352 results about "Softmax function" patented technology

The invention discloses a network text named entity recognition method based on neural network probability disambiguation. The method includes: performing word segmentation on an unlabeled corpus, utilizing Word2Vec to extract a word vector, converting a sample corpus into a word feature matrix, windowing, building a deep neural network for training, adding a softmax function into an output layer of the neural network, and performing normalization to acquire a probability matrix of named entity type corresponding to each word; re-windowing the probability matrix, and utilizing a condition random field model for disambiguation to acquire final named entity annotation. A word vector increment learning method without changing structure of the neural network is provided according to the characteristic that network words and new words exist, and a probability disambiguation method is adopted to deal with the problem that network texts are nonstandard in grammatical structure and contain a lot of wrongly written or mispronounced characters, so that the method has high accuracy in network text named entity recognition tasks.

The invention relates to the field of natural language processing, deep learning and multi-modal sentiment analysis, in particular to an attention fusion-based online short video multi-modal sentimentrecognition method, which comprises the following steps of: obtaining each single-modal feature in a short video; preprocessing the features by adopting a bidirectional GRU to obtain modal internal information; obtaining each advanced modal feature in combination with the interaction between the modal internal information and the modals; determining the contribution degree of each mode accordingto an attention mechanism to obtain a total feature vector, and inputting the total feature vector into a softmax function to obtain a bidirectional GRU multi-mode emotion recognition model based on attention fusion; training the model, and inputting a to-be-recognized short video into the trained model to obtain an emotion recognition result; according to the multi-modal emotion recognition method and system, all single-modal features are well fused, emotion information expressed in the video is effectively mined, and therefore the accuracy and efficiency of multi-modal emotion recognition are improved.

The invention discloses a commodity target word oriented emotional tendency analysis method, which belongs to the field of the analysis processing of online shopping commodity reviews. The method comprises the following four steps that: 1: corpus preprocessing: carrying out word segmentation on a dataset, and converting a category label into a vector form according to a category number; 2: word vector training: training review data subjected to the word segmentation through a CBOW (Continuous Bag-of-Words Model) to obtain a word vector; 3: adopting a neural network structure, and using an LSTM(Long Short Term Memory) network model structure to enable the network to pay attention to whole-sentence contents; and 4: review sentence emotion classification: taking the output of the neural network as the input of a Softmax function to obtain a final result. By use of the method, semantic description in a semantic space is more accurate, the data is trained through the neural network so as to optimize the weight and the offset parameter in the neural network, parameters trained after continuous iteration make a loss value minimum, at the time, the trained parameters are used for traininga test set, and therefore, higher accuracy can be obtained.

The invention discloses a text emotion analysis method based on attention mechanism, which comprises the following steps: 1, preprocessing text data; 2, constructing a vocabulary and constructing a word vector by using a GloVe model; Thirdly, the sentence vector is encoded by intrinsic attention, and the target word vector is encoded by interactive attention. The two encoded vectors are fused by GRU, and the fused representation is obtained after average pooling. Fourthly, according to the fusion representation, the abstract feature of context vector is obtained by point-by-point feed-forwardnetwork (FFN), and then the probability distribution of affective classification labels is calculated by full connection and Softmax function, and the classification result is obtained. Fifthly, the preprocessed corpus is divided into training set and test set, the parameters of the model are trained many times, and the model with the highest classification accuracy is selected to classify the affective tendencies. The method of the invention only uses the attention mechanism to model the text, and strengthens the understanding of the target word, so that the user can understand the emotionaltendency held by the specific target word in the text.

The present invention relates to a fine-grained zero-sample classification method based on multi-layer semantic supervised attention model. The local visual features of fine-grained images are extracted by a convolution neural network, and the local visual features of fine-grained images are supervised by the text description information of categories as category semantic features, and the local visual features of fine-grained images are weighted gradually. The loss function of the multi-level semantic supervised attention model is obtained by mapping the class semantic features to the hiddenspace local visual features. The global visual features of fine-grained images are combined with the local visual features weighted by the multi-layer semantic supervised attention model as the new visual features of images. The category semantic features are embedded into the new visual feature space, and the visual features and semantic features of the output of the multi-layer semantic supervised attention network are aligned, and the softmax function is used for classification. The method of the invention can input the extracted visual features and category semantic features, and output the classification result of the image.

