498 results about "Overfitting" patented technology

The present invention relates to the field of deep learning techniques, in particular to a deep convolution neural network training method and a device. The deep convolution neural network training method and the device comprise the steps of a, pretraining the DCNN on a large scale data set, and pruning the DCNN; b, performing the migration learning on the pruned DCNN; c, performing the model compression and the pruning on the migrated DCNN with the small-scale target data set, In the process of migrating learning of large-scale source data set to small-scale target data set, the model compression and the pruning are performed on the DCNN by the migration learning method and the advantages of model compression technology, so as to improve the migration learning ability to reduce the risk of overfitting and the deployment difficulty on the small-scale target data set and improve the prediction ability of the model on the target data set.

A training method and a training apparatus for a neutral network for image recognition are provided. The method includes: representing a sample image as a point set in a high-dimensional space, a size of the high-dimensional space being a size of space domain of the sample image multiplied by a size of intensity domain of the sample image; generating a first random perturbation matrix having a same size as the high-dimensional space; smoothing the first random perturbation matrix; perturbing the point set in the high-dimensional space using the smoothed first random perturbation matrix to obtain a perturbed point set; and training the neutral network using the perturbed point set as a new sample. With the training method and the training apparatus, classification performance of a conventional convolutional neural network is improved, thereby generating more training samples, reducing influence of overfitting, and enhancing generalization performance of the convolutional neural network.

The invention provides an object-oriented high-resolution remote-sensing image classification method. The method comprises the steps of S1, conducting segmentation processing on images to be processed to obtained a plurality of subimage objects; S2, obtaining feature information of subimage objects; and S3, classifying subimage objects according to the obtained feature information, wherein images to be processed are high-resolution remote-sensing images, the feature information of subimage objects comprises spectral information, shape information and texture information of subimage objects. According to the method, on the basis of object-oriented classification, a classification method combining probabilistic latent semantic analysis and a support vector machine is introduced, the problem that 'the same features with different classifications' and 'the same classifications with different features' are not high in identification ratio in the prior art is solved, the classification precision of high-resolution remote-sensing images is greatly improved, advantages of latent semantic analysis (LSA) and advantages of probabilistic latent semantic analysis (PLSA) are combined, and the problems of overfitting and local optimum which are caused by random initialization are effectively solved.

The invention discloses an adversarial attack defense method based on adversarial sample training for an adversarial attack defense problem of a deep learning network. The adversarial attack defense method includes the steps: firstly, prior distribution of training sample category labels is introduced, and the labels are subjected to smooth correction, so that the network does not excessively believe that the training samples are completely correct during training; secondly, adversarial samples are added in the network training process, meanwhile, an original loss function is modified, the contribution of the adversarial samples to the loss function is reflected, so that the network has resistance to gradient attacks; and finally, random inactivation processing is performed on the networkneurons, so that over-fitting is reduced to a certain extent. According to the adversarial attack defense method, overfitting is inhibited, and the generalization ability of the network is enhanced, so that the robustness of the network for defending against resistance sample attacks is improved.

The invention particularly relates to a sentiment dictionary learning based text sentiment analysis method and system. The method includes steps: subjecting initial text data to normalization processing to generate preprocessed text data, and clustering the preprocessed text data to a preset field; adopting a seed sentiment dictionary based sentiment word discovery method to form a special sentiment dictionary in the preset field; subjecting the preprocessed text data to retrieval according to the special sentiment dictionary to acquire target text data serving as initial training corpus in the corresponding preset field, and classifying inputted to-be-classified texts through formed multiple classifiers. Labor cost is reduced, the problem of overfitting caused by a single classifier is avoided, and accuracy in text sentiment analysis is improved by consideration of the text related fields.

