152 results about "Backward propagation" patented technology

The invention relates to a method and a system for detecting and defending multichannel network intrusion, wherein a half-polling manner is used in a network card driver at a network side to capture a data packet and from a rule for judging abnormal traffic. At an operating system side, secure access control is realized by additionally setting authority control in a Capability module of Linux, meanwhile, monitoring and control treatment is performed on a kernel layer of an operating system, and the monitoring of Trojan and other abnormal operations or virus destruction is realized by carrying out credential privilege arbitration on processes, operating an i-node and carrying out secure control on an application layer. Formed feature data are gathered and sent to a Bayes model for classification, an improved backward propagation neural network (BPNN) is guided in to carry out data training so as to make the produced rule capable of defending corresponding attacks. The system disclosed by the invention comprises three modules and five submodules, wherein the three modules include a network data packet processing module, an operating system layer detecting module ad a detection center module.

The present invention discloses a deep learning-based insulator identification method. The insulator identification method comprises the steps of pre-processing an aerial image, and secondly, extending the data via the methods, such as the geometric transformation, the contrast enhancement, an analog noise adding method, etc.; acquiring the insulator samples, aiming at the insulators of different types, classifying to acquire; determining a to-be-trained model structure; inputting the samples in the to-be-trained model, and continuously adjusting the weights and the bias parameters by the forward propagation and backward propagation methods, and finally determining an optimal model parameter, based on the trained model, taking a to-be-detected image as an input signal, and by the network multi-layer convolution, pooling and full-connection operations, obtaining a final detection identification result. According to the present invention, by a deep learning method, the insulator characteristics are learned continuously, a learning network model is determined, the different insulators are identified under different background environments, and support is provided for the electric power maintenance decisions.

The invention discloses a radar echo extrapolation method based on a dynamic convolution neural network. The method comprises a step of offline convolutional neural network training which comprises the steps of carrying out data preprocessing on a given training image set to obtain a training sample set, initializing a dynamic convolution neural network model, training a dynamic convolution neural network by using the training sample set, calculating an output value through network forward propagation, and updating network parameters through backward propagation such that the dynamic convolution neural network converges. The method also comprises a step of online radar echo extrapolation which comprises the steps of converting a test image set into a test sample set through data preprocessing, testing the trained dynamic convolution neural network by using the test sample set, and carrying out convolution of a laster radar echo image inputted into an image sequence and a probability vector obtained in the network forward propagation to obtain a predicted radar echo extrapolation image.

The invention discloses a weak supervision fine-grained image classification method of a multi-branch neural network model. The weak supervision fine-grained image classification method is characterized by the steps of: firstly, randomly dividing a fine-grained image data set into a training set and a test set in proportion; secondly, positioning a local region with potential semantic informationby using a local region positioning network; thirdly, respectively inputting an original image and the positioned local region into a deformable convolution residual network and a rotation invariant coding direction response network to form a feature network of three branches, respectively training the three branches, and respectively carrying out backward propagation learning on the three branches based on cross entropy loss; and finally, combining intra-branch loss and inter-branch loss to optimize the whole network, and performing classification prediction on the test set. According to theweak supervision fine-grained image classification method, the negative influence of various changes such as posture, visual angle and background interference on a classification result is reduced, and a better effect is achieved on a fine-grained image classification task.

The invention discloses an image content question and answer method based on multi-modality low-rank dual-linear pooling. The image content question and answer method comprises the following steps of 1, conducting data preprocessing on an answer text of a question described with an image and a natural language; 2, conducting characteristic fusing on a multi-modality low-rank dual-linear pooling model; 3, establishing a neural network structure based on an MFB pooling model and a collaborative concern model; 4, training the models and utilizing a backward propagation algorithm to train neural network parameters. The neural network model aiming at an image question & answer is provided, and especially a method for conducting unified modeling on cross-media data in a question-answer in the field of the image question and answer and a network structure for studying the 'collaborative concern' to conduct modeling description on an image and question fine-grained characteristics are provided, and the best effect currently in the field of the image question and answer is obtained.

The invention provides an image recognition and recommendation method based on neural network depth learning. The method obtains pictures and classification from an image database, inputs to a convolution neural network, trains the neural network through repeated forward and backward propagation, improves image recognition accuracy, and extracts a 20-layer neural network model. By using this model, the object recognition and classification is carried out by collecting static pictures. Results are recognized, and by combining with the personalized characteristics of the input, the input probability of interest is analyzed. By using the machine learning model based on the effective recognition and classification of the material cloud database, and using the recommendation system algorithm, the predicted content material is pushed to the image inputter for cognitive learning. The method of the invention has the advantages of high image recognition rate, multiple recognition types and accurate content recommendation, and can be applied to the electronic products of a computer with a digital camera, a mobile phone, a tablet and an embedded system, so that people can photograph and recognize the objects seen in the eyes and actively learn the knowledge of recognizing the objects.

