32 results about "Softmax function" patented technology

A computer-implemented method for performing a code generation for calculating a Softmax function of a neural network includes (i) providing a displacement s and a multiplier for the quantized representation of the input tensors of the Softmax function of the neural network, (ii) creating a second lookup table to replace a nested function to calculate the EXP_ON_NEG (MUL_SAT ( )) function from a CMSIS NN library depending on the displacement and the multiplier, wherein element values of an input tensor normalized to a negative value range between 0 and a minimum value are used as arguments, wherein 0 is assigned to a maximum possible element value of the input tensor and the minimum value is assigned to the smallest possible element value of the input tensor, and (iii) implementing an access to the second lookup table in a code generated to calculate the Softmax function, so that it replaces the function call of the EXP_ON_NEG (MUL_SAT ( )) function.

The application relates to a dual-branch hyperspectral image classification method based on a graph convolutional neural network and an attention mechanism, which comprises the following steps: inputting a hyperspectral original image to data for pretreatment; inputting the data after dimension reduction into a CNN attention mechanism branch, and obtaining processed branch features through an attention mechanism module, a data processing module, an attention mechanism module and convolution operation; inputting the data after dimension reduction into a GCN branch, performing simple linear iterative clustering, segmenting into k superpixels, encoding the k superpixels, obtaining superpixel graph nodes, inputting the graph nodes into a GCN module for feature extraction, decoding the extracted superpixel node features into pixel-level branch features; performing feature fusion on the two branches, using a cross-entropy function as a loss function to train the model, and using a softmax function to obtain a class label probability. The application takes limited hyperspectral images as research objects, improves high classification precision, and ensures classification speed.

The present application relates to the technical field of fault type identification acceleration, in particular to a kind of multi-modal sound feature fusion's fluid transmission equipment fault classification acceleration method and system;After sound sensor collects fluid transmission component operation sound signal, pre-processing is removed interference noise;Time domain feature extraction module extracts multi-scale space-time feature by multi-scale time convolution fusion unit;Frequency domain feature extraction module is converted by FFT, amplitude and phase separation processing and IFFT restoration, and time domain feature after frequency domain processing is obtained;Dual-domain feature fusion module fuses dual-domain feature by real Gabor wavelet function, and then excavates deep feature by Transform module;Fault diagnosis module outputs fault classification result by convolution, fully connected and Softmax function, and hardware platform outputs recognition accuracy by UART serial communication.This application detects comprehensively, responds quickly, has strong stability, low power consumption, good portability, can be adapted to different environments, identifies multiple fault types, and can also be extended to other related fields.

This invention discloses an adversarial defense method and system for testing a visual language model, comprising: applying diverse transformations across multiple domains to an input image to generate a set of transformed images containing the original image; filtering the image features of the transformed images using entropy values ​​to obtain a reference feature set; calculating the defense direction and directional bias score based on the reference feature set to quantitatively distinguish between adversarial and clean samples; designing a dual-stream feature reconstruction strategy based on a threshold division of the directional bias score; generating reconstructed features using hyperparameters and the average displacement distance of the reference features for samples with directional bias scores exceeding the threshold, and using the mean of the reference features as the reconstructed features for samples with directional bias scores less than or equal to the threshold; inputting the reconstructed features into a visual language model, calculating the cosine similarity between the reconstructed features and the category text features generated by the text encoder of the visual language model, outputting the classification probability through a softmax function, and selecting the category with the highest classification probability as the final prediction result; the advantage of this invention is that it balances robustness and accuracy.

This invention discloses an intelligent lithology classification method based on a multi-well graph neural network. First, engineering parameters and lithology data from multiple wells are collected and preprocessed to construct graph nodes at depth points. Then, a graph structure containing intra-well edges and inter-well edges is constructed, where intra-well edges connect adjacent nodes within the same well, and inter-well edges connect nodes in different wells with similar logging responses. Next, a graph attention network is used to aggregate neighborhood information and update node feature representations. The updated features are then passed through a fully connected layer and a Softmax function to output lithology classification probabilities. Finally, a weighted cross-entropy loss function is used to train the model, and micro-batch sampling is employed to reduce computational overhead. This method introduces inter-well geological similarity through inter-well edges and adaptively fuses neighborhood features using an attention mechanism, effectively solving the problem of traditional methods neglecting inter-well correlations. It significantly improves the accuracy of cross-well lithology classification and the ability to identify rare lithologies, possessing significant engineering application value.

