49 results about "Learning by example" patented technology

The invention provides a visual defect detection method and system based on context example learning, and relates to the field of power equipment detection.The method comprises the steps that on the basis of a power equipment image, text description of the power equipment image is generated through a large model; inputting the image and the text description of the power equipment into a trained anomaly detection network, judging the running state of the power equipment, and obtaining a detection result of whether the power equipment is abnormal or not; wherein the anomaly detection network takes a query image and a text description thereof as well as a plurality of sample images and text descriptions thereof of the same equipment as input during each training, and performs judgment by comparing image-text features of the query image and image-text features of the sample images; learning whether a query image is abnormal or normal under the condition of a given context example; according to the method, the multi-mode learning capability of a large model is combined, and the visual defects of the power equipment are efficiently detected through context example input of images and texts.

The invention relates to a business process information extraction method based on a large language model. The method comprises the following steps: collecting business process text data; calling a large language model to pre-process the collected business process text data; performing context example learning on the large language model, and extracting business process information in the preprocessed business process text data through the learned large language model; generating a visual business flow chart based on the business flow information; according to the method, the automation degree and efficiency of business process information extraction are improved, the accuracy and consistency of the information are remarkably improved, complex natural language description can be processed, key information can be automatically extracted, a standardized business process model is generated, high flexibility and adaptability are achieved, and the method is suitable for popularization and application. Adjustment and optimization can be carried out according to specific requirements of different enterprises and industries, and diversified business process modeling requirements are met.

The invention provides a weak supervision video anomaly detection method and system based on prototype orthogonality, and the method comprises the steps: inputting a visual feature sequence into a selective state space model, so as to filter redundant information in a time sequence, capture a key dynamic state, and output a high-value time sequence feature; then, a learnable prototype codebook containing a normal category prototype codebook and an abnormal category prototype codebook is given, end-to-end training is carried out by adopting a multi-instance learning framework of a video-level label, prototype orthogonality constraint is applied in the training process, and after training is completed, the normal category prototype codebook and the abnormal category prototype codebook are subjected to end-to-end training; according to the method, only the distance between the high-value time sequence feature and the nearest prototype in the abnormal category prototype codebook is needed, and the video anomaly score is obtained according to the distance, so that video anomaly detection is realized. According to the method, the global geometric constraint of prototype orthogonality is introduced, and the selective state space model (S3M) with efficient calculation is combined, so that the extremely light weight of the model is realized while the high detection precision is ensured.

The invention discloses a hierarchical optimization multi-instance learning method for digital pathological section brain tumor classification. The hierarchical optimization multi-instance learning method comprises the following steps: extracting and aggregating multi-instance features through alternate training of a feature encoder and an aggregator; a region of interest is automatically searched in a low-magnification image, and then feature extraction and classification are performed on a high-magnification region of interest. According to the method, the region-of-interest is detected by using low magnification, and then the tumor category is comprehensively judged in multiple magnification near the region-of-interest in stages. More importantly, in a model training link, a feature extractor is trained by utilizing a pseudo tag of patch in a region of interest, so that features are better extracted for a specific task and a data set. In order to reduce the noise of the false label, the patent provides a label correction mechanism to ensure the purity of the false label. In general, the HOMIL combines low-magnification tissue structure characteristics and high-magnification cell characteristics, comprehensively predicts brain tumors, and greatly reduces the model reasoning time.

The invention relates to the field of computer vision processing, in particular to a video retrieval method based on a deep neural network model and multi-instance learning, which comprises the following steps: pre-training a query text to obtain an initial feature, and extracting an I 3D-RGB feature, an ROI feature and a connection feature from a video; updating to obtain frame-level visual features and word-level text features; constructing a training graph, and performing graph attention network learning to obtain word-level text features; calculating feature residual errors of the word-level text and the word-level text, wherein the mean value of the residual errors is sentence-level text features; performing segment dimension average operation on the frame-level visual features to obtain pipeline-level visual features; calculating alignment scores of the sentence-level text features and the pipeline-level visual features, constructing a positive sample pair and a negative sample pair, and training a video retrieval network; according to the method, discriminability features in a multi-query text are obtained through deep learning features to construct the graph neural network, so that text features with more representation meanings and multi-modal alignment supervision signals are provided under weak supervision setting.

