged to recognize standard representation and fuzzy representation, the distance between the standard representation and the fuzzy representation is shortened, and meanwhile, learning is supervised to prevent the standard representation and the fuzzy representation from approaching, confrontation is formed, and when confrontation is balanced, an optimal generation result is found, so that the text abstract generation accuracy is improved.

The invention discloses a hierarchical neural network query recommendation method and device. The hierarchical neural network query recommendation method comprises the steps of: performing query recommendation on a hierarchical neural network; establishing two neural networks, namely a session layer neural network for modeling the short-term query record of the user and a user layer neural networkfor modeling the long-term query record of the user, wherein the state vector of the session layer neural network at the current moment is used as the input of the user layer neural network at the current moment, and the state vector of the user layer neural network at the current moment is used as the input of the session layer neural network at the next moment; and outputting query recommendation content for a query session according to the session layer neural network and the user layer neural network. According to the scheme provided by the invention, the query recommendation efficiency and accuracy can be improved.

The invention provides a category prediction method and device based on theme information. The method is applicable to criminal name prediction of legal documents and includes, on the basis of the hierarchical neural network, describing the overall association between each vocabulary in each sentence and the case text according to the theme information of the case description text, determining theimportance of each sentence through the theme information, performing weighted summation to obtain semantic vector representation of the case description text, and inputting the semantic vector representation into a classifier to predict a criminal name corresponding to the case. According to the method, key words and sentences in the case description text are mined by using theme information, and more effective semantic vector representation of the case description text can be obtained, so that a better effect is achieved on low-frequency criminal name prediction.

An apparatus for, by inputting data to a hierarchical neural network and performing operation processing in each layer of the network, calculating a feature plane in the layer, comprises an operation unit, a feature plane holding unit including at least one memory that holds a feature plane to be processed, a unit configured to control to arrange the feature plane in the memory based on network information as information about each layer undergoing the operation processing and to manage reading / writing from / in the memory, and a processor configured to access, via a bus, the feature plane holding unit which is address-mapped in a memory space. The processor calculates, based on the network information, an address address-mapped in the memory space, reads out the feature plane, and processes the feature plane.

The invention discloses a method for detecting the content of heavy metals in lettuce based on multi-scale images. The method is technically characterized by comprising the following steps of: 1, acquiring images of the lettuce by using three cameras, and normalizing the images; and 2, establishing a neural network learning model by taking the normalized lettuce image obtained in the step 1 as input, inputting the collected sample data into the neural network for learning, finally obtaining a multi-scale hierarchical neural network learning model, detecting and identifying the input lettuce image by using the multi-scale hierarchical neural network learning model obtained by learning, and judging whether the heavy metal content exceeds the standard or not. Data are collected through the optical camera, nondestructive and non-contact detection of the content of the heavy metal in the edible lettuce is achieved through analysis and modeling of imaging data of the camera, the speed is high, operation is convenient, environmental protection and safety are achieved, the industrial requirements of modern food safety can be met, and effective replacement of a traditional detection method is achieved.

The invention discloses a SAR image classification method based on layered sparse filter convolutional neural network. The steps are: 1. Divide the SAR database sample set into training data set and test sample set; 2. Learn the first layer sparse dictionary from the training data set; 3. Use the first layer sparse dictionary to extract the first layer sparse feature map and perform Non-linear transformation; 3. Learning the second-layer sparse dictionary from the first-layer nonlinear transformation feature map; 4. Using the second-layer sparse dictionary to extract the second-layer sparse feature map and performing nonlinear transformation; 5. Cascade first, The second-layer nonlinear transformation feature trains the SVM classifier; 6. Use the first and second-layer sparse dictionaries to extract the sparse features of the test set, and use the SVM classifier to classify. The invention solves the problems of complex design, poor universality and noise resistance, and low classification precision in the prior art, and can be used for target recognition.