The invention discloses an event joint extraction method fusing local features and deep learning. The event joint extraction method fusing local features and deep learning comprises the steps of performing entity extraction, part-of-speech and dependency analysis and PV-utilization; training document vectors by using a PV-DM model learning hidden features, and entering the GCN; thirdly, identifying a memory unit between the event element and the event trigger, and helping to identify an event type and an event argument; and finally, marking the final event type by using the CRF, and identifying event elements by using a Softmax function in a full connection layer. The method has the beneficial effects that the subject information in the article is learned by means of the document vector, then the relation between events is mined by means of the graph convolutional network, and finally the information between the event type and the event element is learned by means of the memory unit and the local feature, so that the event joint extraction is completed, and the recognition performance is improved.

The invention discloses an answer generation method based on a multilayer Transformer aggregation encoder, and the method comprises the steps: receiving input information which comprises paragraph article information and question information; converting the input information through a character embedding layer and a word embedding layer to obtain corresponding character vectors and word vectors; splicing the character vector and the word vector to obtain a spliced word vector; performing addition splicing on the spliced word vector and the position coding vector to obtain an input sequence; inputting the input sequence into a multi-layer Transformer aggregation encoder to obtain higher-level semantic information; inputting higher-level semantic information into a context-question attentionlayer, and learning question and answer information; inputting a learning result into an encoding layer comprising three multi-layer Transformer aggregation encoders, and obtaining a starting position and an ending position through a softmax function; and taking the content determined by the starting position and the ending position as a target answer. By applying the embodiment of the invention,the problems of information loss and insufficient performance in the prior art are solved.

The invention discloses a memory-based convolutional neural network system, comprising an input module, a convolutional layer circuit module, a pooling circuit module, an activation function module, afully connected layer circuit module and an output module, wherein a convolution kernel value or a synaptic weight value is stored in a NOR FLASH unit; the input module converts an input signal intoa voltage signal required by a convolutional neural network; the convolutional layer circuit module performs an convolution operation on the voltage signal corresponding to the input signal and the convolution kernel value and transmits the result to the activation function module; the activation function module activates the signal; the pooling layer circuit performs a pooling operation on the activated signal; and the fully connected layer circuit module multiplies the signal subjected to pooling operation with the synaptic weight value to perform classification, and a softmax function module normalizes the classification result to a probability value and takes the classification result as the output of the entire network. The system disclosed by the invention can satisfy the requirements of real-time data processing, and has a low hardware cost.

The invention discloses a pedestrian re-identification model, method and system for adaptive difficulty mining. The identification method comprises the steps of: randomly dividing sample pictures intoa training set used for each iteration, inputting the training set into a convolutional neural network, obtaining the probability that each sample pair belongs to a positive or negative sample pair by using a softmax function, and then obtaining the loss of each sample pair by using a multinomial logistic function; obtaining a difficult sample pair by using the loss of each sample pair; and training the convolutional neural network by using the difficult sample pair until the current number of iterations reaches the upper limit of the number of iterations, thus obtaining the pedestrian re-identification model. The pedestrian re-identification model is used to extract features of each picture in a picture set to be identified, and then a similarity order of the sample pairs in the pictureset to be identified is obtained. The pedestrian re-identification model, method and system avoid over-fitting and under-fitting, and have high recognition accuracy.

The invention discloses a medical text-oriented entity relationship joint extraction method. The method comprises the following steps of: recombining a medical text entity relationship data set according to a triple mode; vectorization representation is carried out on the medical text statement; a parameter sharing layer is constructed by adopting a bidirectional long-short-term memory network anda self-attention mechanism; predicting a head entity label by adopting a softmax function, performing relation-tail entity joint decoding by adopting CNN-softmax, enhancing joint decoding representation by combining a head entity prediction result, and training a parameter sharing layer and a joint decoding layer by adopting a joint loss function optimization mode. According to the method, the problems that entities and relationship categories and positions in the medical text are distributed unevenly, and multiple pairs of relationships appear in the same sentence at the same time are solved, and the quality and efficiency of entity relationship extraction of the medical text can be improved.