This invention belongs to the field of marine equipment health management technology, and discloses a method and system for assessing the health status of marine bearings. The method includes the following steps: S1, collecting marine bearing operating status data through multi-channel sensors and processing it to obtain a standardized health status sequence; S2, performing soft partitioning on the standardized health status sequence to obtain a soft-label sequence; S3, extracting and fusing time-series features from the soft-label sequence to obtain bidirectional time-series features; S4, based on the bidirectional time-series features, quantifying the prediction uncertainty using a Monte Carlo mechanism and generating a health status prediction probability through a Softmax function to obtain the marine bearing health status prediction result. This invention introduces a Monte Carlo mechanism, enabling the model to output the confidence interval of the prediction result in a noisy environment, effectively identifying low-confidence samples. It not only solves the problem of traditional methods lacking risk assessment basis, but also supports hierarchical operation and maintenance strategies, significantly enhancing the robustness of the system in complex marine environments.

This invention relates to an intelligent classification method for policy texts based on a dual-channel neural network and an attention mechanism, comprising the following steps: S1. Preprocessing the collected policy texts; S2. Embedding the preprocessed text information into word vectors using the Skip-gram model in Word2vec word vectors; S3. Inputting the word vectors into a CNN neural network and a BILSTM neural network with an attention mechanism, respectively, to obtain important local features and global features; S4. Merging the obtained important local features and global features to obtain policy text information features, inputting the obtained policy text information features into a fully connected layer neural network, and using the softmax function to calculate the classification probability of the policy text to obtain the classification result.

The application discloses an image spatial domain steganalysis system and method, relates to the technical field of image steganalysis, and comprises a CDNet model module, wherein the CDNet model module is composed of a feature extraction unit, a feature analysis unit and a classification unit; the feature extraction unit carries out pretreatment on an input image through an SRM high-pass filter, embeds a light channel attention mechanism into a residual structure, and constructs a residual channel attention block structure; four stacked composite down-sampling blocks of the feature analysis unit are jointly composed of a normal convolution, a detail enhancement convolution and an average pooling layer; the classification unit converts feature maps obtained through the preceding modules into a feature vector through a global covariance pooling layer, and outputs a classification result through a full connection layer and a softmax function, effectively enhances noise residual features extracted by a pretreatment layer, effectively reduces the loss of steganographic signals in a down-sampling process, and guarantees detection accuracy.

This invention claims protection for a speech emotion recognition method based on attention-based MCNN combined with gender information. The method includes the following steps: S1, obtaining three-dimensional dynamic Mel-frequency cepstral coefficients (MFCCs) from the preprocessed speech signal as input to a gender recognition network; S2, using MCNN for gender recognition and classifying the speech signal into male and female categories; S3, based on the output of the gender classification, extracting three-dimensional dynamic MFCC features from the male and female speech signals as input to the emotion recognition model. To focus on channel and spatial location information and address long-term dependency issues, a coordinated attention mechanism is introduced into the original MCNN model to establish the speech emotion recognition model; S4, to better capture emotional features and temporal information, A_GRUs are added to the emotion recognition model, and finally, a softmax function is used for emotion classification, providing emotion recognition results for different genders. This invention effectively solves the problems of low recognition rate and weak generalization ability of emotion recognition models caused by differences in male and female acoustic features, improving the accuracy and robustness of emotion recognition.

This invention relates to the field of perioperative neurological monitoring technology, and particularly to a postoperative cognitive quantitative assessment system and method that integrates electroencephalography (EEG) and near-infrared spectroscopy. The system includes: a baseline anchoring module, which extracts preoperative scalp EEG and near-infrared spectral signal features, concatenates them into a first vector, and calculates the covariance to generate a preoperative baseline manifold; a matrix construction module, which generates a cross-modal manifold matrix based on real-time bimodal signals within a time window; a feature extraction module, which calculates the geodesic distance between the cross-modal manifold matrix and the preoperative baseline manifold, projects it onto the tangent space using a logarithmic mapping to generate a tangent matrix, and combines the geodesic distance to generate a topological feature vector; a quantitative assessment module, which inputs the vector into a softmax function layer to output a probability distribution and calculates a cognitive assessment index; and a closed-loop intervention module, which outputs an intervention command when the index meets preset conditions and provides a timestamp to reset the time window. This system achieves closed-loop manifold-based assessment and dynamic monitoring of postoperative cognitive status.

