42results about How to "Reduce the amount of training data" patented technology

The invention belongs to the technical field of computer vision and artificial intelligence, and particularly relates to a large-scale face recognition method based on a depth convolution neural network model. The method comprises steps of putting forward a residual error learning depth network model facing large-scale face recognition, wherein the residual error learning depth network model is formed by a convolution layer, a residual error layer and a full connection layer, and the residual error layer is formed by adding one path of multiple convolution layer cascade data and one path of original data to calculate the sum; and carrying out normalization operation in batch after each convolution layer in the model. According to the invention, by use of the characteristics of strong learning ability and good residual error learning convergence of the depth convolution neural network, layers of the model are increased in the aspect of the layer number of the network model; and in the aspect of residual error layer structure, the invention provides a highly efficient residual error layer structure. In the field facing the large-scale face recognition, the accuracy of the provided method is greatly improved compared with a base line model, and the accuracy of face retrieval in a million-class face recognition database can reach 74.25%.

The invention provides a power battery system residual available life prediction method based on a relevance vector machine and particle filter; the method comprises the following steps: using the relevance vector machine to extract collected power battery capacity declining characteristic vector values; building a power battery system aging model; using a particle filter theory to predict the power battery system residual available life through the aging model. The method can effectively reduce the training data bulk, can improve the algorithmic prediction precision, can ensure the RUL estimator stability, and can be expected to obtain accurate and reliable prediction results in real applications.

The invention discloses a state space probabilistic multi-time sequence prediction method based on a graph neural network, and the method comprises the steps: (1) obtaining a multi-time sequence, carrying out the preprocessing of the time sequence to construct a training set, and constructing a graph structure; (2) constructing a generation model for generating prior distribution and a time sequence of the hidden state according to a graph neural network and a multilayer perceptron, and constructing an inference network for generating approximate posteriori distribution of the hidden state according to the graph neural network and a recurrent neural network; (3) constructing a loss function according to the prior distribution and the approximate posteriori distribution of the implicit state, and optimizing a generation model and deducing parameters of the network by taking maximization of the loss function as a target; and (4) during application, obtaining the hidden state estimation of the to-be-predicted sequence at the latest moment by utilizing the inference network, then calculating to obtain the prior distribution of the hidden state by utilizing the generation model according to the hidden state estimation at the latest moment, and then calculating to obtain a time sequence observation estimation value according to the prior distribution of the hidden state.

The embodiment of the invention provides a computer-executed machine learning model training method, device and equipment, and the method comprises the steps: obtaining incremental data in a current time period, and enabling the incremental data to serve as a training sample set; based on the training sample set, conducting incremental training on the machine learning model trained in the last time period, and obtaining an initial machine learning model. And inputting each test sample in the test sample set into the initial machine learning model to obtain a test result. And based on the testresult, determining the accuracy of the initial machine learning model. And if the accuracy is greater than the first threshold, taking the initial machine learning model as a machine learning model trained in the current time period. And if the accuracy is not greater than the first threshold, adding test samples with wrong test results in the test sample set into the training sample set to obtain an updated training sample set, and training the initial machine learning model based on the updated training sample set to obtain a machine learning model trained in the current time period.

The invention discloses a sample data processing method for a cervix uteri image. The method comprises the following steps: classifying and establishing; preprocessing the data; performing segmentation; data enhancement: classifying the target image data, determining differences among the target image data, and performing enhancement processing for the differences; equalization processing: aimingat the total quantity difference among various target image data, supplementing a few types of samples by adopting data fitting to realize total quantity balance among the various target image data; constructing a data set: for each type of target image data after equalization processing, randomly dividing the target image data into a training data set, a verification data set and a test data setin proportion; and model construction: on the basis of the training data set and / or the verification data set and / or the test data set, mapping the training data set and / or the verification data set and / or the test data set to a contrast data set to obtain a classification corresponding to the sample data. According to the method, the problem of data imbalance in cervical image data classificationis solved, the precision and efficiency of image classification are improved, and the effect and quality of auxiliary diagnosis are improved.

The invention relates to the field of FPGA infrastructure optimization. In order to give comprehensive consideration on key parameters influencing circuit delay, allow coordinated exploration for variability of infrastructure-level parameters and transistor-level parameters in FGPA infrastructure design phase, embody and maintain inter-parameter physical significance, reduce training data quantity for a neural network and arrive at quick, precise, low-input delay estimation, the invention provides a precise low-speed low-input FPGA delay estimation method, comprising the steps of 1) determining fitting parameter Alpha and effective mobility Mu; 2) determining load capacitance of each transistor in each sub-circuit; 3) establishing an FPGA-macro delay model for each sub-circuit in the FGPA; 4) analyzing and normalizing; 5) solving weights Omega and Phi and number m of hidden neurons so that minimum training error Et and verification error Ev are achieved. The method of the invention is applied mainly in design and manufacture.

