302 results about "Probability vector" patented technology

A data structure visualization tool visualizes a data structure such as a decision table classifier. A data file based on a data set of relational data is stored as a relational table, where each row represents an aggregate of all the records for each combination of values of the attributes used. Once loaded into memory, an inducer is used to construct a hierarchy of levels, called a decision table classifier, where each successive level in the hierarchy has two fewer attributes. Besides a column for each attribute, there is a column for the record count (or more generally, sum of record weights), and a column containing a vector of probabilities (each probability gives the proportion of records in each class). Finally, at the top-most level, a single row represents all the data. The decision table classifier is then passed to the visualization tool for display and the decision table classifier is visualized. By building a representative scene graph adaptively, the visualization application never loads the whole data set into memory. Interactive techniques, such as drill-down and drill-through are used view further levels of detail or to retrieve some subset of the original data. The decision table visualizer helps a user understand the importance of specific attribute values for classification.

The invention discloses a training method and a training device of a convolutional neural network (CNN) model, and belongs to the field of image recognition. The training method comprises the steps of respectively carrying out a convolution operation, a maximum pooling operation and a horizontal pooling operation on a training image so as to acquire a second feature image; determining a feature vector according to the second feature image; carrying out processing on the feature vector so as to acquire a category probability vector; calculating a category error according to the category probability vector and the initial category; adjusting model parameters based on the category error; and continuing the model parameter adjusting process based on the adjusted model parameters, and using model parameters at the moment when the number of iterations reaches a preset number of times as model parameters of the well trained CNN model. According to the invention, the convolution operation and the maximum pooling operation are carried out on the training image on different levels of convolution layers, and then the horizontal pooling operation is carried out. The horizontal pooling operation can extract a feature image marking a horizontal direction feature of the image from the feature image, so that the well trained CNN model is ensured to recognize images of any size, and the application range of the well trained CNN model in image recognition is expanded.

Devices and methods of determining a proximity of a receiver to a tag in a predetermined region. A signal characteristic is sensed at the receiver from the tag and an assisting tag. Zones are defined representing proximity of the receiver to each tag. A presence probability vector for the receiver and zones of each tag is estimated based on the signal characteristic. For the assisting tag, a further presence probability vector for the receiver and zones of the tag is estimated, given the presence probability vector for the assisting tag, based on a spatial relationship between the tag and the assisting tag. A combined probable proximity vector for the receiver and zones of the tag are calculated, using the presence probability vector for the tag and the further presence probability vector via a Bayesian network. The proximity of the receiver to the tag is based on the combined vector.

A method of retrieving similar patient cases from a medical database includes matching a current patient case against a plurality of clinical profiles resulting in a set of matching clinical profiles. The method further includes determining a degree of membership of the current patient case in each clinical profile from the set of matching clinical profiles based upon a degree of match between the current patient case and the clinical profile. For at least one clinical profile from the set of matching clinical profiles, the method includes retrieving those similar patient cases from the medical database that have a substantially corresponding degree of membership as the current patient case.

The invention discloses a radar echo extrapolation method based on a dynamic convolution neural network. The method comprises a step of offline convolutional neural network training which comprises the steps of carrying out data preprocessing on a given training image set to obtain a training sample set, initializing a dynamic convolution neural network model, training a dynamic convolution neural network by using the training sample set, calculating an output value through network forward propagation, and updating network parameters through backward propagation such that the dynamic convolution neural network converges. The method also comprises a step of online radar echo extrapolation which comprises the steps of converting a test image set into a test sample set through data preprocessing, testing the trained dynamic convolution neural network by using the test sample set, and carrying out convolution of a laster radar echo image inputted into an image sequence and a probability vector obtained in the network forward propagation to obtain a predicted radar echo extrapolation image.

A system and method for profiling subnet-level aggregate network data traffic is disclosed. The system allows a user to define a collection of features that combined characterize the subnet-level aggregate traffic behavior. Preferably, the features include daily traffic volume, time-of-day behavior, spatial traffic distribution, traffic balance in flow direction, and traffic distribution in type of application. The system then applies machine learning techniques to classify the subnets into a number of clusters on each of the features, by assigning a membership probability vector to each network thus allowing panoramic traffic profiles to be created for each network on all features combined. These membership probability vectors may optionally be used to detect network anomalies, or to predict future network traffic.

The invention relates to a log sequence anomaly detection framework based on nLSTM-self attention, and the framework comprises a training model and an anomaly detection model. The training model comprises: assuming that one log file contains k log templates E = {e1, e2L ek}, wherein the input of the training model is a sequence of the log template, the log sequence lt-h,...lt-2, lt-1 with the length of h comprises a log template li belongs to E, t-h < = i < = t-1, and the log template number | lt-h,...lt-2, lt-1 | in one sequence is equal to m < = h; enabling each log template to correspond toone template number, generating a log template dictionary, generating an input sequence from a normal log template sequence, and feeding the input sequence and target data into an anomaly detection model for training. The detection stage comprises the following steps: the data input method is the same as the training stage, anomaly detection is carried out by using the model generated in the training stage, the model output is a probability vector P = (p1, p2L pk), pi represents the probability that the target log template is ei, if the actual target data is within the prediction value, it isjudged that the log sequence is normal, otherwise it is judged that the log sequence is abnormal.