Systems and methods are disclosed that facilitate producing probabilistic outputs also referred to as posterior probabilities. The probabilistic outputs include an estimate of classification strength. The present invention intercepts non-probabilistic classifier output and applies a set of kernel models based on a softmax function to derive the desired probabilistic outputs. Such probabilistic outputs can be employed with handwriting recognition where the probability of a handwriting sample classification is combined with language models to make better classification decisions.

The invention discloses a network rumor detection method based on a pre-trained language model. The network rumor detection method comprises the steps of obtaining a to-be-detected source text and forwarded texts of a plurality of other users; preprocessing the source text and forwarding texts of a plurality of other users respectively, and connecting the preprocessed forwarding texts to obtain aset of the forwarding texts; regarding a set of the preprocessed source text and the forwarded text as a pair of sentences; and constructing a linear sequence, inputting the linear sequence into the pre-trained language model, mining a semantic relationship between the source text and the forwarded text through the pre-trained language model, and obtaining the probability that the source text is arumor or a non-rumor through a full connection layer and a softmax function. According to the method, helpful high-level semantic features can be automatically learned and acquired without dependingon specific prior knowledge, so that the method has good generalization. The method does not need to depend on a large amount of forwarding / commenting information related to the source text, and earlydetection can be achieved.

The invention discloses a laser radar point cloud multi-target ground object identification method based on deep learning, and relates to the field of point cloud identification methods. The method comprises the following steps: sequentially carrying out region segmentation, feature representation and label marking on a point cloud scene to obtain point cloud data comprising a plurality of three-dimensional spaces; establishing a network model comprising an input layer, N convolution layers, a full connection layer and a Softmax function, inputting a test set in the data set, training the model to obtain an optimal model, and inputting the test set in the data set into the optimal model to obtain an identification result; searching suspected misclassification points according to the depthinformation, the high-level difference, the spatial relationship of the power tower beside the power line and the relationship between the adjacent three-dimensional spaces, and classifying again to obtain a final identification result. According to the method, the problems of large calculation amount, difficult feature extraction and low recognition accuracy of the existing neural network due tomassive, sparse and disordered point clouds are solved.

The invention provides a graph convolutional network relationship extraction method based on dependency analysis and keywords, which comprises the following steps of: firstly, performing sentence dependency analysis on a structured text to generate a dependency analysis graph, and trimming sentences through keywords; obtaining a word embedding vector and a position embedding vector of the sentenceby utilizing word2vec, and performing splicing to obtain a word vector sequence; performing a bidirectional GRU neural network on the word vector sequence to obtain an output vector matrix; processing the output vector matrix by adopting a graph convolution network to obtain dependency representation of sentences; combining the output vector matrix and the dependency representation of the sentence through a multi-head attention mechanism to obtain a representation vector of the sentence; and establishing a relationship and entity type prediction model by adopting a softmax function, and taking the representation vector of the sentence as the input of the prediction model, thereby performing training to obtain a relationship with the maximum prediction probability as an extraction result.

The invention discloses a text classification method combining a title and a text attention mechanism. The method comprises the following steps: firstly, carrying out word segmentation preprocessing on a title and a main body of each document to obtain a title word set and a main body word set; training a word vector by adopting a word2vec CBOW model, the expression of each word combined with context semantics is learned by using a bidirectional recurrent neural network, and the potential semantic vector of one word is obtained through serial word vectors and the expression of left and right contexts of the word vectors; respectively carrying out maximum pooling processing on the potential semantic vectors of each word in the title word set and the text word set to obtain a title vector and a text vector; obtaining an attention vector by using a title and text attention mechanism; and after the vector representation of the whole document is calculated, outputting the category of the probability prediction text through a softmax function. The method solves the problem that the classification result is low in accuracy because the importance of the title content is ignored and the title is taken as one part of the text or the title information is ignored in the existing text classification with the title.