The application relates to the field of computer vision processing, in particular to a video retrieval method based on a deep neural network model and multi-example learning, which comprises the following steps: obtaining initial features by pre-training a query text, extracting I 3D-RGB features, ROI features and connection features from a video; updating frame-level visual features and word-level text features; constructing a graph for training, learning word-level text features by using a graph attention network; calculating the residual error of the word-level text features and the word-level text features, and taking the mean value of the residual error as a sentence-level text feature; performing segment dimension average operation on the frame-level visual features to obtain pipeline-level visual features; calculating the alignment score of the sentence-level text features and the pipeline-level visual features, constructing positive sample pairs and negative sample pairs, and training a video retrieval network; and the application constructs a graph neural network by acquiring discriminative features in multiple query texts through deep learning features, so as to provide text features with more representation meanings and multi-modal alignment supervision signals under weak supervision settings.

The present invention discloses a method and device for detecting abnormal network traffic based on PU-MIL, which relates to the field of network security technology. The method includes: dividing a network data stream containing multiple data packets into positive packets and unlabeled packets containing multiple examples; constructing an autoencoder model with a two-component loss function; through deep learning backpropagation training, iteratively solving the minimum value of the two-component loss function to obtain a data stream-level classifier and a data packet-level classifier; using the data stream-level classifier to discriminate abnormal data streams, and using the data packet-level classifier to detect abnormal data packets. By introducing an abnormal traffic detection technology combining positive sample, unlabeled sample learning and multi-instance learning, the present invention can efficiently and accurately identify abnormal traffic in the network when there are only a small number of known abnormal samples and a large amount of unlabeled data.

Techniques and architecture are described for converting tree structured data such as, for example, JavaScript Object Notation (JSON) data, into multiple feature vectors to train multiple instance learning (MIL) models for providing cybersecurity in networks. In particular, a data set is provided, wherein the data set comprises a sample configured as a hierarchal tree. The sample is converted into a set of path and value pairs, e.g., flattened into a set of path and value pairs, where the path is a sequence of field names and array indices encoding a position of a value. Each path and value pair of the set of path and value pairs is converted into a respective feature vector to form a set of feature vectors. The set of feature vectors is used to train a multiple instance learning (MIL) model, wherein each feature vector has a same, fixed length.

This invention discloses a method and system for identifying encrypted traffic based on spatiotemporal features and semantic alignment. The method first extracts the spatial and temporal feature sequences of the network flow, and uses a byte-pair encoding algorithm to convert the spatial packet length into discrete symbols. Then, the discrete symbols and temporal features are mapped to a high-dimensional space and fused together. A global spatiotemporal feature vector is extracted through a network using a concatenated one-dimensional convolution and multi-head self-attention mechanism. Next, the text semantic bullseye matrix of fine-grained behaviors of various known applications is obtained offline from a large language model. Finally, the similarity between the spatiotemporal features and the text bullseye is calculated, and a multi-instance learning max-pooling mechanism is introduced for dynamic routing. Based on this, a contrastive learning loss function optimization model is constructed or cross-modal inference is performed. This invention completely overcomes the conceptual drift problem caused by changes in encrypted features, achieving extremely high generalization accuracy and feature interpretability across generations.

This invention provides an online monitoring method and system for airborne flight runaway precursors of civil aircraft based on flight test data. The method includes: constructing a set of input flight parameters for identifying airborne flight runaway precursors; acquiring daily operational data and flight test data of the target aircraft model; and extracting physical feature information of airborne flight runaway by referencing the aerodynamic mechanism model and extreme flight envelope boundary of the target aircraft model. Based on the domain adaptation concept, an offline precursor recognition model is constructed that integrates physical feature information embedding, meta-learning, and multi-instance learning. Based on knowledge distillation technology, the recognition capability of the offline precursor recognition model is transferred to a lightweight network constructed from gated recurrent units to generate an online precursor monitoring model, enabling real-time precursor probability calculation and early warning. Finally, an intelligent agent model based on a dual-delay deep deterministic policy gradient algorithm is constructed to verify the effectiveness of the online precursor warning. This invention overcomes the cross-domain data gap and meets the requirements of online lightweight computation.