The invention discloses an angle-dependent complex array error calibration method based on a deep neural network, and aims at the problem that a traditional signal processing method is difficult to process calibration of an angle dependent complex array error. In order to process an off-grid target, an adopted local array flow pattern interpolation method can better adapt to the angle dependent array error based on grid point data measured in a darkroom in comparison with a global array flow pattern interpolation method; moreover, the input characteristics of the deep neural network select phases under complex conditions instead of actual phases so that the jump problem of the phases at + / -pi edges can be avoided; and finally, in order to enable the neural network to adapt to noisy signals, only training data needs to be generated on data of a single signal-to-noise ratio, and training data does not need to be generated on multiple signal-to-noise ratios so that the training data volume is reduced, and the training time is shortened. Compared with the traditional signal processing method, the method has the advantages of being smaller in residual array error after calibration and better in calibration performance.

The invention relates to a positioning method and apparatus, a computer readable storage medium and a computer device. The method comprises the steps of obtaining a to-be-positioned target signal set;converting the target signal set into a target fingerprint vector through a trained vectorization model; obtaining the similarity between the target fingerprint vector and each fingerprint vector ina fingerprint library; selecting at least one fingerprint vector from the fingerprint library according to the obtained similarity; and determining position information corresponding to the target signal set according to the position information corresponding to the selected at least one fingerprint vector. According to the scheme provided by the invention, the positioning accuracy can be improved.

The invention provides a cow face detection and recognition method based on deep learning. The method comprises the steps of bovine face data acquisition and preprocessing, bovine face data conversion, bovine face feature extraction model construction and bovine face feature vector database construction, bovine face detection model construction, bovine face part detection and feature extraction byusing the constructed bovine face feature extraction model and bovine face detection model, and bovine face recognition by using an image retrieval technology. According to the method, the data volume required by model training can be effectively reduced, the cattle face part is accurately drawn, the influence of environmental factors of the non-cattle face part on the recognition accuracy is reduced, multiple cattle face feature values are comprehensively used for cattle face recognition, and the cattle face recognition accuracy is improved.

The present invention relates to the field of intelligent traffic, and relates to a camera masking intelligent detection method, in particular to a camera masking detection method using video images and based on machine learning method. The specific technological process of the method comprises the following steps: image down sampling, Gabor filtering processing, PCA dimensionality reduction processing, data training and masking judgment. The uniform LBP features of an image are used as features of learning classification, so that a support vector machine is trained. The trained parameters are used for judging whether the new image is masked or not. A collected image sample is subject to Gabor filtering processing, so that stability and accuracy of the uniform LBP features of the image sample are improved. The uniform LBP features are subject to PCA dimensionality reduction processing to reduce trained data size of the support vector machine. The method is reliable in principle, rapid in judgement, high in judging accuracy, low in implementation cost, and friendly in applying environment.

The invention relates to an FPGA architecture exploration method. The method aims at combining a precise model with a GA algorithm, and achieves trade-off optimization between delay and area by adjusting size of the transistor. To achieve the goal, the FPGA architecture exploration method comprises the following steps of (1) determining key parameters that affects the delay of the FPGA; (2) establishing corresponding ELmore delay model for each circuit; (3) combining the Elmore delay model of the FPGA with a nerve net, establishing a KBNN delay model, conducting training to the KBNN delay model, determining a weight omega and phi that minimize a training error Et and a verification error Ev, and the number of hidden neurons m; (4) establishing an improved minimum width transistor area model, and estimating the area of the FPGA island; (5) combining the delay model, the area model and the GA algorithm, achieving fast adjustment of the size of the transistor. The method is mainly applied to an FPGA design occasion.

The invention discloses a massage manipulation recognition method based on deep learning, and the method comprises the steps: collecting the force distribution and force size information of a hand during massage through a flexible touch sensor, extracting the manipulation features through a neural network, and achieving the recognition of a massage manipulation. A variational auto-encoder is adopted to realize data enhancement; extracting key frames of input data by using a frame difference method, and removing input redundant frames; extracting and training spatial domain and time domain features of the massage dot matrix thermodynamic diagram group through a two-dimensional convolutional neural network and a recurrent neural network; a frame attention mechanism is introduced after the convolutional neural network, and the massage manipulation recognition precision of the network is improved. According to the invention, the original sensor data is expanded without increasing the data acquisition cost; key frames of graph group data are extracted, the network overfitting phenomenon is reduced, and the network generalization ability is improved; the neural network extracts time domain information between the video frames to obtain time domain features of the massage manipulation; and by introducing a frame attention mechanism, the recognition precision is effectively improved.

