1382results about How to "Increase training speed" patented technology

The invention provides a deeply differential privacy protection method based on a generative adversarial network. The deeply differential privacy protection method is adopted to solve the problem that attackers use own coding and other methods for restoring training set data when a deep learning model is in use, and by adopting the deeply differential privacy protection method, the purpose of protecting users' privacy in a training dataset is achieved. The method includes the steps that according to the potential dataset scale of the input training dataset, a sensitivity degree and the maximum attacker attacking probability are queried to calculate an upper bound of privacy budgets; during deep network parameter optimizing calculation, a differential privacy thought is integrated, noise data is added, on the basis of the characteristic that differential privacy and Gaussian distribution can be combined, the privacy budget of each layer of a deep network is calculated, and during stochastic gradient descent calculation, Gaussian noise is added to make the overall privacy budget minimum; the generative adversarial network is adopted to generate optimal results which the attackers can get, by comparing attack results with initial data, feedback regulation is conducted on parameters of a deeply differential privacy model, and the balance between dataset availability and privacy protection degrees is achieved.

The invention discloses a deep learning super-resolution reconstruction method based on residual sub-images; residual sub-images are effectively combined with deep learning method based on convolutional neural network, super-resolution reconstructed images are clearer, and reconstruction speed is higher. By increasing the depth of convolutional neural network, a network model acquired by learning has higher nonlinear expression capacity and image reconstructing capacity; in addition, by introducing residual sub-image process, preprocessing based on traditional interpolation algorithm is removed, and fuzzy effect due to the interpolation algorithm is avoided. By making ingenious use of residual sub-images, it is possible to transfer deep learning convolutional operation from high-resolution space to low-resolution space, and accordingly reconstruction efficiency of super-resolution algorithm is increased at the premise of improving super-resolution reconstruction effect.

The invention provides a parameter synchronization optimization method and system suitable for distributed machine learning. A machine learning algorithm achieved in a parameter server distribution mode is used for overcoming bottlenecks, such as a large amount of parallel machine learning training time delay caused by insufficient network and parameter server resources, of an existing algorithm in the parameter synchronization process. The system comprises a resource monitoring and distributing module at the parameter server end, a parameter maintaining module at the parameter server end, server resource request modules of all working nodes, parameter synchronization time interval control modules of the working nodes, non-synchronization time accumulation modules of the working nodes, parameter calculation modules of the working nodes and parameter synchronization modules of the working node. According to the parameter synchronization optimization method and system, different synchronization time intervals are selected for the different working nodes to avoid request emergency situations by monitoring resource occupancy conditions of a parameter server; meanwhile, it is guaranteed that the selected time intervals can meet the requirements for communication frequency reducing and training accurate rate guaranteeing at the same time, and the bottlenecks of an existing distributed machine learning system in the parameter synchronization process are effectively avoided.

The invention provides a human body behavior identification method based on an inertial sensor. The method comprises the steps of acquiring human body behavior data of testees by means of the inertial sensor, conducting sliding window segmentation on the acquired human body behavior data, conducting feature extraction on a triaxial accelerated speed subsequence and a triaxial angular speed subsequence which are obtained after sliding window segmentation, conducting feature fusion on a feature vector to form a sample set of the human body behaviors of the testees, conducting feature selection on the sample set by means of the least correlated maximum redundant algorithm and the Bayes regularization sparse polynomial logistic regression algorithm, obtaining the classification feature vectors of all human body behaviors of all the testees, obtaining a classifier model of each human body behavior by means of a fuzzy least square support vector machine, and obtaining a human body behavior identification result after the human body behavior data are tested by means of the fuzzy least square support vector machine. By means of the human body behavior identification method, self-adaptation and identification efficiency can be improved.