The invention discloses an automatic question generation method based on deep learning. The method comprises the following steps: constructing a training set (articles, answers and questions), a verification set (articles, answers and questions) and a prediction set (articles and answers); creating Encoder-based construction by using deep learning framework tensorflow A sequence of decoders to a sequence neural network model; Carrying out word segmentation, word list making and word embedding operations on sentences in the data set; Wherein the data set comprises a training set, a verificationset and a prediction set; The training set is used for training the model, the verification set is used for detecting whether the currently trained model is over-fitted or not, and if over-fitting isachieved, stopping training; Otherwise, continuing training; And decoding the prediction set by using the trained model to generate a problem. The method is good in generalization effect and low in labor cost, the generated questions are better matched with articles and answers, and the method can be widely applied to the fields of intelligent teaching, intelligent questions and answers, knowledge question and answer games and the like.

The invention discloses a deep learning-based crack identification method. According to the method, correct identification of crack positions and crack attributes is realized at the same time by utilizing a deep convolutional neural network which comprises branches; on the trunk of the network, convolutional layers and de-convolutional layers are combined to realize end-to-end crack position prediction; and on the branches, attribute identification is realized. In order to overcome the difficulty that the labelling of crack samples wastes time and labor, a simulation crack drawing algorithm is designed to realize the automatic drawing and labeling of cracks, so that the workload of manual labeling is greatly lightened, training samples in large data volume are provided for deep learning, overfitting of deep network models is avoided, and the convergence and convergence efficiency during real crack training are improved; and according to the method, the identification correctness is greatly enhanced, the universality is stronger, the reliability is higher and demands of industrial detection can be satisfied.

The invention provides a speech recognition method adopting a rectified deep auto-encoder network. The method comprises the following steps: first, carrying out training on the deep auto-encoder network by adopting a rectification linear unit as an activation function which replaces the traditional Sigmoid function, then, introducing L2 regularization for optimizing the overfitting problem easily occurring during the deep model training process, and finally, carrying out the layer-by-layer greedy and unsupervised 'pretraining' and supervised 'fine tuning', thus obtaining weights with character representation for speech recognition. According to the method provided by the invention, the strong capacity of learning substantive characteristics of data sets from the minority of samples of the deep neural network is fully utilized, thus the problems of gradient disappear and overfitting during the training process are solved, and the recognition precision of the system under the noise environment is improved.

New computational technologies generating systematic investment portfolios by coordinating forecasting algorithms contributed by researchers are provided. Work on challenges is efficiently facilitated by the algorithmic developer's sandbox (“ADS”). Second, the algorithm selection system performs a batch of tests that selects the best developed algorithms, updates the list of open challenges and translates those scientific forecasts into financial predictions. The algorithm controls for the probability of backtest overfitting and selection bias, thus providing for a practical solution to a major flaw in computational research involving multiple testing. Third, the incubation system verifies the reliability of those selected algorithms. Fourth, the portfolio management system uses the selected algorithms to execute investment recommendations. A dynamically optimal portfolio trajectory is determined by a quantum computing solution to combinatorial optimization representation of the capital allocation problem. Fifth, the crowdsourcing of algorithmic investments controls the workflow and interfaces between all of the hereinabove introduced components.

The invention discloses an image classification method and system based on migration learning. The method comprises the steps that 1, a feature similarity known training set A is utilized to make a training set B of a migration network through a support vector machine; 2, a migration learning network is constructed; 3, the training set B classified in the step 1 is used as a training learning setof the migration learning network, and a migration learning network model with high robustness and good accuracy is obtained through training; and 4, a to-be-classified dataset is introduced into thetrained migration learning network model, and a final classification result is obtained and marked with a tag. Through the image classification method and system, the requirement that a big sample dataset is needed to serve as input when a common RGB image is trained through deep learning is overcome, the problems of overfitting and locally optimal solutions in the training process are avoided, and classification precision is improved to some extent compared with a traditional classification algorithm.

The invention discloses a transformer fault diagnosis method based on an improved cuckoo search optimal neural network. According to the method, first, the concentrations of DGA characteristic gases are collected and subjected to normalization processing; the number of neurons in an implicit layer of a BP neural network, a training function and a transfer function from an input layer to an outputlayer are determined, and a fault diagnosis model based on the BP neural network is established; an improved cuckoo search algorithm is adopted to optimize parameters of the BP neural network, an optimal weight threshold parameter is obtained, and an optimal BP neural network model is obtained; training samples are utilized to train the optimal BP neural network model, and an improved cuckoo search neural network diagnosis model is obtained; and the improved cuckoo search neural network diagnosis model is adopted to predict test samples, and the output of the model is a transformer fault diagnosis result. Through the method, the problems that the existing BP neural network is slow in overfitting and convergence speed and a solution in a CS algorithm is poor in quality and low in diagnosisprecision are solved.