The invention discloses a clothes classifying method based on a convolutional neural network. The method comprises the following steps of acquiring clothes image samples, and dividing the samples into training samples and testing samples; preprocessing the training samples and the testing samples; constructing a convolutional neural network model; performing training of two stages including a forward propagation stage and a backward propagation stage on the convolutional neural network model through preprocessed training samples, finishing the training when the error calculated during the backward propagation stage reaches a desired value, and acquiring a parameter of the convolutional neural network model; testing the preprocessed testing samples by using the trained convolutional neural network model and outputting final clothes classifying results. The convolutional neural network model can make clothes images directly serve as network inputs, extract image features in an implicit way and establish a global feature expression. Compared with a manually designed feature extraction way, the method is more convenient and accurate. The problem that a conventional algorithm leads to low clothes classifying accuracy is solved.

The invention discloses an OIN (Optimal Input Normalization) neural network training method for a mixed SVM (Support Vector Machine) regression algorithm. The method mainly comprises the steps of an OIN forward propagation part, a SVM regression part and an OIN backward propagation part; and finally, an optimized OIN / SVM mixed model is obtained through various trainings; and in a testing stage, a testing sample is input into the optimized OIN / SVM mixed model to obtain a predicated result, so that the classification of the sample or the regression of the sample is predicated. According to the invention, a latest designed OIN artificial neural network training method is adopted; and through the adoption of the method disclosed by the invention, the convergence of the traditional backward propagation algorithm can be greatly improved.

Systems and method of compensating for transmission impairment are disclosed. One such method comprises receiving a wavelength-division multiplexed optical signal. The received optical signal has been distorted in the physical domain by an optical transmission channel. The method further comprises propagating the distorted optical signal backward in the electronic domain in a corresponding virtual optical transmission channel. The backward propagation fully compensates for fiber dispersion, self-phase modulation, and cross-phase modulation (XPM) and partially compensates for four-wave mixing (FWM).

The invention discloses a deep learning-based face recognition and face verification supervised learning method. The method comprises the following steps: a soft maximum loss function is used to increase a between-class distance for full connection layer output characteristics of a convolutional neural network model, a center is learnt for the depth characteristics of each class through a centralloss function, a super parameter is used to balance the two functions to thus jointly supervise the learning characteristics; backward propagation of the convolution neural network model is calculated, a stochastic gradient descent algorithm based on minimum batch processing is adopted to optimize the convolutional neural network model, and a weight matrix and the depth characteristic center of each class are updated; and after the depth characteristics are subjected to principal component analysis and dimension reduction, the cosine distance between each two characteristics is calculated to calculate a score, wherein the score is used for target matching in nearest neighbor and threshold comparison, and a face is recognized and verified. The identification ability of the neural network learning characteristics can be effectively improved, and a face characteristic recognition and face verification mode with robustness is acquired.

The invention discloses a method for identifying a genetic relationship of figures based on a self-encoder. The method comprises the following steps: inputting face images and pre-treating the face images; confirming identity characteristics of figures according to the face images; constructing the self-encoder and forming a self-encoder neural network; repeatedly performing forward propagation and backward propagation on the identity characteristics in the self-encoder neural network; updating weight till a cost function is minimized and a related characteristic of the identity characteristics is acquired; and identifying the genetic relationship between the face images according to the related characteristic. According to the invention, the genetic relationship between the figures can be identified.

An embodiment of the invention discloses a method for predicting a failure of an aero-engine, which comprises the following steps: collecting parameter data in an engine recording system to obtain anengine vibration signal, and determining a training data set and test data based on the parameter data; constructing an LSTM neural network model by using a python platform, and determining a signal margin spectrum according to the training data set and the LSTM (Long Short Term Memory) neural network; constructing a BP neural network model by using the python platform, training a BP (Backward Propagation) neural network by using a marginal spectrum characteristic training set, and acquiring and storing parameters of the BP neural network after convergence of training; and predicting a failureof the aero-engine according to the trained BP neural network model. Through adoption of the method of the invention, the failure of the aero-engine can be predicted.

