31 results about "Layered learning" patented technology

A design application allows one or more authors to create a learning path associated with a given topic. The learning path includes content curated from various online resources. The design application allows the authors to aggregate the content from the various online resources. The design application allows the author to curate content to include or associate with nodes. The design application also allows the authors to arrange the nodes in a hierarchy. When the learning path is complete, the design application presents the learning path to users. Users access the content within the learning path according to the hierarchy of the nodes. Advantageously, the author is able to present content, which may otherwise be distributed across the internet, within a cohesive learning path.

The invention discloses a land utilization change and carbon reserve quantitative estimation method based on remote sensing data. The method comprises the following steps: downloading an image; preprocessing the image; using and classifying land; calculating ground object carbon density according to ground survey data; making correlation analysis on the carbon reserves in the sample plots and the characteristic values, and selecting the characteristic values with significant correlation for modeling; and normalizing the screened characteristic values as an input layer of the convolutional neural network, putting the calculated carbon density of each sample plot into a network output layer, carrying out network training, and carrying out carbon reserve quantitative estimation on a to-be-studied region by utilizing a trained model. The invention is based on a hierarchical learning architecture of the multi-scale convolutional neural network, so that a land utilization classification result is better. On the basis of different feature values in the image and the carbon density obtained from ground survey data, the nonlinear relation between the feature variables and the carbon reserves is better fitted, and the final quantitative estimation result of the regional carbon reserves is improved.

The invention provides a micro-grid analog simulation training system based on the virtual reality technology. The micro-grid analog simulation training system comprises a virtual reality host and anoperation module, wherein the virtual reality host comprises a scene creating module, a mode selection module, a feedback module and a background module; the operation module is used for operating thebehavior of a virtual character, controlling the advancing of the virtual character, operating a device button and a switch, and operating the content in microgrid control software. For the system, by virtue of the virtual reality technology, the micro-grid training mode is changed, so that the training environment is vivid and controllable, the learning motivation of students is improved, the virtuality and reality combination and virtuality and reality interaction between virtual training and real training are achieved, and the gradual and layered learning is realized.

The invention discloses an apple virus recognition method based on deep learning, and the method comprises the steps: building a fruit damage recognition network through the improvement of an originalresidual network structure and a loss function based on deep learning; for a training problem under a small sample, carrying out model training by combining strategies of transfer learning and a hierarchical learning rate in a network training process; and compressing the finally obtained model, so as to reduce the deployment cost of the model and improve the recognition efficiency. Theoretical guidance and technical support are provided for crop disease recognition.

The invention discloses an image classification method based on a deep ridgelet neural network, and mainly solves problems that the prior art based on the neural network is long in image classification training time, and is not high in classification precision. The method comprises the implementation steps: 1, selecting 10% of data in an image library as a training sample, wherein the remaining data serves as test samples; 2, building a network structure of the deep ridgelet neural network, and enabling the training sample to serve as the input of the network; 3, carrying out the layered learning of parameters of each layer in the deep ridgelet neural network through a ridgelet auto-encoder; 4, enabling a parameter result of layered learning to serve as the initial values of parameters in the deep ridgelet neural network, carrying out the training of the parameters in the whole network through a gradient descending method, and obtaining a trained network; 5, inputting the test samples into the network, and obtaining a class label of each test sample. The method is high in classification precision, is high in training speed, and can be used for target detection and analysis and the detection of social activities.

The invention provides a training method and a device for determining a neural network of a molecular inverse synthesis route. The method comprises the following steps: for a plurality of molecules, when the inverse synthesis route of each molecule is determined, a hierarchical learning concept is adopted, the training process of the molecular inverse synthesis route which needs to be explored toa large depth is split into multiple layers for training, and the complete inverse synthesis reaction process is replaced by the multiple layers of molecular inverse synthesis routes; after the training of one layer of molecular reverse synthesis route is completed, representative molecules are selected as starting molecules of the next layer of molecular reverse synthesis route by means of a molecular screening mode, so that the exploration efficiency of the molecular reverse synthesis route can be effectively improved, and the accurate cost value information of the molecules can be extractedmore efficiently. By means of the layering mode, a large amount of calculation expenditure caused by determination of the molecular reverse synthesis route is greatly reduced, and on the basis that the accuracy of the molecular reverse synthesis route is guaranteed, the time for determining the molecular reverse synthesis route is shortened.