The present invention discloses a deep convolutional network-based airborne ground penetrating radar target identification method, relates to the machine learning and ground penetrating radar application technologies, in particular to the application of a deep learning method in the airborne ground penetrating radar target identification. The method comprises the following steps of acquiring and pre-processing the radar data, designing the multiple layers of structures of a neural network, selecting a hyper-parameter, preventing the overfitting, activating a function, training a convolutionalmodel and displaying a prediction result. The airborne ground penetrating radar target identification method of the present invention identifies an airborne ground penetrating radar target, can automatically extract the parameters of the updated network during the training process, and reduces the manual intervention during the processing process. Meanwhile, the convolutional model of the presentinvention can extract the two dimensional filter characteristics of the different levels of the target, and the characteristics can represent the characteristics, such as the target, the background, the interference, etc. The deep convolutional network-based airborne ground penetrating radar target identification method enables the accuracy of the airborne ground penetrating radar target signal identification to be improved.

The invention discloses a centrifugal pump fault diagnosis method based on complete ensemble empirical mode decomposition and random forest. The method comprises the following steps: (1) decomposing a centrifugal pump vibrating signal obtained by a sensor into a series of IMF categories by utilizing CEEMD; (2) taking the sample entropy of the IMF categories as characteristic vector of the signal; and (3) carrying out fault diagnosis by taking the characteristic vector obtained by the CEEMD-sample entropy as input of a random forest classifier. According to the invention, the CEEMD and the sample entropy are used for characteristic extraction of the centrifugal pump vibrating signal, on one hand, the phenomena of modal aliasing and end effect occurring in EMD decomposition are avoided as far as possible; and on the other hand, characteristic extraction is relatively convenient and simple, calculated amount is small, and the characteristic extraction is not sensitive to data length and noise, thus being high in applicability. According to the invention, the random forest classifier is used for fault mode identification of the centrifugal pump, thus avoiding the phenomenon of overfitting caused by the fact that conventional classifiers depend too much on training samples, and improving the classification accuracy as far as possible.

The present invention provides a target tracking method based on a long short-term memory network, and relates to the computer vision technology. The method comprises the steps of: performing pre-estimation of a candidate target state by employing a fast matching method based on similarity learning, screening out high-quality candidate target states, and performing classification of the high-quality candidate target states by employing a long short-term memory network. The long short-term memory network comprises a convolutional layer used for feature extraction and a long short-term memory layer used for classification. The convolutional layer is obtained through offline training on a large-scale image data set ILSVRC15 to avoid a risk of overfitting of the target tracking data set. The long short-term memory layer is obtained through online learning and fully employs the time correlation included by an input video sequence so as to have good capacities on adaption of target forms andmotion change. The speed is observably improved, the long short-term memory network capable of adapting target change is utilized to target tracking.

The invention provides an abnormal behavior detection method and an abnormal behavior detection device based fused characteristics. The method comprises steps that, according to a detection tracking processing result of a motion target in a to-be-detected video, a behavior type of the motion target is determined; multi-dimensional characteristics of the motion target are extracted, including a pixel point change degree, a pixel point arrangement tightness degree, an integral shape, a frame image similarity degree, motion characteristics, position characteristics and form characteristics; the multi-dimensional characteristics are analyzed and processed according to a characteristic fusion model corresponding to the behavior type, whether the motion target has abnormal behaviors is determined according to the processing result; according to innovative characteristics of the multiple abnormal behaviors, algorithm robustness and stability can be effectively improved, according to the characteristic fusion model acquired through learning and training large-scale abnormal behaviors, the multi-dimensional characteristics are analyzed and processed, problems of algorithm overfitting and insufficient fitting can be effectively avoided, the method is suitable for multiple types of complex application scenes, time cost and manpower cost are greatly saved, and the method has high popularization values.

