76 results about "Batch training" patented technology

The invention discloses a supervised snore source identifying method. The method includes data pretreatment, training and identification; the main steps include, firstly, performing Mel frequency conversion on actual measured snore data, and acquiring a data sample; secondly, arranging a structure of a convolution neutral network, quantity of output feature drawings of a convolution layer and convolution core size, pooling size, weight vector updating studying rate, batch training sample number, and training iterations; thirdly, using a snore time frequency spectrogram of a training set as the convolution neutral network to input; performing network initialization on the arranged network structure; completing the training process through forward process, direction bias transmission, weight value updating and offset until the appointed iterations; at last, delivering the testing set to the trained network model and acquiring the identifying result. The supervised snore source identifying method can effectively identify the snore source, and is exact in identifying result and good in performance.

The invention discloses a machine learning-based system and a learning method. The system includes a data management module, a data conversion module, an algorithm training module, a result evaluationmodule and a model prediction module. According to the learning method, the data management module automatically determines the data type of a data source uploaded by a user and preprocesses the data; the user labels the data; the data conversion module substitutes the preprocessed data into a feature conversion algorithm set to obtain feature-converted data; the algorithm training module automatically substitutes the feature-converted data into a classification, clustering, or regression algorithm for batch training so as to obtain a batch of classification, clustering, or regression models;the result evaluation module evaluates the standard indicators of the above models so as to obtain an optimal model; and the model prediction module predicts the latest data source with the optimal model. The system can automatically perform batched data processing and algorithm training, manual intervention in an intermediate process is not required, and therefore, the threshold of learning is decreased, and efficiency can be improved.

The invention provides a method for link prediction of a complex network, an end-to-end link prediction model based on a graph attention network (GAT), and a batch training method of the model. The key of this model is to learn the attention distribution of network nodes to neighbors. The steps of model training and model prediction include: (1) inputting the topological structure of unweighted undirected homogeneous network; 2, sampling all nodes according to the topological structure of the training set, so as to batch the network; 3, inputting the batched training set into the model to train the model parameters; Step 4, inputting the point pair to be predicted, and the model outputs the probability of connecting edges between the point pairs. The model of the invention has the characteristics of end-to-end. The batch training method makes the model suitable for large-scale complex networks.

The invention provides an image generation method and system based on a capsule network. The method comprises the following steps: firstly, constructing training data according to attribute characteristics of an image to be generated; Constructing an image generation model; Training the constructed image generation model by adopting a batch training method according to the training data to obtaina trained image generation model, and constructing a random noise vector; And finally, taking the attribute vector and the random noise vector of the to-be-generated image as the input of the trainedimage generation model, and generating a new image with the same size as the image in the training data. The method has the beneficial effects that the capsule network is added into the image generation model to serve as the encoder network, so that the faster convergence of the model training process is facilitated; And on the other hand, compared with the pooling process of the convolutional neural network, the dynamic routing algorithm of the capsule network has stronger generalization robustness on characteristics, and more diversified and real images can be generated.

The invention specifically discloses an electric power metering system network intrusion detection method based on a federated learning framework. The method comprises the following steps: S1, enabling a concentrator to obtain a local model and setting a target; s2, acquiring user electric meter data and sending the user electric meter data to a concentrator to construct a plurality of batches of training samples; s3, enabling each concentrator to download a global model weight parameter from the data center and carrying out training on the global model weight parameter and the selected batch of training samples to obtain a corresponding local model weight parameter; s4, transmitting the local model weight parameters to a data center for aggregation averaging to obtain a next round of global model parameters, feeding back the obtained next round of global model weight parameters to the concentrator, and optimizing respective local model weight parameters; s5, repeating the step S4 to obtain a detection model; and S6, judging whether network intrusion exists in the real-time operation system or not according to the obtained detection model. According to the method, the privacy of the user can be protected, the detection rate can be ensured, and the communication time and computing resources are reduced.