The invention discloses a high spectral image denoising method combining Bayes layered learning with space-spectrum combined priori; the method comprises the following steps: carrying out three dimensional slide block segmentation for a high spectral image according to a space spectrum correlation and non-local self-similarity thereof, and using relative distance priori based on fused features tonon-locally select a plurality of block data pieces to serve as synergy block data, wherein the selected block data is most similar to to-be-observed block data; using layered priori to build a Bayeslow-rank decomposition model, and realizing synergy block data learning and expression. The mode uses low-rank decomposition to carve statistics characteristics of synergy block data, and combines Dirichlet process mixing Gaussian distribution to express noise statistics characteristics; the method uses a variation Bayes method to solve the model, thus effectively reducing image noises. An existing method cannot simultaneously inhibit a plurality of noises in the high spectral image; the high spectral image denoising method combining Bayes layered learning with space-spectrum combined priori can solve said problems, is accurate in result, and can provide strong analysis basis for follow up applications.

Techniques are disclosed for building a dictionary of words from combinations of symbols generated based on input data. A neuro-linguistic behavior recognition system includes a neuro-linguistic module that generates a linguistic model that describes data input from a source (e.g., video data, SCADA data, etc.). To generate words for the linguistic model, a lexical analyzer component in the neuro-linguistic module receives a stream of symbols, each symbol generated based on an ordered stream of normalized vectors generated from input data. The lexical analyzer component determines words from combinations of the symbols based on a hierarchical learning model having one or more levels. Each level indicates a length of the words to be identified at that level. Statistics are evaluated for the words identified at each level. The lexical analyzer component identifies one or more of the words having statistical significance.

Systems and methods are provided for performing medical imaging analysis. Input medical imaging data is received for performing a particular one of a plurality of medical imaging analyses. An output that provides a result of the particular medical imaging analysis on the input medical imaging data is generated using a neural network trained to perform the plurality of medical imaging analyses. Theneural network is trained by learning one or more weights associated with the particular medical imaging analysis using one or more weights associated with a different one of the plurality of medicalimaging analyses. The generated output is outputted for performing the particular medical imaging analysis.

The invention discloses a fast optimization method for dynamic dispatching of a cross-regional interconnected power grid based on knowledge transfer. Firstly, the optimized optimal knowledge matrix ofeach source task is stored in a knowledge base as empirical knowledge in the pre-learning stage of the source task; then the source task with the highest similarity to the target task is obtained inthe learning stage of the target task, and the initial knowledge matrix of the target task is obtained by transferring its optimal knowledge matrix so as to perform fast optimization of the target task; and finally the optimal knowledge matrix of the target task is stored in the knowledge base as the empirical knowledge. The dispatching mechanism can select a reasonable action scheme according tothe actual operation state of the power grid at the dispatching time under the obtained strategy to realize dynamic dispatching of the cross-regional interconnected power grid. The mechanism of hierarchical learning and knowledge transfer can avoid the problem of "dimension disaster" of reinforcement learning to a certain extent, accelerate the convergence speed of the algorithm and promote fast solving of the dispatching strategy.

The invention discloses a time delay and energy consumption compromise model in an extended unmanned aerial vehicle network and a hierarchical learning algorithm. The time delay and energy consumptioncompromise model is characterized in that: an unmanned aerial vehicle network is divided into clusters, and each cluster comprises a large unmanned aerial vehicle serving as a cluster head, a relay unmanned aerial vehicle and a small unmanned aerial vehicle group serving as an extended unmanned aerial vehicle group; the relay unmanned aerial vehicles form a core network, the extended unmanned aerial vehicles form an extended network, idle unmanned aerial vehicles serve as the relay unmanned aerial vehicles, the relay unmanned aerial vehicles are used for assisting the extended unmanned aerialvehicles in transmitting information to the cluster heads, and the relay unmanned aerial vehicles are further used for adjusting a bandwidth strategy of an unmanned aerial vehicle network system to improve the bandwidth utilization rate and realizing compromise between time delay and energy consumption of unmanned aerial vehicle network communication according to a coupling relationship between power and bandwidth resources of unmanned aerial vehicle network communication. In combination with other structures or methods, the defect that the time delay and throughput performance and the resource utilization rate are not considered at the same time when the system delay problem of the unmanned aerial vehicle networks is optimized in the prior art is effectively avoided.