Provided is a gait recognition method based on a similar rule Gaussian kernel function classifier. The method comprises the steps that a camera collects a current background image and an original gait image sequence of a detection target in real time, image preprocessing is carried out by the adoption of an Euclidean distance method and the like, and a standard gait image sequence is obtained; one gait sequence is divided into three gait subsequences by application of an interval frame grabbing technology, feature extraction is carried out, and gait feature vectors are obtained; similar rule construction is carried out by utilization of the feature vectors in a gait feature vector database; the gait feature vectors of the detection target are classified through the Gaussian kernel function classifier corresponding to the similar rule construction, and a recognition result is counted and output. The method can rapidly remove the background, and adaptability under different situations is improved by application of the image normalization and the interval frame grabbing technology. In addition, the similar rule Gaussian kernel function classifier can effectively avoid the problems of overfitting, dimension disasters and the like, and improve the integral recognition precision.

The invention relates to an improved extreme learning machine, particularly relates to an improved extreme learning machine combining learning thought of a least square vector machine, and belongs to the technical field of the artificial intelligence. On the basis of minimizing empirical risk of the traditional extreme learning machine, the improved extreme learning machine provided by the invention is combined with the learning thought of the least square vector machine, and is additionally provided with a structural risk control item, and calculates the result through efficiently adjusting the ratio of two kinds of risks, so that the risk of the overfitting generated by a module is greatly lowered. Three experiments of practically applying the method in the Sinc data set, the Boston Housing data set and the oxygen dissolving forecasting in the aquaculture indicate that compared with the ELM algorithm and the EOS-ELM algorithm, the forecast error of the method is relatively approximate with the training error, so that the problem of the overfitting is efficiently lowered, and the forecast precision is enhanced to a certain extent.

The invention discloses a neural network photovoltaic power generation prediction method and system suitable for small samples. The method comprises the steps that firstly, historical (small sample) photovoltaic power generation power data and meteorological data are acquired; Establishing a BP neural network photovoltaic power generation prediction model according to factors influencing photovoltaic power generation; In order to solve the problem of overfitting of a prediction model caused by too few training data, a Dropout strategy is adopted to optimize a neural network, and in order to solve the problem that a BP neural network is prone to being caught in a minimum value, a genetic algorithm is adopted to optimize the neural network. The sample data is divided into the training data set and the test data set, the training data is used for training the neural network photovoltaic power generation prediction model, the test data set is used for testing the network, and the generalization ability of the neural network photovoltaic power generation prediction model is improved. By adopting the method provided by the invention, the problem of low precision caused by over-fitting ofthe neural network photovoltaic power generation prediction model under the condition of small sample historical data can be effectively solved.

The invention relates to a lightweight remote sensing target detection method based on SEYOLOv3, which belongs to the technical field of target detection, and comprises the following steps: 1, takinga YOLOv3 algorithm as a basic model framework, and in order to reduce network parameters and improve network reasoning speed, designing a lightweight trunk feature extraction network; 2, in order to improve the scale invariance of the features and reduce the over-fitting risk, a spatial pyramid pooling (SPP) algorithm is provided, and pooling of three scales is carried out to obtain an output feature vector with a fixed length; a spatial attention model SE module is introduced, useless information is further compressed, and useful information is enhanced; and 3, updating parameters through iterative training to obtain a final network model, adopting multi-scale prediction by utilizing the model, and predicting a final result through detection heads of three scales. According to the method,while the reasoning speed of the network is effectively improved, the precision is ensured, the feature expression capability of the network is enhanced, and the scale invariance is improved.