According to the embodiment, the invention provides a compressed speech recognition model optimizing method. The method comprises the following steps: based on a speech recognition model before compressing, confirming a teacher model, and generating a student model based on the compressed speech recognition model and un-annotated speech data in a speech database; extracting annotated speech data sequences from the speech database, so that a training data set is obtained, implementing neural network forward propagation on the student model via the training data set, and conforming a first posterior probability of the student model; implementing forward-backward computing on the teacher model via the training data set, and confirming a second posterior probability of the teacher model; comparing the first posterior probability and the second posterior probability, and conforming errors between the student model and the teacher model; and implementing neural network backward propagation on the student model in accordance with the errors when the errors are not convergent, so that the student model is optimized. According to the embodiment, the invention also provides a compressed speech recognition model optimizing system. In the embodiment, the compressed model is optimized in accordance with a source model.

The invention discloses an image matting method and device, and belongs to the field of digit image matting. The method includes: inputting an input picture to a preset full-convolution network to obtain a prediction value Fs belonging to a foreground area, a prediction value Bs belonging to a background area and a prediction value Us belonging to an unknown area of each pixel in the input image,calculating a foreground probability matrix F and a background probability matrix B corresponding to the input image according to the Fs, Bs and Us of each pixel in the input image, inputting a presetimage matting realization function into the foreground probability matrix F and background probability matrix B to perform the image matting to obtain a transparency value matrix of the input image.The image matting realization function is obtained after utilizing a first sample image which employs a preset backward propagation algorithm to perform a training, the more accuracy image matting result is obtained without performing a three-value diagram calibration to the input image for multiple times, and the automatic digital image matting is realized.

In a light amplification system and technique, a pump source provides pump power at a source wavelength. The pump power is launched as an input into a cascaded Raman resonator. A wavelength-dependent loss element is connected such that it precedes the cascaded Raman resonator. The wavelength-dependent loss element is configured to transmit light power at the source wavelength with low loss, and to provide high loss at the first Stokes shift. The wavelength-dependent loss element prevents buildup of light power between the pump source and the cascaded Raman resonator, thereby preventing backward propagation of light power back into the pump source.

Polarization independent optical isolator / circulator devices based on Mach-Zehnder interferometers. The devices utilize either polarization splitting and nonreciprocal polarization conversion or nonreciprocal phase shift within the interferometric arm. For devices with nonreciprocal phase shift, the relative phase difference is 0° in the forward propagation direction and 180° in the backward propagation direction, or vice versa, so that light goes into a bar or cross port depending on the propagation direction. The devices have advantages over previous designs in the use of inexpensive device components, simple alignment, minimal space requirement, and negligible polarization mode dispersion or polarization dependent loss. In addition, the devices can be made in a waveguide form with minimal loss and with high fabrication ease. An additional phase compensator and / or a variable attenuator can be integrated in order to relax the fabrication tolerances.

The invention discloses a digital holographic compression transmission method of a quantum backpropagation neural network. The original image is recorded as a digital hologram through a Fresnel off-axis holographic calculation method, a quantum backpropagation (QBP) neural network is constructed, and a digital hologram is used. Train the QBP neural network, use the trained QBP neural network to compress, transmit and decompress the digital hologram to obtain a reconstructed hologram, and reproduce the reconstructed hologram to obtain a reconstructed image. The method provided by the invention has faster parallel processing speed and stronger ability to store data, and is suitable for the calculation and processing of digital holograms with a large amount of data; it can use fewer training times to complete the training of the compressed transmission network structure, and improve the digital hologram. The compression transmission speed of the hologram; the image compression rate is adjustable, and the reproduced image quality is better.

The invention discloses an adaptive digital signal processing algorithm for compensating optical fiber transmission nonlinear damages, and relates to the field of optical fiber transmission. The algorithm comprises following steps of setting three nonlinear coefficients after a receiving end receives signals; respectively substituting into an adaptive digital backward propagation algorithm for processing, thus obtaining three signal amplitude values as initial values; estimating the nonlinear coefficients by utilizing the strength variances of the signals, wherein after many times of iteration, an optimum nonlinear coefficient gamma <opt > for nonlinear compensation can be estimated; taking the optimum nonlinear coefficient gamma <opt > as the input of adaptive digital backward propagation; then stopping iteration; carrying out CR (carrier recovery) and decode processing to the signals, thus recovering the signals of a transmitting end; and calculating a BER (bit error ratio). The algorithm provided by the invention can be applied in coherent light communication system of an M-QAM (M-ary quadrature amplitude modulation) high order modulation format; the optimum nonlinear coefficient can be determined more precisely; and the algorithm complexity is reduced.