The invention discloses a method for improving a BP (back propagation) neutral network and based on a genetic algorithm. The method includes coding the BP network to determine structure of the neutral network, wherein the structure includes the number of hidden layers and the number of units of each layer; adopting real-number coding to code by taking weight and threshold as genes, wherein each neutral network corresponds to a chromosome after coding; using the genetic algorithm to perform selection optimization on the network, wherein selection optimization includes the steps of selection, crossing and variation; training the BP network to acquire a final result; decoding an optimal individual selected by the genetic algorithm to generate a new neutral network, and training the new network by applying a BP training algorithm to acquire a final result. The method combines the genetic algorithm with the BP network, thereby being capable of fully utilizing advantages of the both, the problem that initial weight and threshold of the BP network are difficult to determine can be solved, searching range can be narrowed, training speed of the BP network can be increased, and the problem of local minimum can be improved.

The invention discloses a text classification method based on a bidirectional cyclic attention neural network, and belongs to the technical field of learning and natural language processing. The method comprises the following steps: step 1, preprocessing data; Step 2, according to the preprocessed data, generating and training a word vector of each word through a Word2vec method; Step 3, performing text semantic feature extraction on the word vector according to the word vector, fusing an attention mechanism and a bidirectional recurrent neural network, calculating the overall weight of each word, and converting the weight into an output value Y (4) of a model; And step 4, taking the feature vector Y (4) as the input of a softmax classifier according to the feature vector Y (4), and carrying out classification identification. According to the method, the attention mechanism is fused in the text feature learning model, the effect of keywords can be effectively highlighted, the performance of the model is greatly improved, and the text classification accuracy is further improved.

The invention discloses a generative adversarial network-based pixel-level portrait cutout method and solves the problem that massive data sets with huge making costs are needed to train and optimizea network in the field of machine cutout. The method comprises the steps of presetting a generative network and a judgment network of an adversarial learning mode, wherein the generative network is adeep neural network with a jump connection; inputting a real image containing a portrait to the generative network for outputting a person and scene segmentation image; inputting first and second image pairs to the judgment network for outputting a judgment probability, and determining loss functions of the generative network and the judgment network; according to minimization of the values of theloss functions of the two networks, adjusting configuration parameters of the two networks to finish training of the generative network; and inputting a test image to the trained generative network for generating the person and scene segmentation image, randomizing the generated image, and finally inputting a probability matrix to a conditional random field for further optimization. According tothe method, a training image quantity is reduced in batches; and the efficiency and the segmentation precision are improved.

The invention discloses a collaborative filtering recommendation method and a collaborative filtering recommendation system of a convolution neural network integrating an attention model, which relates to the technical field of data mining recommendation, improves feature extraction efficiency and scoring prediction accuracy, reduces operation and maintenance cost, simplifies cost management mode,and is convenient for joint operation and large-scale promotion and application. The invention relates to a collaborative filtering recommendation method of a convolution neural network integrating an attention model, comprising the following steps: step S1, splicing a user feature vector and an item feature vector into a new vector; S2, sending the new vector as an input vector into the multi-layer perceptron to learn and predict the score; The attention model is fused into the object potential vector to obtain the convolution neural network of the object feature vector or the hidden layer of the multi-layer perceptron.

The invention discloses a multi-unmanned aerial vehicle path collaborative planning method and device based on hierarchical reinforcement learning. The method comprises the steps of extracting a characteristic space of each unmanned aerial vehicle in a plurality of unmanned aerial vehicles; layering tasks needing to be executed in task targets according to the task targets of the unmanned aerial vehicles in the plurality of unmanned aerial vehicles, and dividing the tasks into a plurality of subtasks, wherein each subtask is realized by a neural network; forming each neural network composed bythe plurality of subtasks, and initializing parameters of each neural network to obtain each initial neural network; associating each neural network; taking difference between output results and target output as a loss function; carrying out parameter updating on each neural network through gradient descent; finishing training each neural network when a value of the loss function is smaller thana given threshold or the appointed step number arrives; passing each neural network through utilization of characteristic vectors in the respective characteristic space in sequence, thereby obtainingeach output value; selecting an action which enables an operation value to be maximum as a control signal of each unmanned aerial vehicle, and realizing multi-unmanned aerial vehicle path collaborative planning.