The present teaching relates to method, system, medium, and implementations for machine learning for time series via hierarchical learning. First, global model parameters of a base model are learned via deep learning for forecasting time series measurements of a plurality of time series. Based on the learned base model, target model parameters of a target model are obtained by customizing the base model, wherein the target model corresponds to a specific target time series from the plurality of time series for forecasting time series measurements of the specific target time series.

The invention provides a multi-AUV (Autonomous Underwater Vehicle) collaborative navigation method based on reinforcement learning. According to the method, a collaborative navigation process is divided into two main processes: firstly, a trajectory planning process: a trajectory of a slave AUV in a cluster is obtained through manual planning according to a task to be executed, and a trajectory of a master AUV is obtained through a trajectory planning method based on hierarchical Q learning; and then, a navigation calculation process: a proper nonlinear filtering algorithm is selected to carry out actual navigation calculation. Test verification shows that after the master AUV sails according to the trajectory planned by the method and then the slave AUV performs navigation calculation by using nonlinear filtering, the positioning error of the slave AUV can be obviously reduced.

The invention relates to an anti-interference channel selection method and system for a cognitive satellite-ground network, and the method comprises the steps: modeling an anti-interference decision problem as a Stackelberg game based on an obvious layering behavior between cognitive users and external malicious interference, building a lower-layer sub-game model of a graph game according to the characteristic that the interference between the cognitive users presents local influence, and carrying out the calculation of the anti-interference channel selection of the cognitive satellite-ground network. And finally, the corresponding sub-game model is converged based on the hierarchical learning algorithm of local information interaction and the Q learning algorithm, compared with other algorithms, the convergence performance is outstanding, and the system average throughput of the algorithm is close to the optimal NE solution, so that the problems that the algorithm adopted in the prior art is poor in convergence effect and inaccurate in channel selection result are solved, and the channel selection efficiency is improved. The current situations of frequency shortage and low spectrum utilization rate are relieved, and the spectrum resource utilization rate is improved.

The invention discloses a communication network architecture generation method and device, electronic equipment and a medium. By applying the technical scheme of the invention, the distributed unit DU and the centralized unit CU of the base station equipment can be utilized, and the distributed unit DU and the centralized unit CU form a communication network architecture together with the edge node. In this way, the model parameters of the device nodes in the communication network aggregation network can be aggregated subsequently to obtain aggregated model parameters, and the aggregated model parameters are used for hierarchical learning training to obtain a hierarchical federated learning model for service processing at a user equipment end or a base station end. And thus, the purpose of optimizing the service processing efficiency by using the communication network architecture is achieved.

The invention discloses a multi-extreme learning machine-based post-fault transient stability hierarchical evaluation method for a power system. The multi-extreme learning machine-based post-fault transient stability hierarchical evaluation method comprises the steps: extracting key features capable of representing the transient stability condition of the current power system from historical big data of the power system according to a maximum correlation minimum redundancy algorithm MRMR; carrying out hierarchical learning training on the ELMs by using the extracted key features of the transient stability condition of the power system; and then inputting the key feature group of the power system after the fault into the trained ELMs of each layer, and finally obtaining a final transient stability evaluation result by utilizing a decision evaluation criterion of each layer. The multi-extreme learning machine-based post-fault transient stability hierarchical evaluation method has high precision and meets the rapidity requirement of big data of the smart power grid.

The invention discloses a spectrum overhead-based presale system model and a hierarchical learning algorithm. According to the model, a spectrum manager sells resources to a user to profit, a presalemechanism helps lower impact caused by uncertain user demand, so as to increase profits of a spectrum resource manager and the user. The algorithm comprises steps of setting up a multi-user hierarchical game, wherein participants are spectrum managers and users; joining in a neighbor alliance randomly by a user, calculating expected utility, comparing utility functions before and after the user joins the alliance, selecting the superior resolution, and forming a stable alliance structure; pre-purchasing spectrum resources from the spectrum manager by the alliance, and updating resource provision volume by the spectrum manager; and cycling and iterating till users make up a stable alliance. The method effectively increases direct gains of the spectrum manager and the user.