The invention provides a vibration model establishment method applied to a fiber sensing system. The method comprises steps that a vibration signal is adjusted through a fiber vibrator into a fiber interference signal which is outputted, intensity and the period of the output signal are utilized in combination with the signal generation time and a signal generation distance, a signal state graph is established through taking a distance between the fiber position and a fiber start point as a first coordinate axis and the time as a second coordinate axis; signal state graphs under different conditions are acquired to carry out manual label marking for the data, the data information including manual marking labels and the signal state graphs is inputted to a convolutional neural network model to carry out image identification and CNN training, and a vibration model applied to the fiber sensing system is acquired; image identification and CNN training of the convolutional neural network model comprise two stages, including forward propagation and backward propagation. The method has advantages of high identification rate, good identification accuracy and strong identification capability.

Systems, apparatuses, and methods for managing buffers in a neural network implementation with heterogeneous memory are disclosed. A system includes a neural network coupled to a first memory and a second memory. The first memory is a relatively low-capacity, high-bandwidth memory while the second memory is a relatively high-capacity, low-bandwidth memory. During a forward propagation pass of the neural network, a run-time manager monitors the usage of the buffers for the various layers of the neural network. During a backward propagation pass of the neural network, the run-time manager determines how to move the buffers between the first and second memories based on the monitored buffer usage during the forward propagation pass. As a result, the run-time manager is able to reduce memory access latency for the layers of the neural network during the backward propagation pass.

The invention relates to a method for identifying an airplane structure load based on flight parameter monitoring. The method comprises the following five steps: (1), recording a flight parameter and load data of an airplane structure dangerous part through a flight test; (2), performing correlation analysis on the flight parameter and the load according to the flight parameter and the load data, and selecting a load identification parameter; (3), establishing a flight parameter and load identification model according to the flight parameter and the load data of an airplane by utilizing a polynomial reconstruction technology; (4), establishing a flight parameter and load identification model according to the flight parameter and the load data of the airplane by utilizing a backward propagation artificial neural network method; and (5), substituting the flight parameter obtained by a flight parameter sensor into a polynomial identification and artificial neural network identification model so as to obtain a load to be identified. The method disclosed by the invention has the characteristics of being high in calculation precision, low in cost, convenient and rapid and capable of obtaining the load data of the airplane structure dangerous part in real time; and thus, requirements of health management and residual life detection can be satisfied.

The invention provides a channel estimation method based on neural network prediction. The method comprises: firstly, constructing a neural network based on a channel estimation method of neural network prediction to approach an ideal channel state, and using a large amount of orthogonal frequency division multiplexing (OFDM) channel data for training the neural network; secondly, predicting channel states of other limited data positions through a neural network according to the received pilot signals; and finally, estimating the channel frequency domain response of the residual data positionby utilizing low-order spline interpolation. According to the invention, a backward propagation (BP) neural network is utilized, and the problem of channel estimation under an OFDM channel is solved.Due to the fact that the neural network has the self-adaptive self-learning capacity, the channel estimation precision of the system can be improved.

The invention relates to the technical field of deep learning algorithms and specifically relates to a hybrid precision deep learning algorithm. The to-be-solved technical problem is ensuring calculation precision and improving calculation efficiency at the same time. The technical scheme includes steps of: S101, utilizing a single-precision many-core processor for forward propagation calculation and calculating the value of each nerve cell for each network layer; S102, utilizing the single-precision many-core processor for backward propagation calculation and calculating error residual values for each network layer; S103, utilizing the single-precision many-core processor for calculating weight increment; S104, updating the weight increment calculated by the single-precision many-core processor to weight increment calculated by a high-precision many-core processor and implementing calculation of an irritation. The invention is suitable for the deep leaning field.

The invention belongs to the technical field of electronic signal processing, particularly relates to a training method and system for an RNN transducer model, and a device, and aims to solve the problem that the RNN transducer model cannot learn alignment information of speech data well. The training method comprises the following steps of extracting the features of speech training data to obtaina speech feature sequence; forcedly aligning the speech feature sequence through a GMM-HMM (Gaussian Mixture Model-Hidden Markov Model) model to obtain an alignment annotation, and splicing differentframes of speech features; training the RNN transducer model based on the spliced speech feature sequence and text annotation training data to obtain the probability distribution and the negative logarithmic loss value of each word in a preset word list; acquiring an alignment loss value; carrying out weighted averaging on the alignment loss value and the negative logarithmic loss value to obtaina combined loss value, and updating the parameters of the model through a backward propagation algorithm; and iteratively training the model. Through the adoption of the training method and system, the alignment information of the speech data can be learned accurately.