The invention discloses a video enhancement and transmission method based on deep learning. Downsampling is carried out on a high-definition video at a video source end, thereby obtaining a low-definition video; the low-definition video is compressed in an existing video coding mode; and the compressed low-definition video is transmitted. The size of the video data is greatly reduced through downsampling and video coding, so the video traffic needing to be transmitted is correspondingly reduced, and an effect of reducing the bandwidth cost is achieved. At a user receiving end, a user receives the low-definition video, reconstructs the low-definition video by employing a super-resolution image reconstruction method of the deep learning and restores the low-definition video into a high-resolution video for the user to watch, so the video transmission bandwidth cost is effectively reduced. According to the method, the video transmission bandwidth cost is reduced by at least 50%; the resolution of the live broadcast video is improved; and the watching experience of a user is improved.

The invention provides a small-sample face recognition method which relates to a method for face recognition by means of electronic equipment. According to the face recognition method, a single-layer multi-scale convolutional neural network structure is used for performing multi-characteristic fusion and classification. The method comprises the steps of preprocessing a face image; extracting a multi-characteristic graph of the face image; extracting a DWT low-frequency subband graph, an Sobel edge characteristic graph and an LBP pattern characteristic graph; performing multi-characteristic fusion by means of the single-layer multi-scale convolutional neural network structure; and predicating a classification result by means of a Softmax classifier, thereby realizing face recognition. The small-sample face recognition method overcomes a defect of low classification recognition rate caused by a classification method in settling a small-sample face recognition problem.

The invention provides a medical image segmentation method based on an improved convolutional neural network, and belongs to the field of image processing. According to the method, an improved convolutional neural network is constructed, and a multi-channel fusion module, a multi-channel dense connection module, a down-sampling module and an up-sampling module are designed on the basis of an original framework of a standard U-Net network. The depth of the network is increased while the network training calculation amount is controlled, and redundant calculation is reduced. Gradient disappearance is avoided by improving the internal structure of the network while the network is deepened. The experimental result is that the segmentation DICE coefficient of the output prediction graph of thetrained neural network is 98.57%, the DICE coefficient segmented by the original network is 98.26%, and the effectiveness of the method is reflected.

The present invention discloses a facial emotion recognition method based on a depth sparse self-encoding network. The method comprises the steps of 1, acquiring and pre-processing data; 2, establishing a depth sparse self-encoding network; 3, automatically encoding / decoding the depth sparse self-encoding network; 4, training a Softmax classifier; and 5, finely adjusting the overall weight of the network. According to the technical scheme of the invention, sparseness parameters are introduced. In this way, the number of neuronal nodes is reduced, and the compressed representation of data can be learned. Meanwhile, the training and recognizing speed is improved effectively. Moreover, the weight of the network is finely adjusted based on the back-propagation algorithm and the gradient descent method, so that the global optimization is realized. The local extremum and gradient diffusion problem during the training process can be overcome, so that the recognition performance is improved.

A multi-level anomaly detection method based on exponential smoothing, sliding window distribution statistics and an integrated learning model comprises the following steps of a statistic detection stage, an integrated learning training stage and an integrated learning classification stage, wherein in the statistic detection stage, a, a key feature set is determined according to the application scene; b, for discrete characteristics, a model is built through a sliding window distribution histogram, and a model is built through exponential smoothing for continuous characteristics; c, the observation features of all key features are input periodically; d, the process is ended. In the integrated learning training stage, a, a training data set is formed by marked normal and abnormal examples; b, a random forest classification model is trained. The method provides a general framework for anomaly detection problems comprising time sequence characteristics and complex behavior patterns and is suitable for online permanent detection, the random forest model is used in the integrated learning stage to achieve the advantages of parallelization and high generalization ability, and the method can be applied to multiple scenes like business violation detection in the telecom industry, credit card fraud detection in the financial industry and network attack detection.