The invention provides a method for synthesizing three-dimensional human motion by using a recurrent neural network based on hierarchical learning. The method comprises a model training step and a model testing step, the model training step comprises the following sub-steps: constructing a low-layer motion information extraction network by adopting a GRU unit; establishing a high-level motion synthesis network by adopting a GRU network; adopting motion data with different styles as input of a high-level motion synthesis network, combining the skeleton features of the motion data with the motion features extracted by the low-level motion information extraction network as input, and training the high-level motion synthesis network to learn skeleton space-time relationship information of motions with different styles; and inputting the first 30 frames of data of the motion data in the test set into the trained high-level motion synthesis network, and finally performing verification. The method not only can be used for synthesizing the motion conforming to the input track, but also can be used for generating the transition motion between two different types of motions, and can also beused for synthesizing motion sequences with different emotion styles.

The invention discloses a multi-region active power distribution system peak regulation optimization method considering power consumption demand elasticity. The method comprises the following steps: firstly, establishing a mathematical model of a photovoltaic and all-vanadium redox flow battery energy storage system and a multi-region flexible load scheduling unit in an electric power elastic environment; then, establishing a DTMDP model for the peak regulation optimization problem of a multi-region active power distribution system considering power consumption demand elasticity; and finally, solving the mathematical model in combination with reinforcement learning and an intelligent algorithm to obtain a multi-region scheduling optimization control strategy meeting peak regulation requirements. According to the invention, the hierarchical learning mechanism in the invention avoids the problem of curse of dimensionality of reinforcement learning to a certain extent, and promotes rapid solution of a scheduling strategy; the exploration capability of the algorithm is further enhanced through combination of reinforcement learning and an intelligent algorithm, and the optimal peak regulation strategy can be obtained; and potential dispatching information of the active power distribution system can be further obtained by considering electricity demand elasticity, and stable and safe operation of the system is promoted.

The present application provides a neural network training method and device for determining molecular retrosynthetic routes. The method includes: for multiple molecules, when determining the retrosynthetic route of each molecule, the concept of layered learning is adopted to disassemble the training process of the molecular retrosynthetic route that needs to explore a large depth into multiple layers for training, and multi-layer The molecular retrosynthetic route replaces the complete retrosynthetic reaction process. After completing the training of a layer of molecular retrosynthetic route, the molecular screening method is used to select a representative molecule as the starting molecule of the next layer of molecular retrosynthetic route, which can effectively Improve the exploration efficiency of molecular retrosynthetic routes, and then more efficiently extract accurate molecular cost value information. By layering, it greatly reduces the large amount of calculation overhead caused by determining the molecular retrosynthetic route, and reduces the time for determining the molecular retrosynthetic route on the basis of ensuring the accuracy of the molecular retrosynthetic route.

The invention discloses a small sample target detection method for structured filter learning, and the method comprises the following steps: preparing a small sample target detection data set and a base class target detection data set, and building a network structure; the backbone feature extraction network uses CSPDarknet53, and a general feature enhancement module is introduced; constructing a small sample target detection deep neural network learned by a structured filter; training the network by using the base class data set; on a small sample target detection data set, using a KSVD algorithm to hierarchically learn a dictionary of an input feature map of the structured filter, using the dictionary as a weight parameter of the filter, and completing initialization of the parameters of the structured filter through a forward propagation process; continuously training on the small sample data set by using the initialized network model; and completing a detection task by using the trained neural network. According to the method, the target detection deep neural network with good generalization performance can be trained only through a small amount of sample data.

The invention provides a cascading neural network-based membrane pollution multi-index identification method, which is characterized in that the method considers the interference of irrelevant input variables on output indexes, can avoid the influence of the irrelevant variables on output based on a cascading structure, and adjusts the parameters of a cascading neural network through a hierarchical learning algorithm. The membrane pollution multi-index prediction is realized, the problem that the membrane pollution multi-index prediction precision cannot meet the requirement is solved, and the method has relatively high prediction precision and identification effect.