The invention discloses an audio and video mutual retrieval method based on user click behaviors. The method comprises the following steps: preprocessing input audio and video data; Sending the preprocessed audio data into a deep convolutional neural network to obtain an audio representation vector and attention weight distribution; Sending the preprocessed video key frame into a deep convolutional neural network to obtain a key frame representation vector, and sequentially sending the key frame representation vector into a time sequence processing network based on an attention mechanism to obtain a representation vector of the video and attention weight distribution; Calculating the similarity of the audio and video representation vectors and sorting the audio and video according to the similarity; performing Annotating according to the attention weight distribution to provide explainable basis for sorting; Calculating the loss function through a user click behavior, and carrying outmodel training by adopting a backward propagation method; And carrying out retrieval matching on audios and videos in the media library based on the trained model. According to the method and the device, matched audios and videos in the media library can be retrieved under the condition of giving videos and audios.

Disclosed is a light illumination measurement and intelligent control system based on a convolutional neural network. The system comprises an image acquisition device, a sample database creation unit,a convolutional neural network training unit and a single-chip microcomputer intelligent control unit; the image acquisition device includes a camera, a video line and a main unit; the sample database creation unit is used for obtaining a class label of a test sample set and the test sample set; the convolutional neural network training unit is used for inputting sample images into a convolutional neural network model, conducting a convolution operation through a convolution layer, completing a downsampling operation through a pooling layer, conducting forward propagation to calculate neuronoutput values and error backward propagation to adjust weight values, inputting training data for simulation verification, and determining completion of the network training; the single-chip microcomputer intelligent control unit is used for comparing the ambient light illumination with the optimal light illumination and selectively adjusting the output of a single-chip microcomputer according toa comparison result to achieve control over the ambient light illumination. The system is high in measurement precision and good in adaptability to the measurement environment.

The invention discloses a method for realizing the range-based localization of an indoor target based on Taylor series expansion. With the method adopted, the problem of low positioning accuracy of wireless transmission signal-based indoor localization technologies under a wireless signal transmission model can be solved. The method includes the following steps that: 1, the ranging reference point of the indoor target is set; 2, the valid ranging signal sample set of the indoor target is constructed; 3, the number of valid ranging access points of the indoor target is calculated; 4, whether the total number of the valid ranging access points is larger than 3 is judged; 5, whether the total number of the valid ranging access points is larger than 5 is judged; 6, the position coordinates of the indoor target are calculated through adopting the least squares method; and 7, the Taylor series expansion is adopted to obtain a locating circle with the indoor target adopted as a center, and the coordinates of the centroid of the locating circle are calculated and are adopted as the position coordinates of the indoor target. According to the method of the invention, a backward propagation neural network model and the Taylor series expansion are combined, so that the localization accuracy of the indoor target can be improved.

The invention provides a BP (Back Propagation) neural network algorithm based method for analyzing coating aging. The method has the advantages of higher flexibility and forecast precision and better hereditability and comprises the processes of signal forward propagation and error backward propagation, wherein in the forward propagation process, an input sample is imported from an input layer and then transmitted to an output layer after the sample is processed through various buried layers layer by layer; if the actual output of an output layer does not accord with an expected output, the process is turned to an error backward propagation stage; in the error backward propagation process, an output error is backwards transmitted to an input layer through the buried layers in a certain form layer by layer, and the error is shared by all units of all the layers so as to obtain error signals of all the units of all the layers, wherein the error signals are used as references for correcting the weight values of all the units; and the weight value adjustment process of all layers of signal forward propagation and error backward propagation is carried out in cycles until network output errors are reduced to an acceptable degree or a preset number of times is finished. The method is characterized in that a momentum item delta W(t)=eta delta X+alpha delta W(t-1) is added, wherein alpha is a momentum factor alpha belonging to the set of (0, 1); the learning rate is adaptively regulated, if a total error E rises after the adjustment of a batch of weight values, eta is equal to beta eta (theta>0), and if the total error E drops after the adjustment of a batch of weight values, eta is equal to theta eta (theta>0); and a steepness factor is introduced, and when an error curve plane enters a flat area, a changed output quantity is set, wherein lambada is the steepness factor, in the flat area, lambada is larger than 1, and after quitting the flat area, lambada is equal to 1.

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 decentralized parameter server optimization algorithm, which is characterized in that a parameter server in a centralized parameter server algorithm in the prior art is removed, working threads of an original parameter server are dispersed into each working node, and when model training is carried out, a random gradient descent algorithm is adopted to carry out parameterupdating of multiple iterations. The iteration process comprises the following steps of firstly, carrying out gradient calculation on each working node by adopting forward propagation and backward propagation; then, each working node collecting gradients of a network layer set by a part of models in charge of the distributed working threads of the parameter server and updating the parameters by using a random gradient descent algorithm; and finally, taking out the updated parameters of all the working nodes to form parameters of a global network model, and performing multiple iterations to complete model training.