The invention discloses a melody MIDI accompaniment generation method based on a deep neural network, and belongs to the technical field of artificial intelligence and music science and technology. The melody MIDI accompaniment generation method comprises the steps of collecting an MIDI data set with a body tailoring label, processing the MIDI data set to obtain a data set, encoding through MuMIDIto obtain entries, inputting the entries into a GC-Transformer model for training until a total loss function converges, completing training of the GC-Transformer model, finally encoding MIDI fragments only containing melody, and inputting the MIDI fragment into a trained GC-Transformer model, and finally outputting the generated accompaniment MIDI fragment. The melody MIDI accompaniment generation method has the advantages of short generation time, high generation quality, less hardware resource consumption, less training data volume and the like.

A training data sampling method for establishing a word translation model includes: first, randomly sampling a first proportion of example sentences for a target number of the word from raw data in afirst round of sampling; Labeling the example sentences obtained in the first round of sampling and storing the labeled sentences in a label data pool; Performing word embedding preprocessing on tag data in the tag data pool and obtaining data center points of each category corresponding to respective interpretations of the words; Heuristic clustering of raw data using different types of centers;enabling The example sentence acquired in the first round of sampling to be subjected to data post-processing and the processing result is fed back to the next round of sampling, so that the total number of samples reaches the target number of samples.

The invention relates to the field of vehicle intelligent perception, and discloses a vehicle behavior identification method. The method comprises the steps of obtaining the current transverse distance of a front vehicle relative to a main vehicle and the first duration of transverse distance change of the front vehicle; wherein the front vehicle is a vehicle which runs in front of the host vehicle and is adjacent to the host vehicle, and the front vehicle and the host vehicle run on different lanes; determining a first lane changing two-degree-of-freedom chi-square value of the front vehicle based on the current transverse distance and the first duration, and determining a first lane changing probability of the front vehicle relative to a lane where the host vehicle is located based on the first lane changing two-degree-of-freedom chi-square value; and based on the first lane changing probability and a first set lane changing probability threshold value, identifying the lane changing behavior of the front vehicle. According to the method, the cut-in and cut-out behaviors of the vehicle are identified by using the chi-square distribution of two degrees of freedom, and the vehicle behavior identification efficiency is improved on the basis of ensuring the vehicle behavior identification accuracy.

The invention belongs to the technical field of model training, and discloses a machine learning model training method and system for a large-scale machine learning system. The machine learning model training system for the large-scale machine learning system comprises a data acquisition module, a data preprocessing module, a parameter range determination module, a central control module, a model training module, a model test module, a model evaluation module, a model optimization module, a data storage module and an update display module. The training sample set is processed through the data preprocessing module, the feature subsets of the training sample set are obtained, and the model training data volume is reduced; the machine learning model is trained based on incremental learning, so that the accuracy of model training can be improved; the optimal parameter value is determined in the value range of each parameter through the model evaluation module and the model optimization module, and the model parameters are adjusted, so that the training efficiency of the model in machine learning is improved.

The invention discloses an electric quantity prediction method under edge equipment based on sparse anomaly perception. The method comprises the following steps: S1, collecting electric quantity dataof K target buildings by edge equipment; S2, performing sparse anomaly marking on the abnormal data at the edge equipment end by adopting a sparse anomaly sensing method; S3, calculating a sparse abnormal discard probability and obtaining a total training set; S4, with a machine learning regression algorithm and a five-fold cross validation training model, randomly discarding abnormal data by utilizing a sparse abnormal discarding probability during each-fold cross validation and not participating in training; and S5, carrying out electric quantity prediction on to-be-predicted data by utilizing the machine learning model, and multiplying model prediction output by the sparse abnormal discarding probability to obtain final prediction. According to the method, a problem that a machine learning regression model based on mean square error loss is sensitive to abnormal data is solved; and meanwhile, the training data volume is reduced, the model training speed is increased, randomness is introduced, and the prediction precision and generalization ability of the model are improved.