The invention discloses an open-pit mine slope stability index construction method based on dynamic and static data fusion. The method comprises the following steps: acquiring slope dynamic monitoring data of a site, and inputting the slope dynamic monitoring data into a convolutional neural network for processing to obtain a feature map; a data set of the multi-instance learning model is a dynamic feature instance package, the multi-instance learning model outputs a prediction result of the dynamic feature instance package and converts the prediction result into a softmax probability, and a stability level is obtained; obtaining slope static monitoring data, inputting the slope static monitoring data into the GBDT model to obtain a prediction result, and converting the prediction result into a softmax probability to obtain a static stability index of a corresponding station; and the multi-layer perceptron MLP combines the stability level and the static stability index into an input vector to carry out stability level prediction processing so as to obtain a final comprehensive prediction result. According to the slope stability evaluation method, the multi-example learning model, the GBDT model and the multi-layer perceptron MLP are combined, the efficient and accurate slope stability evaluation method is formed, and high-timeliness and high-precision slope stability evaluation can be provided through fusion of dynamic and static data.

The invention discloses a PU-MIL-based network abnormal flow detection method and device, and relates to the technical field of network security. The method comprises the following steps: dividing a network data flow containing a plurality of data packets into a positive packet and an unlabeled packet containing a plurality of examples; constructing an automatic encoder model with a two-component loss function; iteratively solving the minimum value of the two-component loss function through deep learning back propagation training to obtain a data flow level classifier and a data packet level classifier; an anomalous data stream is discriminated with a data stream level classifier, and an anomalous data packet is detected with a data packet level classifier. According to the invention, by introducing an abnormal traffic detection technology based on the combination of positive sample learning, unlabeled sample learning and multi-instance learning, the abnormal traffic in the network can be efficiently and accurately identified under the condition that only a small number of known abnormal samples and a large number of unlabeled data exist.

The invention relates to a lung cancer EGFR (Epidermal Growth Factor Receptor) gene mutation prediction method based on multi-instance learning and Transform technology, which comprises the following steps of: segmenting a lung CT (Computed Tomography) image into image blocks for instance-level feature extraction, providing an extra supervision signal for instance-level features by utilizing a soft pseudo-marking technology in a self-generating manner, guiding a multi-instance learning model to more effectively distinguish instances, and predicting the mutation of the lung cancer EGFR gene mutation. The identification capability of instance-level features is enhanced; unique relative spatial position information is integrated into a self-attention mechanism, so that the dependency relationship between instances can be flexibly and clearly mined, and the expression capability of heterogeneity tumors can be greatly improved; the importance scores of different instance features are automatically learned through the self-adaptive gating instance aggregation module, dynamic weighting of feature aggregation instances is facilitated, and the generalization of the model on different data sets is improved through the method. According to the prediction method provided by the invention, heterogeneity characteristics in the tumor can be caught and analyzed finely, and it is ensured that prediction of the EGFR mutation state is more accurate and high in pertinence.

Disclosed embodiments of the present invention provide a computer-implemented method for predicting peptide binding and presentation by MHC molecules, the method comprising: collecting or generating training data comprising a set of MHC molecules present in a biological sample and a set of observed peptide sequences presented by at least one of the MHC molecules present in the biological sample, where it is unknown to which particular MHC molecules the peptide sequences bind, and the training data is organized as a plurality of bags, each bag having a set of training instances, where labels are known for the bags but unknown for the training instances; θ We train the MIL classifier f at the instance level and then apply the labels of multiple new instances to the MIL classifier f θ and / or predict the labels of multiple new bags by directly applying the MIL classifier f θ and making predictions by applying the criterion 1 to each instance in each bag and aggregating the results across all instances in each bag.