The invention relates to a human face identification method. The method comprises the following steps of 1, reading a human face image sample data set; 2, establishing a deep convolutional neural network, wherein a residual error unit is introduced in the deep convolutional neural network; 3, updating parameters of the deep convolutional neural network by utilizing a gradient descend algorithm, and firstly mapping human face images of the sample data set into 512-dimension eigenvectors through the deep convolutional neural network; calculating a loss function and a gradient of the loss function, and updating the parameters of the deep convolutional neural network according to whether a gradient descend distance of the loss function is smaller than a preset threshold or not, wherein the loss function is formed by weighting a Softmax function and an A-softmax function; and 4, performing human face identification through the deep convolutional neural network subjected to the parameter updating. Not only the between-class distance but also the within-class distance are considered, so that the human face identification rate is increased.

The invention discloses an implementation method based on various hardware platforms (such as CPLD, FPGA, special chips) for calculating a softmax function. The softmax function is widely applied to multi-classification tasks, attention models and the like for deep learning, wherein related e-exponent and division calculations need to consume many hardware resources. According to a design method,by conducting mathematical transformation on the function, the e-exponent calculation is simplified to one constant multiplication, one exponential operation of 2 with a fixed input range and one shift operation; n division operations are simplified to one highest nonzero digit detection operation, one reciprocal operation with a fixed input range, one shift operation and n multiplication operations. The exponential and reciprocal operations of 2 are achieved according to a specially-designed lookup table, and the same precision can be achieved with a smaller storage space. The method is usedin the attention models and the like for deep learning, the calculation speed can be greatly increased on the premise of almost not influencing the precision, and the consumption of calculation resources and storage resources is reduced.

The invention discloses a method for classifying surface EMG signals based on CNN and LSTM. The method includes utilizing a sliding window to convert a time sequence into a 'data-tag' pair, applying aHamming window to the surface EMG signals in each time window, using the fast Fourier transform to calculate the time-frequency spectrum Spectrogram, superimposing and integrating the time-sequence spectrum data along the time axis direction, sending the data to a convolutional neural network to complete the local spatial high-level feature extraction and obtain high-dimensional features, expanding the high-dimensional features along the data superposition dimension, restoring the data to the time sequence, feeding the data into a long and short time memory network, extracting the sequence features, sending the sequence features into a fully connected network for further feature extraction and integration to obtain the fully extracted high-dimensional features and feeding the fully extracted high-dimensional features into a Softmax function to get a final classification result. The core of the method is based on a deep learning algorithm, and the classification decoding accuracy is obviously improved by further analysis and extraction on the traditional manual extraction features.

The invention relates to an expression recognition method, in particular to an expression recognition method based on a multi-branch cross-connection convolutional neural network. The invention aims to solve the problems of low efficiency, serious resource waste and incomplete feature extraction of an existing traditional expression feature extraction method. The method comprises the following steps of: 1, preprocessing a facial expression image data set; 2, a multi-branch cross-connection convolutional neural network is constructed and used for extracting facial expression image features, andthe process is as follows: the multi-branch cross-connection convolutional neural network is composed of a first convolutional layer, a module 1, a module 2, a module 3, a forty-th convolutional layer, a batch standardization BN and a Relu activation function; and 3, classifying the image features extracted by the network by adopting a Softmax classification algorithm, namely connecting a globalmean value pooling after the constructed multi-branch cross-connection convolutional neural network, and carrying out multi-classification by using a Softmax function after a global mean value poolinglayer. The method is applied to the field of expression recognition.