The invention relates to a video classification processing method and device, computer equipment and a storage medium. The method comprises the steps of obtaining a target video; extracting multi-modal data of the target video; performing classification prediction on the data of each mode to obtain probability vectors which respectively correspond to each mode and are used for representing the probability that the target video belongs to each preset category; combining the probability vectors corresponding to the modalities; and predicting the final category of the target video according to the combined vectors. According to the scheme, the accuracy of video classification can be improved.

The invention provides an event detection model training method and device and an event detection method. The event detection model training method comprises the steps that training image frames in anumber of training videos with tags are acquired; a target neural network is used to extract feature vectors for the training image frame in each training video; in the unit of each training video, aself-attention mechanism processing network is used to carry out at least two rounds of weight assignment on a feature vector matrix formed by the feature vectors of each training video; the weight-assigned feature vector matrix is input into a class prediction network for class prediction to acquire a probability vector of the event classification result of the training video; and an event detection model is trained according to the probability vector of the event classification result and the result of the comparison between tag vectors formed by the tags of the training videos. According tothe method, computational resources and training time consumption are reduced without affecting the accuracy of the model.

The invention discloses a markov-model-based position prediction method. The method comprises: a historical track is collected, data centralization probabilities of all sampling positions are determined, normalization processing is carried out, discrete probability distribution is determined, and a variable-order global markov model is established; according to a historical track of each moving object, an individual markov model of each moving object is constructed; and on the basis of linear regression, the global markov model and the individual markov models are combined to generate a probability vector linear combination, time period division is carried out, all tracks are mapped to the time periods according time stamps, probabilities of falling into all time periods by all objects are calculated, clustering is carried out, and then a next position is predicted by combining a clustering result and the markov models. According to the invention, the time factor is taken into consideration and different models are trained at different time periods. When a next position is predicted, a proper model is selected based on the time stamp, so that the prediction accuracy is improved substantially.

The invention discloses a text keyword extracting method based on a subject model. The method comprises the following steps: firstly obtaining a probability matrix WT of the lexical item and the subject of a training text set between the lexical item and the subject obtained through training by the subject model from a great deal of text training sets by using the subject model method ; further obtaining a probability matrix B of the lexical item and the subject of candidate keyword composed of the set of probability vectors of the subject and the lexical item in a candidate keyword set A, and obtaining a word frequency weight vector D of the candidate keyword corresponding to the candidate keyword set, cyclically computing by using the probability matrix B of the subject of the candidate keyword through the weight vector of the lexical item of the candidate keyword and the subject vector of the text to obtain the finally modified text subject vector and lexical item weight proportion vector, and thus extracting the keyword of the text. According to the text keyword extracting method based on the subject model, the error in keyword extraction due to different lengths of texts is reduced, and the keyword more proper to represent the text content is extracted.

The invention relates to a deep neural network in artificial intelligence, and provides a default probability prediction method and device, computer equipment and a storage medium. The method comprises the steps of acquiring an insurance source identifier, and inquiring whether a matched target entity exists in a preset knowledge graph or not according to the insurance source identifier; When thematched target entity exists, obtaining a to-be-predicted entity associated with the target entity in a preset knowledge graph; And based on the state attribute information of the to-be-predicted entity and the association relationship between the to-be-predicted entity and the target entity, obtaining a risk early warning feature vector, inputting the risk early warning feature vector into a trained risk early warning model for prediction, obtaining a default probability vector, and obtaining the default probability of the to-be-predicted entity according to the default probability vector. Byadopting the method, the efficiency and accuracy of default probability prediction can be improved.

The invention relates to a high throughput WLAN (Wireless Local Area Network) Mesh network rate selection method. A terminal firstly distinguishes impact loss from noise jamming loss according to a selected data sending mode; the loss is distinguished by adding negative frame NACK (Negative Acknowledgement) for a basic access mode, and the loss is distinguished by the sending and receiving conditions of an acknowledgement frame ACK and a clear-to-send frame CTS for an RTS/CTS (Request To Send/Clear To Send) mode. The accurate estimation of channel state is realized on the basis of self learning of local information of a local MAC (Media Access Control) layer acknowledgement frame to instruct the selection of the rate of a physical layer. The terminal utilizes the maximum system throughput as a target function by maintaining a rate selection probability vector. When each rate in the system is selected for certain times, the terminal starts to update the probability vector so as to ensure the astringency and the stability of the rate selection algorithm, and then, the terminal selects any one rate for data transmission with the probability in the probability vector.