The invention relates to a DCNN (Deep Convolutional Neural Network)-DNN (Deep Neural Network) and PV-SVM (Paragraph Vector-Support Vector Machine)-based multi-modal depressive disorder estimation and classification method. The method comprises the following steps: preprocessing audio and video features through a displacement range histogram and an Opensmile tool, extracting hidden layer abstract features of audio and video statistical features through a DCNN, performing depressive disorder estimation through a DNN, performing high-dimensional feature mapping on textile information through a PV method, inputting an obtained high-dimensional feature expression into an SVM for binary classification, connecting a depressive disorder estimation result with a binary classification result in series, inputting the whole into a random forests model for training, and performing a final depressive disorder classification task through the trained random forests model, namely judging a depressive disorder or a non depressive disorder. By the adoption of a DCNN model for extraction of the hidden layer abstract features from a primary audio / video, an original high-dimensional feature is more compact, and included information is richer; therefore, the model is more effective, and the phenomenon of overfitting caused by extremely high dimension of the feature is avoided.

The invention provides a flight delay model establishment method and device and a prediction method and device, and relates to the technical field of machine learning, and the method comprises the specific steps: obtaining historical data, and carrying out the data preprocessing of the historical data; constructing input characteristics of a flight delay model for the historical data after data preprocessing; wherein the input characteristics comprise non-weather characteristics of the flight and weather characteristics of the flight; performing random stratified sampling on the historical data after data preprocessing to obtain a training set, a first test set and a second test set; determining a target weather feature from the weather features of the flight based on a preset algorithm; establishing a corresponding basic model according to the non-weather characteristics of the flights of the training set and the at least one target weather characteristic; over-fitting of the basic model on the training set is prevented through expression of the basic model on the first test set; and combining the plurality of basic models into a flight delay model through the second test set. Overfitting can be prevented, and an ideal prediction effect is obtained.

The invention discloses an equipment residual life prediction method and system. The method comprises the following steps: establishing an equipment degradation mathematical model based on a nonlineardiffusion process; obtaining an estimated value of the equipment degradation parameter under the acceleration stress as first data; calculating a degradation parameter value under normal working stress according to the first data to obtain second data; determining the distribution type of the second data through goodness-of-fit test; obtaining a posteriori distribution function and an expected value of the second data according to the distribution type of the second data; obtaining a first residual life probability density function according to the posteriori distribution function, the expected value and the equipment degradation mathematical model; correcting the first residual life probability density function according to a total probability formula to obtain a second residual life probability density function; and predicting the residual life of the equipment according to the second residual life probability density function. The prediction precision of the residual life of the equipment can be improved.

The invention discloses a pedestrian re-identification method based on a random erasure pedestrian alignment network, and the method comprises the following steps: 1, carrying out the preprocessing ofan original image in a data set through employing a random erasure data enhancement algorithm, increasing the shielding, and reducing the influence caused by the overfitting; Step 2, sending the preprocessed image into a basic branch of the network; Step 3, enabling the images with the image dislocation phenomenon to enter an affine estimation branch to be aligned; 4, the aligned image enters analignment branch to be repositioned and subjected to feature extraction; 5, pedestrian re-identification is carried out in the three large pedestrian re-identification databases, and values of Rank-1and Map are used for representing effects of the re-identification test; According to the method, the image dislocation phenomenon is well improved, the overfitting influence is reduced, and the pedestrian re-identification rate can be well improved through the trained network.

The invention discloses a wavelet transform and particle swarm optimized grey model-based short-term wind speed forecasting method. The method comprises the following steps: decomposing and analyzing wind speed data by utilizing wavelet decomposition and reconstruction; inputting the reconstructed data into a particle swarm optimized grey wind speed forecasting model one by one to obtain forecasting values; adding the forecasting values to obtain a wind speed forecasting value; and finally evaluating the wind speed forecasting ability. According to the method provided by the invention, the problem of high-frequency component overfitting is solved on the basis of ensuring the low-frequency component fitting; and less information is required and the forecasting can be completed under the condition of being relatively lack of data, so that the precision and stability of the method is higher than that of the traditional grey model forecasting method.

The invention discloses a residual network and attention mechanism-based drug relationship extraction method. The method comprises the following steps of: S1, carrying out vector representation on words in a drug entity relationship data set; S2, carrying out time-series modeling on a drug relationship statement by utilizing a two-layer bidirectional long-short term memory model neural network; S3, importing residual connection into the constructed two-layer bidirectional long-short term memory model neural network; S4, decomposing a deep semantic meaning automatically obtained by the two-layer bidirectional long-short term memory model neural network into a memory space and an attention space, fusing memory information and attention information, and inputting the fused information into aSoftmax classifier to extract a drug relationship. According to the drug relationship extraction method disclosed by the invention, dependency relationships between long-distance words are effectivelysolved, gradient diffusion is overcome, model overfitting is prevented, the model robustness is good and the classification effect is good.