The invention discloses a novel power supply train multi-target optimization-oriented intelligent operation control prediction method. According to the method, original data is obtained via static anddynamic big data of a novel power supply train, and parameterization / regularization and normalization are carried out to obtain train information sequence data; train operation state data under an ideal condition is obtained by utilizing local data in the train information sequence data; the train operation state data is input into a long and short-term memory network for training so as to obtaina basic model, the train information sequence data is divided into multiple parts which are input into the basic model in sequence and then retrained to obtain a basic model with knowledge; square loss training is established; and a practically acquired speed is input into the basic model with the knowledge for prediction output so as to obtain traction at the next moment. The method is independent of mass data, so that the training data size is reduced, the model according with the train operation data distribution is obtained, the model correctness is improved and the model can be conveniently and directly used for practical application.

The invention belongs to the field of radar target recognition, and particularly relates to a radar high-resolution one-dimensional range profile target recognition method and system based on a double parallel network, and the method comprises the steps: constructing a target recognition model, and carrying out the training optimization of the model through employing the actual measurement data of a preset waveband broadband radar test field as sample data, the model comprises two branch networks which are parallel to each other and are used for carrying out extraction and mapping classification on correlation characteristics of an input radar high-resolution one-dimensional range profile data sequence, a fusion layer used for carrying out data weighted fusion on output of the two branch networks, and an output layer used for carrying out classification identification on an output sequence after weighted fusion. And identifying a high-resolution one-dimensional range profile target in a to-be-detected range by using the trained and optimized target identification model. According to the method, multi-channel coding is carried out on the HRRP sequence data to extract features, the robustness of a network model can be further improved through weight adjustment, the target recognition accuracy is improved, and the method is convenient for actual scene application.

The invention discloses a culture medium formula development method and system based on transfer learning. The method comprises the following steps: (1) obtaining a culture medium pre-training model based on a deep neural network and a culture medium formula data set for training of cells to be developed; (2) pre-training model migration; and (3) carrying out regression prediction on a culture effect value: carrying out culture effect value prediction on a culture medium formula to be predicted by adopting the final model obtained in the step (2). According to the method, a small amount of training data is adopted to carry out migration training on the culture medium pre-training model, so that the training data volume required by development of a specific cell formula is greatly reduced, the development time of the culture medium formula is greatly shortened from 4-6 months to 1-2 months, and the development cost is remarkably reduced.

The invention provides an ultra-short-term power load prediction method, which is characterized by comprising the following steps of: firstly, randomly generating a sufficient number of low-dimensional delay variable combinations according to multi-dimensional observation variables; then, mapping the data to a load at a specific future time point through a neural network prediction algorithm, obtaining a plurality of networks according to the mapping, and calculating a plurality of prediction values by each network; and finally, calculating a final prediction value through an aggregation estimation method. According to the method, the random distributed embedded framework and the BP neural network are combined, and the distribution information is created by utilizing the interaction between the high-dimensional variables to replace a time sequence lacking due to less training data, so that the amount of training data required by prediction is effectively reduced, the prediction precision is remarkably improved, and the stability is enhanced.

The invention discloses a method for predicting power consumption of edge devices based on sparse anomaly perception, which includes the following steps: S1: the edge device collects power data for K target buildings; Perform sparse anomaly labeling; S3: Calculate the sparse anomaly discard probability and obtain the total training set; S4: Use the machine learning regression algorithm and 5-fold cross-validation to train the model, use the sparse anomaly discard probability to randomly discard abnormal data during each fold cross-validation, and do not participate Training; S5: Use the machine learning model to predict the power of the data to be predicted, and multiply the model prediction output by the probability of sparse exception discarding to obtain the final prediction. The invention alleviates the problem that the machine learning regression model based on mean square error loss is more sensitive to abnormal data; at the same time, it reduces the amount of training data, improves the speed of model training, introduces randomness, and improves the prediction accuracy and generality of the model. ability.

The embodiment of the invention provides a data processing method. The data processing method comprises the steps: obtaining at least one sample of a target object; performing feature extraction on atleast one sample by using a first neural network to obtain at least one piece of first feature data, wherein the first feature data is used for indicating features of at least one sample; performingfeature extraction on at least one piece of first feature data by using a second neural network to obtain at least one piece of second feature data, wherein the second feature data is used for indicating change features of the corresponding first feature data; and making a decision according to at least one second feature data to obtain decision data. According to the data processing method provided by the embodiment of the invention, the first neural network and the second neural network are used for carrying out feature extraction on the sample twice, so the data capable of representing theevent represented by the sample and the decision data are obtained, the training data volume and the training time used by the deep learning model are reduced, and the training efficiency of the deeplearning model is improved.