The present invention discloses a method for obtaining a neural network model after adversarial distillation. The neural network model has a forward network with a feature layer structure and a softmax layer outputting probability vectors in multiple classes. The method is suitable for execution in a computing device, and comprises the steps of: adding a zooming layer between the forward network of an original neural network model and the softmax layer according to a distillation temperature, and generating a first neural network model; employing a first tag of a training sample itself to train the first neural network model, and obtaining a second neural network model; inputting the training sample into the second neural network model, and outputting a second tag expressing the probability vector of the training sample in the multiple classes through the softmax layer; employing the second tag and the first tag to perform constraint training of the second neural network model, and obtaining a third neural network model; and deleting a zooming layer in the third neural network model, and obtaining a neural network model after adversarial distillation. The present invention furtherdiscloses a corresponding computing device.

The invention discloses a construction method of a deep neural network model and a fault diagnosis method and system, and relates to the technical field of communication. The construction method comprises the following steps: determining an alarm root derivation rule based on a service path, and in a service topology of a target network, taking adjacent service nodes in the service path from a source end to a destination end as a relationship between a client layer and a service layer; a unified diagnosis factor matrix used for diagnosing faults in the service path is constructed based on theexpert fault diagnosis data, the unified diagnosis factor matrix comprises a root node and alarms and performance state indexes on service nodes associated with the root node, and the root node is a service node generating a source alarm; and constructing a deep neural network model by taking the unified diagnosis factor matrix as an input and the probability vector of the fault cause type as an output, and training and verifying by using sample data. According to the invention, based on comprehensive and effective alarm and performance state indexes, the constructed model can quickly and accurately perform fault diagnosis and positioning.

The embodiment of the invention discloses a point cloud classification method and device. The method comprises steps of acquiring multiple target obstacle blocks according to original laser point cloud; acquiring static probability vectors and dynamic probability vectors of the target obstacle blocks; and according to the static probability vectors and the dynamic probability vectors, determining the types of the target obstacle blocks. According to the invention, accuracy of point cloud classification can be improved.

The invention provides a multi-label classification method of electronic medical record based on symptom extraction and feature representation. Considering the influence of disease and symptom and thecorrelation between symptoms on the multi-classification problem of disease label of electronic medical record, two different symptom representation methods are adopted: using TF-IDF builds symptom vectors and learns symptom vectors using word2vec. The two symptom vector sequences extracted from the electronic medical record are respectively used as input sequences of the two two-way LSTM models,and the two-way LSTM models are trained. For the electronic medical records of unknown disease tags, two symptom vectors corresponding to the symptoms are extracted to form two symptom vector sequences, and two trained bi-directional LSTM models are input respectively to obtain two probability vectors. The weighted combination of the two probability vectors is used to obtain the final classification vector. This method has good classification effect and applicability.

The embodiment of the invention provides a method for selecting a network in a heterogeneous network, which comprises the following steps that an access network at each decision time is determined according to benefits, state transition probabilities and state probability vectors of each network in various states; and according to the access network at each decision time, the heterogeneous network obtains the largest system benefit, wherein the decision time is the time for sending an access request or a leave request to the network by a user, and the state of the network is the number of various users probably to be accessed to each network. Because the correlativity of state changes of each network in a time quantum is considered, the access network at each decision time is determined according to the benefits, the state transition probabilities and the state probability vectors of each network in various states at each decision time in a plurality of continuous decision times so that the heterogeneous network obtains the largest system benefits during the decision period comprising the plurality of the decision times.

A system for soft-decoding of Reed-Muller coded information has one or more rows of decoding blocks, each decoding block having a soft-output device and a Reed-Muller message passing device. A first soft-output device of a first decoding block processes a coded signal and a zero value probability vector. Each subsequent soft-output device processes the coded information and a non-zero value probability vector. The system for soft-decoding Reed-Muller coded information decodes a code-bit reliability vector from a soft-output device to generate an updated codeword reliability vector, which is used by a next decoding block in a sequence of decoding blocks to reprocess the coded information using the updated reliability vector. The reliability vector is updated through processing in each decoding block to optimize the reliability vector for extraction of the transmitted information from the received information.

The invention discloses a store and commodity recommendation method, a store and commodity recommendation device, equipment and a readable storage medium. The store and commodity recommendation methodcomprises the following steps that S1, acquiring to-be-predicted data, wherein the to-be-predicted data comprises user data and store commodity data; S2, obtaining a user feature vector representingthe user data and a store commodity feature vector representing the store commodity data according to the user data and the store commodity data in the step S1; S3, inputting the user feature vector and the store commodity feature vector into a neural network model, and outputting a probability vector; and S4, predicting the shopping demand of the user according to the probability vector, and recommending corresponding shops and commodities of the user within a certain distance range.