The present invention discloses a feature extraction and state recognition method for one-dimensional physiological signal based on depth learning. The method comprises: establishing a feature extraction and state recognition analysis model DBN of a on-dimensional physiological signal based on depth learning, wherein the DBN model adopts a "pre-training+fine-tuning" training process, and in a pre-training stage, a first RBM is trained firstly and then a well-trained node is used as an input of a second RBM, and then the second RBM is trained, and so forth; and after training of all RBMs is finished, using a BP algorithm to fin-tune a network, and finally inputting an eigenvector output by the DBN into a Softmax classifier, and determining a state of an individual that is incorporated into the one-dimensional physiological signal. The method provided by the present invention effectively solves the problem that in the conventional one-dimensional physiological signal classification process, feature inputs need to be selected manually so that classification precision is low; and through non-linear mapping of the deep confidence network, highly-separable features / feature combinations are automatically obtained for classification, and a better classification effect can be obtained by keeping optimizing the structure of the network.

The invention discloses a commodity target word oriented emotional tendency analysis method, which belongs to the field of the analysis processing of online shopping commodity reviews. The method comprises the following four steps that: 1: corpus preprocessing: carrying out word segmentation on a dataset, and converting a category label into a vector form according to a category number; 2: word vector training: training review data subjected to the word segmentation through a CBOW (Continuous Bag-of-Words Model) to obtain a word vector; 3: adopting a neural network structure, and using an LSTM(Long Short Term Memory) network model structure to enable the network to pay attention to whole-sentence contents; and 4: review sentence emotion classification: taking the output of the neural network as the input of a Softmax function to obtain a final result. By use of the method, semantic description in a semantic space is more accurate, the data is trained through the neural network so as to optimize the weight and the offset parameter in the neural network, parameters trained after continuous iteration make a loss value minimum, at the time, the trained parameters are used for traininga test set, and therefore, higher accuracy can be obtained.

The invention discloses a polarimetric SAR classification method based on a semi-supervised convolutional neural network. According to the method, first, image data is input, and super pixel segmentation is performed; training samples and test samples are extracted; a neighborhood preserving and sparse filtering deep learning network (NDSFN) is adopted to perform unsupervised pre-training on parameters of the convolutional neural network, and a needed filter set is obtained; a feature map is obtained through convolution operation; fuzzy processing is performed on the feature map through downsampling; parameter settings of the NDSFN are adjusted according to the filter size in a second convolution layer to obtain a new feature map; a Softmax classifier is utilized to perform image classification on the training samples; and a small quantity of marked samples are adopted for slight adjustment to obtain a polarimetric SAR terrain classification. Through the method, by constructing a novelsemi-supervised convolutional neural network (SNCNN) model, the problem that a traditional supervised convolutional neural network has a high requirement for marked samples is effectively solved.

The application discloses an information acquisition method and device. An embodiment of the method includes: acquiring a plurality of question-answer pairs in a data set, extracting at least one question word and at least one answer word of each question-answer pair; determining a context of the question word and the answer word; taking the question word, the answer word and the context as a training sample, training a preset model, and obtaining a word vector set; receiving question information to be responded; based on the word vector set, acquiring answer information matching the question information from the data set. The information acquisition method and device may train the word vector through evaluating a correlation of the question-answer pairs from a semantic aspect, and a training speed and accuracy of the word vector are improved. Based on that when the word vector acquires the matched information, a complicated supervised neural network training is not needed, the speed and accuracy of information acquisition can be improved.

The invention discloses a method of deep neural network based on a discriminable region for dish image classification. The method relates to the field of image processing, integrates a significant spectrum pooling operation, and fuses low-level features and high-level features in a network. The method adopts a convolution kernel filling operation, effectively preserves important information on characteristic spectra, and is matched with data dimensions of a full connection layer, so that the full connection layer can utilize a VGG-16 pre-training model at a training state, thereby improving the training efficiency and network convergence speed. Each image to be classified is subjected to normalization processing based on the model which is learned in a constructed database, the image is tested by using a trained convolutional neural network, the classification precision is measured by using Softmax loss, a classification result of the image is obtained, real categories and predicted categories of targets in all test images are compared, and a classification accuracy rate is obtained through calculation. The method is used for testing on a self-established data set CFOOD90, and theeffectiveness and the real-time performance of the method are verified.