The application belongs to the technical field of image recognition, and discloses a weakly supervised interstitial lung disease lesion recognition method based on multiple example learning, which comprises the following steps: step 1: acquiring CT image samples; step 2: selecting part of the CT images, and manually labeling the lung parenchyma in the images; step 3: establishing a lung parenchyma segmentation model through a saliency segmentation algorithm, inputting the manually labeled CT image samples to perform training and testing, and obtaining a trained lung parenchyma segmentation model; step 4: training a lesion recognition model using a multiple example learning algorithm; step 5: acquiring a to-be-recognized CT image sample, inputting the lung parenchyma segmentation model to perform segmentation, and then inputting the segmented data sample into the lesion recognition model to obtain a lesion position. The application can realize the visual labeling of interstitial lung disease through a small amount of labeling, and greatly improves the recognition efficiency.

The present invention discloses a multi-instance breast cancer whole-slice pathology image classification method based on a combination of self-supervision and weak supervision methods, including: the method is divided into two stages: MoBY self-supervised contrastive learning and Transformer-based weakly supervised multi-instance learning. In the first stage, a contrastive learning strategy is used to train on a large amount of unlabeled data. Swin Transformer is selected as the backbone model, and through the self-attention mechanism and hierarchical feature representation, it better adapts to the changes and differences of different breast cancer pathology images, and preliminarily learns the tissue characteristics of breast cancer pathology images. In the second stage, using the weight initialization model of the first stage, a multi-instance learning method is used to regard each full-slice image as a package, and the generated small slices as examples in the package. The top-n examples with the highest probability in each full-slice image are selected through the multi-instance learning method to optimize the model to reduce the impact of category imbalance. The selected key examples are then used to construct an aggregator through the Transformer encoder to more effectively explore the correlation between the selected breast cancer examples and adaptively highlight the most unique example features.

According to the twin multi-instance learning hyperspectral target detection network and detection method based on Transform, through a spectrum embedding module, a semantic feature extraction module and target feature enhancement fusion operation, in the training process, the target features are extracted, and the target features are extracted; the network inputs a positive sample packet pair, namely two positive sample packets or a negative sample packet pair, namely one positive sample packet and one negative sample packet through a twin structure, and extracts spectral features of each sample through a spectrum embedding module to obtain packet embedding; the positive sample packet is embedded through a semantic feature extraction module, the semantic relationship between a target and a background is adaptively learned through a self-attention mechanism, the obtained semantic features are further converted into weight scores through a full-connection network, and the positive sample packet is embedded into weighted fusion to obtain representative target features; the negative sample packet is embedded without semantic feature extraction; jointly optimizing the model through InfoNCE comparison loss and cross entropy loss, back-propagating and updating network parameters, and strengthening the distinguishing capability of a target and a background; according to the method, the coding capability of spectrum embedding is effectively improved, and the detection distinguishing effect of target and background features is enhanced.

This invention relates to a method and system for detecting unbalanced malicious traffic based on coarse-grained data labels. The method includes: S1: collecting malicious traffic data and normal traffic data at different time periods to construct a dataset, labeling each dataset with a coarse-grained label y, and labeling each data point x... i S2: Inherit the label y of its data set; i Multiple clustering algorithms were used to cluster malicious traffic data according to different sub-features, and the estimated score s was calculated. i S3: x i Dense vector representation of e i Input a multi-instance learning model and output each x. i The predicted value p i , using s i Construct a loss function to train a multi-instance learning model; S4: Based on p i Updates i When p i If the traffic is malicious, then increase the value by s. i The value is s, and vice versa. i The method provided by this invention only requires coarse-grained data labeling and utilizes estimated malicious data scores to achieve a malicious data prediction accuracy of the multi-instance learning model that approaches the performance of the supervised learning model with fine-grained data labeling.