The application discloses a millimeter wave security door target recognition model training method, belongs to the technical field of millimeter wave security door target recognition, and solves the problem of low target classification and recognition accuracy caused by the difficulty of capturing the fine-grained characteristics of targets by high-dimensional data characteristics. The method comprises the following steps: a millimeter wave security door target recognition model extracts features and converts dimensions from batch original millimeter wave security door echo data, obtains a feature map of each target category, and obtains a total score of each target category after vector aggregation; the total scores of all target categories are input into a Softmax function, and the prediction probability of each target category is calculated; a loss function is calculated according to the prediction probability output by the Softmax function and the real target category label, and the parameters of the millimeter wave security door target recognition model are updated; if the training end condition of the millimeter wave security door target recognition model is reached, a millimeter wave security door target recognition model that passes the training is obtained.

The invention relates to a computer-implemented method for performing code generation for calculating a softmax function of a neural network, comprising the following steps: - Providing (S1) a shift s and a multiplier m for the quantized representation of the input tensors of the softmax function of the neural network; - Create (S2) a second lookup table (LUT2) to replace a nested function for calculating the EXP_ON_NEG (MUL_SAT ()) function from a CMSIS-NN library depending on the shift s and the multiplier m, using as arguments the element values ​​of an input tensor normalized to a negative range between 0 and a minimum value, where 0 is assigned to a maximum possible element value of the input tensor and the minimum value is assigned to the smallest possible element value of the input tensor; - Implement (S5) an access to the second lookup table (LUT2) in code generated to calculate the softmax function, so that it replaces the function call of the EXP_ON_NEG (MUL_SAT 0) function.

The application provides a chip end face image definition evaluation method and system based on semantic information, comprising: obtaining a chip end face image to be evaluated; inputting the chip end face image to an image degradation feature extraction module, extracting the degradation features of the image and performing global average pooling processing, and obtaining a guide vector through MLP network mapping; constructing an adaptive prompt word pair, processing the image to be evaluated through an image encoder to obtain an image feature vector, and simultaneously processing the adaptive prompt word pair through a text encoder to obtain a prompt word feature vector; calculating the definition score of the image to be evaluated through a Softmax function and comparing it with a preset threshold to obtain a determination result. The application overcomes the problem that the picture definition evaluation scale is not the same under the condition that the texture quantity and density are different, and makes the evaluation score only depend on the overall definition of the picture, so as to avoid the problem that the definition of the defects at different focal depths of the chip end face image causes the score to be biased and leads to inaccuracy.

This invention provides a method for controlling the inference stage of multi-factor customized text generation, comprising: dynamically scoring and fusion calculating each factor based on the currently generated text and context factor verification information to obtain a fusion score; constructing an activation factor set based on the fusion score; normalizing the fusion scores corresponding to the factors in the activation factor set using a softmax function to obtain a final weight set; obtaining the probability distribution of the current position using the currently generated text, the final weight set, and a preset large language model; obtaining the current output word based on the probability distribution of the current position, and adding the current output word to the currently generated text as the current generated text in the above process; and repeating the above process until the final generated text is generated. This method achieves independent management of elements, adaptive stage adjustment, redundancy interference suppression, and stable optimization of effect and quality in multi-factor customized text generation.

In an example embodiment, each sample's loss is monitored during training of the segmentation model, and the amount of each sample's loss is used to determine which samples are mislabeled. For each pixel in each epoch, a logit of the model prediction class and a logit of the label class are determined, and the difference between them (margin) determined and saved. The margin then can be integrated over all epochs to derive an area under the margin (AUM) score for each pixel. A softmax function can then be applied to each sample to obtain an AUM score for each sample. Any images whose AUM score is greater than a predetermined threshold may be determined to be mislabeled and appropriate action can then be taken to fix the mislabeling, thus resulting in more accurate training and validation of the segmentation model.