The invention relates to the technical field of deep learning, in particular to a small sample generation method and device based on a generative adversarial network, aiming at solving the technical problem of how to generate sample data by using the generative adversarial network under the condition of a small amount of sample data. For this purpose, the small sample generation method based on the generative adversarial network provided by the present invention can generate samples corresponding to small sample types based on the generative adversarial network and according to random noise and label information. In this process, the invention adopts the methods of migration learning and batch training to train the generative adversarial network, so that the generative adversarial networkcan be effectively migrated and applied to the generative adversarial network sample generation task of a small number of samples.

The invention belongs to the technical field of applying an electronic device to carry out identification and discloses a verifiable privacy protection single-layer perceptron batch training method and a mode identification system. On the basis of the same group of training samples, the training can be carried out for different modes, thereby obtaining a plurality of different training models. Ineach round of iteration, a small batch of samples instead of a sample are selected for iteration. Through utilization of a stochastic gradient descent method, a sample vector is expanded into a matrixsample for acceleration of a convergence rate of the iteration. In a training phase, a user outsources heavy computing tasks to a cloud server. The user needs to carry out encryption operation beforeuploading an input matrix. Through utilization of a stochastic permutation function and a sparse matrix blind technology, the privacy protection of user data is realized. According to the method, a verification mechanism is taken into consideration in a single-layer perceptron training scheme for the first time, the cloud server returns a wrong computing result, and the user can check the wrong computing result with a probability of 100%.

The invention discloses a multi-mechanism mixed recurrent neural network model compression method. The multi-mechanism mixed recurrent neural network model compression method comprises A, carrying outcirculant matrix restriction: restricting a part of parameter matrixes in the recurrent neural network into circulant matrixes, and updating a backward gradient propagation algorithm to enable the network to carry out batch training of the circulant matrixes, B, carrying out forward activation function approximation: replacing a non-linear activation function with a hardware-friendly linear function during the forward operation process, and keeping the backward gradient updating process unchanged; C, carrying out hybrid quantization: employing different quantification mechanisms for differentparameters according to the error tolerance difference between different parameters in the recurrent neural network; and D, employing a secondary training mechanism: dividing network model training into two phases including initial training and repeated training. Each phase places particular emphasis on a different model compression method, interaction between different model compression methodsis well avoided, and precision loss brought by the model compression method is reduced to the maximum extent. According to the invention, a plurality of model compression mechanisms are employed to compress the recurrent neural network model, model parameters can be greatly reduced, and the multi-mechanism mixed recurrent neural network model compression method is suitable for a memory-limited andlow-delay embedded system needing to use the recurrent neural network, and has good innovativeness and a good application prospect.

The invention provides an air quality prediction method based on a seasonal recurrent neural network. The method comprises the following steps: recording long-time-sequence air quality monitoring station data; preprocessing monitoring station data, analyzing seasonal and periodic change characteristics of air quality long-time sequence monitoring data, and determining seasonal change frequency of the air quality long-time sequence monitoring data; carrying out time sequence decomposition according to seasonal and periodic time sequence change characteristics of the air quality monitoring parameters, and generating a trend component, a seasonal component and a residual component; constructing a large batch of training samples, preliminarily establishing a three-channel recurrent neural network, and obtaining an optimal three-channel recurrent network model through training; predicting the future value of each component according to the historical value by using the optimal three-channel cyclic network model to obtain a trend component predicted value, a seasonal component predicted value and a residual component predicted value; and integrating the prediction results of the components to form a final air quality prediction value sequence.

The invention discloses an OFDM signal detection method based on an RNN neural network, which can be applied to a cognitive radio spectrum sensing technology of wireless mobile communication, and effectively improves the recovery capability of an OFDM system in the aspect of receiver signal demodulation, thereby improving the signal detection performance of cognitive radio spectrum sensing. On thebasis of analyzing an OFDM system, firstly, an OFDM system framework is built, and a time sequence data set for model learning is generated by utilizing the OFDM system framework; then, the time sequence signal data set is learned through an LSTM model of the RNN neural network; an Adam training algorithm is adopted to quickly optimize parameters and a small-batch training mode, so that end-to-end spectrum signal detection is realized, the problem of signal receiving of an OFDM system with NBI and ICI is effectively solved, and the spectrum detection performance of the system is improved.