The invention discloses a neural network based analyzing method for recognizing emotional tendency of text comments, and belongs to the technical field of computer language and work processing. The method is characterized in that the text comment data are processed through CBOW; each sentence is accurately divided into words or terms, and each sentence is provided with a corresponding class label; the emotional tendency of the comments is determined through a long-and-short term memory LSTM model; the label of each sentence is obtained and compared with a true label to obtain the accuracy rate; and a neural network model is trained to obtain the optimal accuracy rate, thus achieving the purpose of recognizing the emotional tendency of the text comments based on the neural network. According to the method, the GPU is utilized to accelerate the neural network training process, so that the emotion classing accuracy rate is increased, and moreover, the mass corpus data training speed is increased; the emotional tendency of the comments can be effectively recognized; and the method particularly has a good application prospect on e-commerce, film and other fields.

The invention discloses a traffic flow rate prediction method based on road clustering and two-layer two-way LSTM (long short-term memory). The method comprises the following steps of (1) providing a mode of using peripheral equalization on a loss value when the training data has a missing value, so that the missing data is filled; the prediction precision is improved; (2) providing a method of performing relevance clustering on the road according to the historical flow rate data; dividing the road into a plurality of groups; simultaneously utilizing the time information and the space information in the data preprocessing stage for improving the prediction precision; (3) designing a two-layer two-way LSTM deep neural network model for improving the prediction precision of the model; (4) providing a method of performing mass training and test on the network model; accelerating the training and test speed of the neural network model; (5) providing a multi-model fusion method for improving the prediction precision. The method provided by the invention has the advantages that the prediction speed and the prediction precision of the deep neural network model in an aspect of traffic flow rate prediction are accelerated and improved at the same time.

The invention discloses a method and device for using a neural network to correct white balance of images, and belongs to the technical field of digital image processing. The method comprises the steps of 1, collecting an original image; 2, calculating parameter values of color temperature of an ambient light source; 3, preprocessing the original image and an image to be processed to obtain a preprocessed image and a second preprocessed image which are used for training respectively; 4, constructing a convolutional neural network model; 5, training the convolutional neural network model; 6, utilizing the convolutional neural network to calculate and obtain a red channel gain value (gainR) and a blue channel gain value (gainB); 7, utilizing the red channel gain value (gainR) and the blue channel gain value (gainB) to conduct white balance correction on the image to be processed, and obtaining a white balance image after correction. The method solves the technical problem that white balance easily loses efficiency in the prior art, the calculation speed of an algorithm is effectively increased, the accuracy is greatly improved, and the model is very good in robustness.

The invention discloses a neural network online learning system based on a memristor. A pulse coding mode of a K-bit input vector is improved; the coding pulse corresponding to each bit is expanded into 2m pulses; In this way, the total number of needed coded pulses is K * 2m, each bit of weighted summation calculation is actually carried out for 2m times, finally, summation averaging operation iscarried out at the output end, the influence of accidental factors and noise on the calculation result in the calculation process is reduced through the mode, and therefore the calculation precisionis improved. The memristor array is simultaneously used for forward weighted summation calculation and weight small storage in the neural network; Different from offline learning, The weight in the memristor array needs to be updated once every time a signal is input through online learning, the variable quantity of the weight is mapped into the number of pulses, then the pulses are applied for one-time weight write-in operation, the neural network training speed can be increased, the hardware cost can be reduced, and the power consumption of neural network training is reduced.