The invention discloses a terahertz composite metamaterial multi-band absorber and a design method of the terahertz composite metamaterial multi-band absorber. The terahertz composite metamaterial multi-band absorber comprises a metal substrate, a dielectric layer is arranged on the metal substrate, and a metal square ring is arranged above the dielectric layer. A metal cross is arranged in the center of the interior of the metal square ring, and a metal circular ring is further arranged between the metal cross and the metal square ring. According to the present invention, the problems that in the prior art, in the design process of a traditional metamaterial absorber, continuous trial and error and case-by-case calculation and analysis need to be carried out, and meanwhile designers need to have previous analog simulation experiences and professional knowledge related to metamaterials, are solved.

The invention provides a near-field source positioning method based on an auto-encoder and a parallel network. The method comprises the steps of: generating near-field source data received by an array under the condition of a single information source, constructing an auto-encoding network and a parallel full-connection network, inputting the near-field source data into a combined network of the auto-encoding network and the parallel full-connection network, and when angle information is obtained, calculating out a single spectrum peak search mode. According to the method, the signal received by the array is subjected to subspace segmentation, and direct output of the DOA spatial spectrum under the condition of unknown signal source number is realized through the parallel full-connection network, so that the efficiency of the algorithm is greatly improved. The capability of multi-source positioning can be obtained only by training single-source data, so that the data volume of training is greatly reduced, and meanwhile, the training difficulty of the neural network is reduced to a certain extent. And an offline training and online testing process is adopted, so that the algorithm complexity in an actual use process is greatly reduced.

The embodiments of this specification provide a computer-executed machine learning model training method, device, and equipment. In the training method, incremental data within a current time period is acquired as a training sample set. Based on the training sample set, the machine learning model trained in the previous period is incrementally trained to obtain the initial machine learning model. Each test sample in the test sample set is input into the initial machine learning model to obtain test results. Based on the test results, determine the accuracy of the initial machine learning model. If the accuracy rate is greater than the first threshold, the initial machine learning model is used as the machine learning model trained in the current time period. If the accuracy rate is not greater than the first threshold, the test samples with wrong test results in the test sample set are added to the training sample set to obtain an updated training sample set, and based on the updated training sample set, the initial machine learning model is trained. To get the machine learning model trained in the current time period.

The invention relates to the field of FPGA architecture optimization. In order to fully consider the key parameters affecting circuit delay, it allows collaborative exploration of the variability of architecture-level parameters and transistor-level parameters in the FPGA architecture design stage, which not only reflects and maintains the physical meaning between parameters, but also Reduce the amount of training data for neural networks to achieve fast, accurate, and low-input latency estimation. For this reason, the present invention, the FPGA time-delay estimation method of accurate fast low investment, steps are as follows: 1) determine fitting parameter α and effective mobility μ; 2) determine the load capacitance of each type transistor in each sub-circuit; 3) Establish the FPGA-macro delay model for each sub-circuit in the FPGA; 4) analyze and normalize; 5) solve the weight Ω and Φ and the number m of hidden neurons, so that the training error E t and validation error E v minimum. The invention is mainly applied to design and manufacture occasions.

The invention relates to an FPGA architecture exploration method. The method aims at combining a precise model with a GA algorithm, and achieves trade-off optimization between delay and area by adjusting size of the transistor. To achieve the goal, the FPGA architecture exploration method comprises the following steps of (1) determining key parameters that affects the delay of the FPGA; (2) establishing corresponding ELmore delay model for each circuit; (3) combining the Elmore delay model of the FPGA with a nerve net, establishing a KBNN delay model, conducting training to the KBNN delay model, determining a weight omega and phi that minimize a training error Et and a verification error Ev, and the number of hidden neurons m; (4) establishing an improved minimum width transistor area model, and estimating the area of the FPGA island; (5) combining the delay model, the area model and the GA algorithm, achieving fast adjustment of the size of the transistor. The method is mainly applied to an FPGA design occasion.

The invention discloses a terahertz composite metamaterial multi-band absorber and a design method of the terahertz composite metamaterial multi-band absorber. A terahertz composite metamaterial multi-band absorber comprises a metal base on which is arranged a In the dielectric layer, a metal square ring is arranged above the medium layer, a metal cross is arranged at the inner center of the metal square ring, and a metal circular ring is also arranged between the metal cross and the metal square ring. The invention solves the problems existing in the prior art that the traditional metamaterial absorber design process requires continuous trial and error, case-by-case calculation and analysis, and at the same time requires the designer to have previous simulation experience and metamaterial-related professional knowledge and other limitations. .