The application discloses a method for knowledge extraction based on TextCNN, comprising: S10, collecting first training data, and constructing a character vector dictionary and a word vector dictionary; S20, constructing a first convolutional neural network, and training the first convolutional neural network based on a first optimization algorithm, the first convolutional neural network comprises a first embedding layer, a first multilayer convolution, and a first softmax function connected in turn; S30, constructing a second convolutional neural network, and training the second convolutional neural network based on a second optimization algorithm, the second convolutional neural network comprises a second embedding layer, a second multilayer convolution, a pooling layer, two fully-connected layers and a second softmax function, the second embedding layer connected in turn; S40, extracting a knowledge graph triple of the to-be-predicted data according to an entity tagging prediction output by the first trained convolutional neural network and an entity relationship prediction output by the second trained convolutional neural network.

The invention discloses an improved hybrid attention module-based crop pest and disease damage fine-grained identification method. The method comprises the following steps of: firstly, inputting a crop disease and insect pest picture, performing feature extraction through a convolution layer after preprocessing, and taking a feature map F obtained by the convolution layer as input of attention I _CBAM by using an Inception thought in combination with a residual connection structure in a forward propagation process to obtain weights MC (F) and MS (F); and finally, obtaining a feature map F2, and generating a final prediction probability by using a softmax function. In order to improve the accuracy of a disease and pest identification model and detect diseases and pests in time, the hybridattention CBAM is improved; through a parallel connection structure of channel attention and space attention, the problem of interference generated by serial connection of channel attention and spaceattention is solved, and the direct generalization of I _ CBAM in different models is ensured while the improvement of the accuracy of a pest and disease damage fine-grained identification model afterattention adding is more stable.

The method is suitable for the technical field of machine vision image detection and deep learning, and provides a identification method and device for mask defects, equipment and storage medium. Themethod comprises the following steps of: performing spatial transformation processing on the mask image to be subjected to mask defect recognition through a first sub-model in the multi-feature fusionconvolutional neural network model; Obtaining a first mask image, performing dimension reduction processing on the mask image through a second sub-model in the multi-feature fusion convolutional neural network model; Obtaining a second mask image, carrying out image fusion on the first sub-image and the second sub-image; Obtaining a third mask image, performing classification prediction on the image features of the third mask image by adopting a Softmax function; obtaining The classification probability corresponding to the mask defect type, and identifying the mask defects of the mask imageaccording to the classification probability, so that the distinguishing degree of mask defect features is improved through the multi-feature fusion convolutional neural network model, and the accuracyof identifying different mask defect types is improved.

The invention relates to a keyword extraction method integrating theme information and bidirectional LSTM. The method comprises the following steps: firstly, combining LDA and Skip-gram model to learnthe topic word vector representation of the word; taking the subject word vector of the word as the input of a bidirectional LSTM model; time memory characteristics of a bidirectional LSTM model arefully utilized, and semantic information of themes of words is modeled at the same time; and finally, outputting a label prediction probability of the word through a softmax function. According to themethod, the context semantic information of different distances can be fully utilized to predict the keywords, the obtained correct rate, recall rate and F value are all good, the keyword recognitioneffect obviously surpassing the prior art is achieved, and the requirements of practical application can be well met.

The invention discloses a coal slime flotation clean coal ash content prediction method based on deep learning. The method comprises the following steps: assembling an image acquisition hardware system; collecting a coal slime flotation froth image and corresponding ash content data; dividing the data set into nine types according to the + / -0.5 interval of the gray unit digit, and performing dataenhancement; adopting a resnet50 network to extract foam surface features, adopting a random gradient descent process to update parameters and softmax function classification, obtaining a high accuracy rate through multiple times of iterative training of a model, and finally making suggestions for on-site working conditions according to prediction results. Compared with manual subjective observation operation, the method has the following advantages: representative high-order abstract detail features can be automatically extracted along with continuous optimization of the model; and in addition, compared with a traditional method, the method has the following advantages: the modeling time is greatly shortened, high-order abstract features are extracted through a convolutional network, a training sample input into the model is more real, an obtained prediction result is more accurate, and the method has an important guiding effect on flotation field production.