The invention provides a CNN based rain measurement identification of a marine radar. The method comprises a model building process, a model optimization process and a model training process. The method is based on a classical LeNet-5 CNN model. A multi-level residual CNN model is established, a training set is established by using samples of light rain, medium rain and heavy rain, and input intothe multi-level residual CNN to train the network model, the loss rate is calculated by using a cross entropy loss function, the minimum loss is obtained rapidly, and achieving a very good convergencestable state by using a batch training mode.

The embodiment of the invention provides a model training method which is applied to the field of artificial intelligence, and the method comprises the steps: obtaining a first neural network model and M batches of batch training samples, M being a positive integer greater than 1; then determining a target incremental training method according to sample distribution characteristics between batches of batch training samples in the M batches of batch training samples, wherein the sample distribution characteristics are related to the degree of catastrophic forgetting generated by the model when incremental training is carried out based on the batches of batch training samples; and using the target incremental training method for realizing catastrophic forgetting resistance when incremental training is performed on the model; and according to the M batches of batch training samples, performing self-supervised training on the first neural network model through a target incremental training method to obtain a second neural network model. According to the method, on the premise that the training time is shortened and the data storage space is saved, the balance between efficiency and performance is realized.

The invention relates to a GAN-based carrier image synthesis steganography method, which comprises the steps of cutting each image in a data set into images with the same size, and forming a real image data set; constructing a generation network G, a discrimination network D, an evidence obtaining network F and an extraction network E, initializing parameters in the generation network G, the discrimination network D and the extraction network E, and setting the parameters in the evidence obtaining network F as preset values; training the initialized generation network G, discrimination networkD and extraction network E by using a batch training mode to obtain a trained generation network G, discrimination network D and extraction network E; inputting secret information to be embedded anda preset secret key into the trained generation network G to obtain a synthetic carrier image; and inputting the synthesized carrier image and the preset secret key into the trained extraction networkE, and extracting secret information. According to the method, the security of the algorithm is enhanced, and the naturalness of the synthesized carrier image is improved.

The invention discloses a data receiving method based on adaptive moment estimation, which comprises the following steps of: eliminating intersymbol interference through a combined method consisting of a time domain feedforward equalizer, a rear filter and maximum likelihood sequence detection improved on the basis of the adaptive moment estimation method; the process specifically comprises the steps that a training sample matrix is constructed through a small number of training sequences, iterative batch training based on a self-adaptive moment estimation algorithm is conducted on time domainequalizer tap coefficients through the matrix, ideal equalizer tap coefficients are obtained, post-filtering and maximum likelihood detection are conducted on equalized signals, and signals with interference eliminated are obtained; an efficient adaptive moment estimation iteration batch equalization training method is adopted to replace a traditional time domain feedforward equalizer one-by-onesample training method, the interference elimination performance is high, and meanwhile the proportion of effective loads is increased.

The invention discloses a target detection model training method and device, electronic equipment and a storage medium, and relates to the technical field of deep learning and computer vision. The specific implementation scheme is as follows: firstly, acquiring a plurality of sample sets of a plurality of targets in a preset scene; determining the sample sampling number of a single target in batchtraining according to the plurality of sample sets of the plurality of targets; then, in each batch of training, performing sampling from the sample set of each target according to the sample sampling number, and forming batch training data; and training the target detection model by using the batch training data until the training of the target detection model is completed. According to the method, when the training data is sampled, each target is sampled according to the sample sampling number, and the training data with the same number or approximately the same number can be obtained for training, so that the detection efficiency and the detection accuracy of the trained target detection model on each target are ensured.

The invention discloses a model parameter updating method and device, and relates to the technical field of deep learning. According to the specific implementation scheme, the method comprises the steps of obtaining batch training time of a model for batch training data to be trained; when the batch training time is greater than or equal to a preset time threshold, improving the priority of the tasks sorted later in the parameter gradient communication task sequence; carrying out parameter gradient communication and parameter updating according to the priority of each parameter gradient communication task in the model, so that the priority of each parameter gradient communication task can be adjusted by combining batch training time, the influence of other training tasks sharing a communication link is avoided, the training time is reduced, and the training speed is increased.