The invention discloses a remote sensing image rapid target detection method based on a deep Hash auxiliary network, and solves the problem that an existing target detection framework cannot carry outefficient detection on a background area which does not have a target and occupies a large area in a wide remote sensing image. The method comprises the following implementation steps: generating anexperimental training image block set and a test image block set; constructing a remote sensing image rapid target detection model based on a deep Hash auxiliary network; training a remote sensing image rapid target detection model based on a deep Hash auxiliary network; and using the trained deep Hash auxiliary network to obtain a test result of the test image block set, and counting the runningtime. According to the method, the Hash auxiliary branch network module is introduced, redundant calculation in target detection is greatly reduced by quickly filtering most non-target background areas in the wide remote sensing image, the detection time of the non-target background areas is greatly shortened, and the detection efficiency is greatly improved. The method is used for target detection of wide remote sensing images.

The invention relates to a feature selection identification method for non-invasive power load monitoring, which is characterized in that the method comprises the following steps: step 1, detecting aload start-stop event from a total active power of a power consumption scene and extracting a load event feature; step 2, establishing a training data set of a machine learning classify; step 3, carrying out feature selection processing on that train data set, and optimally configuring the features used for the train classification model; step 4, establishing a supervised classification model, andtraining that supervised classification model with the train data set in the step 2 and the feature selected in the step 3; step 5, obtaining that value of each characteristic of the unknown kind ofload event through the step 1, analyzing the machine learn classification model established in the step 4, obtaining the classification result, and finally realizing the detection and identification of the load event. The invention can improve the identification performance of the machine learning classification model established for the identification load event through the feature selection.

The invention discloses a method and model for evaluating a network e-commerce loan risk. The method comprises the following steps of collecting historical transaction data of a network e-commerce loan customer as a sample set; building a least square twin support vector machine classification model, training the least square twin support vector machine classification model and building a network e-commerce load credit model; predicting default labels of various samples in the sample set according to a built network e-commerce load risk model; comparing a predicted default label with an actual default label for each sample, determining a prediction error of the sample, determining the magnitude of the weight of each sample according to the prediction error; and building the least square twin support vector machine classification model on the basis of the weight of each sample, rebuilding the network e-commerce load risk model and evaluating the load risk. The evaluation accuracy is high.

The invention discloses a citrus recognition method based on improved YOLOv4. According to the method, a YOLOv4 network model structure is improved, an up-sampling module and a detection feature map sensitive to a small target are added, and citruses with relatively small individuals can be better identified; sparse training, channel pruning and layer pruning are carried out on a network model obtained through training, the defects of large memory consumption, long recognition time and the like caused by module addition are overcome, clustering is carried out by using a Canopy algorithm and ak-means + + algorithm, and a user can obtain an anchor frame parameter value more suitable for a data set of the user. When citrus recognition is carried out, an improved YOLOv4 network structure is adopted to train a citrus data set, and the obtained model can recognize a target with a small individual more accurately; before a network model is trained, through combination of layer pruning and channel pruning, the depth and the width of the model are compressed, and the training speed is improved on the premise that the precision is not lost; citrus on trees in different periods is recognized, the recognition precision is high, the speed is high, and the requirement for real-time recognition can be met.

The invention discloses a feature pyramid-based remote-sensing image time-sensitive target recognition system and method. The system comprises a target feature extraction sub-network, a feature layersub-network, a candidate area generation sub-network and a classification and regression sub-network, wherein the target feature extraction sub-network is used for carrying out multiple layers of convolution processing on a to-be-processed image and outputting the convolution processing result of each layer as a feature layer; the feature layer sub-network is used for overlapped the last feature layer and the current feature layer to obtain the current fused feature layer, wherein the topmost fused feature layer is a topmost feature layer; the candidate area generation sub-network is used forextracting candidate areas from different layers of fused feature layers; and the classification and regression sub-network is used for mapping the candidate areas to different layers of fused featurelayers so as to obtain a plurality of mapped fused feature layers, and carrying out target judgement on the plurality of mapped fused feature layers so as to output a result. According to the systemand method, the hierarchical structures of feature pyramids are utilized to ensure that all the scales of features have rich semantic information.