The invention discloses a cervical OCT image classification method based on a double-channel attention convolutional neural network. On the basis of a convolutional neural network architecture, two attention mechanisms are added and realized, so that the incidence relation between features with relatively long distances on image pixels can be better captured, the weights of different high-dimensional features are learned, and accurate classification of the cervical 3D OCT images is realized. The method comprises the following steps: 1) introducing two attention mechanisms into a convolutionalneural network; 2) introducing a channel attention mechanism, preferentially using global average pooling to extract channel features of the 2D OCT image, and then using a multi-layer perceptron to learn weights of channels; 3) introducing a spatial attention mechanism, referring to a self-attention mechanism, and calculating the similarity between each feature in the feature map and other features to realize similarity calculation of non-adjacent image regions; 4) performing downsampling on the features by using global average pooling, then adding two full connection layers, and finally performing classification by using a softmax function.

The invention discloses an extreme TS fuzzy reasoning method and system based on an extreme learning machine. The method includes the steps that initial-condition attribute value matrixes corresponding to training data sets are clustered with the K-means clustering algorithm, and expansion decision attribute value matrixes are established according to clustering results; a single extreme learningmachine is trained through the expansion decision attribute value matrixes, and the output layer weight and a trained extreme learning machine are obtained; new samples are input into the extreme learning machine, and the triggering strength of a fuzzy rule antecedent and the conclusion truth value of a fuzzy rule consequent are obtained; according to the triggering strength and the conclusion truth value, defuzzification is carried out, and forecasting output of the new samples is obtained. The single extreme learning machine is trained through the expansion decision attribute value matrixes,the training process can be rapidly completed through parameter optimization without iteration, training time is short, and through defuzzification operation based on the softmax function, normativeprocessing of the triggering strength can be effectively achieved, and outputting of forecasting output data is effectively achieved.

The invention provides a fine-grain emotion element extracting method based on local information representation in order to solve the problems that by means of an existing fine-grain emotion element extracting method, when an evaluation object is extracted, the closely following word cannot be well utilized, so that the judgment on the part-of-speech of a phrase is wrong, an extracting result has many omissions, and it is difficult to judge whether the current word is one part of the evaluation object or not. The extracting method comprises the steps that for each word in a preset window size, the vector representation of word features is found through a Lookup Table, and the obtained word vectors are input into an LSTM model respectively; the obtained word vectors are combined into one vector to be input to a feedforward neural network model; the hidden layer feature representation of the LSTM model and the local context feature representation of the feedforward neural network model are merged, and a merged result is obtained; the merged result is input to a output layer and classified with a softmas function as a tag. The fine-grain emotion element extracting method is suitable for a fine-grain emotion element extracting tool.

The invention relates to a commodity recommendation method based on a gated graph convolutional network. The commodity recommendation method comprises: modeling a session sequence into an undirected graph, wherein in the undirected graph, one vertex represents one commodity, each edge represents that the user clicks the commodities at the two ends of the edge in two consecutive clicks of the session, and the weight of the corresponding frequency is given to each edge according to the frequency of occurrence of each edge in the session; initializing commodities in all sessions in the session sequence into a unified embedding space to obtain an embedding representation of the commodities in each session, and learning the embedding representation of the commodities in the sessions through a graph convolution network and a gating cycle unit; learning the embedded representation of the session according to the learned embedded representation of the commodity in the session; multiplying theembedded representation of all the commodities and the embedded representation of each session according to the obtained embedded representation of all the commodities and the embedded representationof each session, then performing normalization processing through a softmax function to obtain recommendation scores for all the commodities of each session, and performing commodity recommendation according to the recommendation scores.

The invention provides a construction method of a medical intelligent question-answering system based on an attention mechanism. The method comprises the following steps: firstly, collecting medical question and answer data, carrying out vectorization representation by using a Word2vec technology, and for each question, selecting a sentence with the maximum semantic similarity from answer data asa standard answer to construct a medical knowledge question and answer training library; constructing a deep learning model of the medical intelligent question-answering system based on the attentionmechanism, wherein the deep learning model comprises semantic coding and splicing of questions and answer data by using bidirectional LSTM, and adding the attention mechanism; and inputting the spliced semantic code into a bidirectional LSTM for feature fusion, and finally outputting a standard answer by a softmax function.