Mechanisms are provided for optimizing a deep learning (DL) computer model for homomorphic encryption (HE) workload processing. The mechanisms receive an original DL computer model architecture that is to be optimized for HE workload processing, and modifying the original DL computer model architecture by replacing a self-attention layer of the original DL computer model with an HE friendly self-attention layer that comprises a Power SoftMax function that does not have exponent terms, to thereby generate a modified DL computer model architecture. The mechanisms execute a machine learning training of the modified DL computer model architecture, approximate one or more elements of the Power SoftMax function with polynomials to generate a trained HE optimized DL computer model, and output the trained HE optimized DL computer model for execution on HE workloads.

The application discloses a time series prediction method and system based on time-frequency characteristics and double-channel processing, and relates to the technical field of time series prediction. The method first acquires signal data of a communication system, divides the signal data into two independent data channels of real parts and imaginary parts, and independently and synchronously processes the two data channels; secondly, the method multiplies original sequence data by correlation weighted by a Softmax function to obtain an enhanced time-domain characteristic matrix; then, the method converts input data into frequency-domain characteristic data by using discrete wavelet transform, performs parallel convolution on the frequency-domain characteristic data by using multiple hollow convolution kernels with different expansion rates, and fuses the frequency-domain characteristic data to obtain a frequency-domain characteristic matrix; finally, the method realizes fusion of time-frequency characteristics by using a single-hidden-layer feedforward network, and outputs a final prediction result. The method realizes efficient processing and high-precision prediction of complex radio frequency data, and has obvious technical advantages and application prospects.

This invention discloses a robust stereo matching method based on feature pyramids and attention perception, comprising: setting a feature extraction pyramid to process the input binocular stereo image, and obtaining multi-scale feature maps of the left and right view input images respectively; constructing a multi-scale cost volume pyramid using the multi-scale feature maps, and implementing feedback interaction and cross-scale cost aggregation between cost volumes of different scales; converting the cost volume pyramid into a probability value pyramid using a softmax function, and processing the probability value pyramid using a soft-argmin function to generate an initial disparity map; designing a saliency attention perception module to extract and generate attention feature maps in the shallow layers of the stereo matching network, and fusing the feature maps with the initial disparity map to obtain a refined disparity map. This invention can effectively correct erroneous matching points, restore lost image details and sharp object edges, improve the robustness and precision of disparity prediction, and generate a fine and accurate disparity map.

This invention discloses a probiotic screening method based on neural networks and metagenomic data, relating to the field of probiotic screening. It involves obtaining metagenomic sequencing data from NCBI, preprocessing the data, and dividing the processed sequence data into training, validation, and test sets. A deep learning model is constructed based on a Transformer architecture combined with a HyenaOperator hybrid mechanism to capture the internal dependencies of sequence fragments at different distances, performing pooling to obtain better sequence features. These features are then input into a multilayer perceptron (MLP) base model for dimensionality reduction, and the predicted probabilities are output using a softmax function. A voting mechanism is used to summarize and determine the predicted results of the sequence fragments. This invention employs the aforementioned probiotic screening method based on neural networks and metagenomic data, enabling efficient and accurate screening of potential probiotics in large-scale microbial genome data.

The application relates to a semantic double-syntactic aspect-level sentiment analysis method, and belongs to the technical field of natural language processing. The method comprises the following steps: S1: encoding processing, using a BERT pre-training model to encode input sentences and aspect words to obtain word vector representation; S2: semantic and syntactic feature extraction, extracting semantic features through dynamic context weighting and multi-head attention mechanism, and simultaneously extracting syntactic features by combining SenticNet knowledge with double syntactic structures of a dependency tree and a constituent tree; S3: feature fusion and optimization, adopting a contrast learning method to balance different syntactic information, and fusing semantic and syntactic features through a double-affine residual block and a factorized bilinear pooling mechanism; and S4: sentiment classification, performing pooling processing on the fused features, and outputting aspect-level sentiment polarity through linear transformation and a softmax function. Through the cooperative processing of semantic and double-syntactic information, the long-distance dependence and semantic mismatch problems are effectively solved.

The application provides a text classification method based on residual semantic enhancement and a multi-scale convolutional network, and the method comprises the following steps: S1, obtaining a feature after semantic enhancement; S2, generating a local enhanced feature and obtaining a wide-area context feature; S3, splicing and fusing the local enhanced feature and the wide-area context feature to obtain a spliced feature vector, and calculating a class probability distribution of the text by using a Softmax function to perform class prediction. The application can enrich the early semantics of the bottom layer word vector representation without destroying the original word sequence, and provide an extremely robust high-dimensional semantic base for subsequent multi-scale mapping; the method can capture rich local feature combinations by using multi-size convolution kernels, and can multiply the receptive field of the model without increasing the parameter amount, so that the network has a powerful long-span chapter-level context dependence capturing capability.