The invention discloses a ship target identification method based on transfer learning. Firstly, a data set is generated through a data enhancement method; optimizing and dividing the data set based on the number of positive and negative samples to obtain a training set and a test set; In order to avoid overfitting caused by a small sample data set, adopting an Inception V3 model to train and testthe data set in a transfer learning mode; taking the quantitative evaluation index as a quantitative evaluation index, selecting a video frame shot in a real scene to carry out a comparison test, obtaining a series of coordinate points by setting a threshold value, drawing a qualitative evaluation index ROC curve, and generating classification information as a qualitative evaluation index. According to the method, high-precision classification and high-operation-efficiency identification of the ship target domain can be realized without manually extracting features, and the method has a higher classification recall rate and a higher operation speed.

The invention relates to a fine-grained image classification method based on a sparse bilinear convolutional neural network, and the method comprises the steps: carrying out the feature channel cutting of the bilinear convolutional neural network, automatically thinning feature channels in a training process, distinguishing the importance of the feature channels for classification, and carrying out the scale cutting according to the importance; inputting the output of the bilinear convolutional neural network into the batch regularization; taking a scaling factor of BN as a scale factor; applying a regularization method to the scale factor, wherein the regularization method has a plurality of types such as L1 and L2, the sparsity of L1 is strong, the sparsity of the feature channels can be realized by jointly training the network weight and the scale factor; finally, performing pruning according to the size sequence of the sparse scale factor, and finally, obtaining a model for finally performing a fine-grained image classification task by utilizing fine tuning. Weak supervision can be realized, redundant parameters are reduced, overfitting is prevented, and the accuracy of fine-grained image classification is effectively improved.

The invention discloses a method of application classification in Tor anonymous communication flow, which mainly solves the problem of acquisition of upper-layer application type information in the Tor anonymous communication flow and relates to the correlation technique, such as feature selection, sampling preprocessing and flow modeling. The method comprises the following steps of: firstly, defining a concept of a flow burst section by utilizing a data packet scheduling mechanism of Tor, and serving a volume value and a direction of the flow burst section as classification features; secondly, preprocessing a data sample based on a K-means clustering algorithm and a multiple sequence alignment algorithm, and solving the problems of over-fitting and inconsistent length of the data sample through the manners of value symbolization and gap insertion; and lastly, respectively modeling uplink Tor anonymous communication flow and downlink Tor anonymous communication flow of different applications by utilizing a Profile hidden Markov model, providing a heuristic algorithm to establish the Profile hidden Markov model quickly, during specific classification, substituting features of network flow to be classified into the Profile hidden Markov models of different applications, respectively figuring up probabilities corresponding to an uplink flow model and a downlink flow model, and deciding the upper-layer application type included by the Tor anonymous communication flow to be classified through a maximum joint probability value.

The invention relates to a PCB labeling printing quality detection method based on deep learning. The method comprises the following steps of carrying out a series of pretreatment such as Gaussian filtering, conversion into a gray scale image, edge detection, hole filling and optimal threshold binarization segmentation on an obtained PCB original image so as to effectively improve recognition detection efficiency; adopting an Otsu algorithm for carrying out optimal threshold segmentation on the image to obtain a binary image, so that the test precision of the method is improved; effectively collecting and expanding image sample data to increase representativeness of samples. The model overfitting phenomenon caused by few data is prevented. A model obtained through training is affected by irrelevant factors as little as possible. Robustness of the model is enhanced, training time of a deep learning network is shorter, and the convergence speed is higher. Target features of the image areextracted efficiently through the deep neural network, and the defects of easy overfitting, long training time, high parameter adjustment difficulty and the like caused by a traditional artificial neural network feature extraction method can be effectively avoided.