This invention relates to the field of natural language processing technology and discloses a sentiment analysis method for online shopping reviews based on multi-scale feature fusion. The method includes the following steps: inputting a pre-processed text sequence into a pre-trained model ERNIE to generate word vector representations; inputting the word vectors into TextCNN, extracting multi-granularity local features through multi-size convolutional kernels, and concatenating them after max pooling to form sequence features; inputting the sequence features into a BiGRU to capture contextual sequence information; assigning weights to features at each time step through an attention mechanism to focus on key sentiment information; and finally outputting the sentiment classification result through a fully connected layer and a softmax function. This invention provides a sentiment analysis method for online shopping reviews based on multi-scale feature fusion, which can effectively mine multi-level textual features of online shopping review texts.

The application discloses a kind of methods for extracting road marking line in laser radar point cloud, comprising the following steps: S1, propose M-Transformer network, the original point cloud of road marking line is processed, the context information of original point cloud and the long-term dependence between different road marking lines are obtained;S2, based on parallel pooling attention mechanism, the feature relationship of the road marking line is aggregated;S3, the position dependence of all scales is integrated, the minimum value and average value of feature sequence are calculated, and they are connected to obtain the two-dimensional feature of road marking line point cloud data;S4, the feature weight of the two-dimensional feature is obtained using Softmax function, further enhance the semantic feature of road marking line point cloud data, obtain accurate road marking line distribution.The present application can quickly and accurately extract the road marking line of large-scale urban road scene.

The application relates to the technical field of image processing, and discloses a window local and global attention-based multi-modal image feature matching method, which uses an FPN architecture to preliminarily extract a group of features after image data enhancement, uses window attention to interact local features, selects important features in each window, interacts global information of all features, completes final feature extraction, uses a bidirectional softmax function to process features after attention interaction, trains the model, and realizes feature matching under multi-modal images. The window local and global attention-based multi-modal image feature matching method interacts local information through window attention, significantly reduces the calculation amount of global information interaction on the basis of excellent local attention interaction, has excellent matching ability and matching accuracy, has very good generalization on various multi-modal data sets, and has very high practical value.

The application discloses a bearing fault diagnosis method and device based on space-time domain fusion. The method comprises the following steps: obtaining an original feature sequence by performing feature extraction on a target vibration data set of a bearing to be measured, and constructing an original graph structure corresponding to the original feature sequence; mixing attention weights based on channel attention weights and spatial attention weights corresponding to the original feature sequence to obtain a target time domain feature; further determining a target space domain feature corresponding to each node based on the target time domain feature, an adjacency matrix and each preset diagnosis label; finally, determining a label prediction probability according to a cross-entropy function and a Softmax function, and determining a fault diagnosis result of the bearing to be measured based on the label prediction probability, so that sufficient extraction and organic fusion of double-domain features are realized, the focusing ability of the model to key time sequence fragments is enhanced, the recognition ability of key fault features under complex working conditions and noise interference is significantly improved, and the diagnosis precision and user's use satisfaction are improved.

This invention discloses a method for constructing an intelligent diagnosis and treatment system based on the Med-PaLM 2 medical large-scale model, comprising the following steps: Fine-tuning of the Med-PaLM 2 medical large-scale model based on prompting learning: setting a task for designing a pre-trained language model, designing input template styles, label styles, and mapping methods to labels, and training and fine-tuning the Med-PaLM 2 medical large-scale model; Implementation of medical large-scale model diagnosis and treatment based on multi-frequency contextual information interaction: using StreamingLLM technology, improving the Softmax function in the Attention mechanism of the Med-PaLM 2 medical large-scale model, optimizing Window Attention and injecting positional information, enhancing its multi-frequency contextual information interaction capability and increasing the text output length; The beneficial effects of this invention are: improving the accuracy of the model in diagnosing lung X-ray diseases; simultaneously, the application of multi-frequency contextual information interaction enables the model to comprehensively analyze the patient's condition information at different time points, providing doctors with more comprehensive and accurate diagnostic basis, reducing the occurrence of missed diagnoses and misdiagnoses.