According to the pathological image classification method and device based on multi-instance learning, better feature representation can be generated, and the problems of domain migration, data hunger and thirst and model overfitting can be effectively solved. The method comprises the following steps: (1) preprocessing a pathological image; (2) feature extraction; (3) feature aggregation; the step (2) comprises the following steps: (2.1) training a feature extractor of a full-section pathological image WSI by adopting self-supervised contrast learning; (2.2) performing data enhancement based on clustering representation; and (2.3) gradually discarding a network layer by adopting a regularization technology.

The invention discloses a full-section pathological image classification method based on diffusion model feature transformation, and relates to the field of computational pathology. The method comprises the following steps: acquiring a paired HE dyeing image and IHC dyeing image, and extracting image features thereof; building a dynamic feature conversion network model, wherein the model comprises a feature encoder and a cross-modal dynamic diffusion module; the feature encoder comprises an HE stream and an IHC stream which are respectively used for extracting HE features and IHC features; and the cross-modal dynamic diffusion module takes the HE features as conditions, ensures the consistency of diagnosis semantics by comparing semantic bridging strategies, adaptively processes cross-modal distribution differences by using the frequency domain expert hybrid module, and finally generates target IHC features through a conditional denoising diffusion process. According to the feature conversion method, the IHC features with high quality and consistent semantics can be generated, the classification performance of a multi-instance learning framework is remarkably improved, an intermediate pixel image does not need to be generated, the calculation efficiency is high, and a new normal form is provided for calculation of biomarker prediction.

Systems, methods, and computer programs disclosed herein relate to training a machine learning model and using the trained machine learning model to classify images, preferably medical images, using multiple-instance learning techniques. The machine learning model can be trained, and the trained machine learning model can be used for various purposes, in particular for the detection, identification and / or characterization of tumor types and / or gene mutations in tissues.

The present invention discloses an abnormal data stream classification and detection method and device based on unbalanced multi-instance learning, which relates to the field of network security technology. The method includes: dividing network traffic into a positive packet set and a negative packet set; performing oversampling operations on all positive packets in the positive packet set to obtain an expanded new positive packet set; performing undersampling operations on all negative packets in the negative packet set to obtain a reduced new negative packet set; constructing a cost-sensitive loss function based on cost weights and class weights; constructing a classifier according to the cost-sensitive loss function; training the classifier, and updating the cost weights of the classifier according to the number of missed alarms and false alarms in each round of training; and identifying network traffic using the optimized classifier. The present invention can efficiently and accurately detect abnormal data streams in network traffic.

This invention discloses a multimodal data scene recognition method based on multi-level interactive fusion. Using video data and vehicle data collected by onboard sensors in autonomous driving scenarios, three single-modal features are extracted: 2D-level features from the video are obtained through multi-instance learning based on a two-stage attention mechanism; 3D spatiotemporal features from the scene video are extracted through a multi-layer spatiotemporal attention network, and onboard information feature vectors are added for training and interaction; and the onboard information feature vectors are trained. After feature extraction of the three modalities, similarity loss is calculated. During training, the similarity of the three modalities is maximized, and the features of the three modalities are interacted based on a multi-layer self-attention network. Finally, classification is performed. This method can utilize existing video and onboard information interaction to supplement information and improve the speed and accuracy of scene recognition.

The present invention relates to an industrial Internet patent identification method based on multi-example learning. Natural language processing technology is used to divide the abstract information in the patent into sentences, and a text topic sentence extraction algorithm based on a sentence relationship graph is used to extract the topic sentence in the abstract, which can effectively reduce the computational overhead. At the same time, by combining the topic sentences extracted from the title and the abstract, the patent is converted into a sentence package, where each patent is regarded as a package and each sentence in the package is regarded as an example. Finally, the category of the new sample is predicted by adopting the K Nearest Neighbors (KNN) algorithm. The method of the present invention can effectively improve the industrial Internet patent identification effect, greatly reduce the cost of manual review, and has a very important significance for patent retrieval.

The invention provides a hierarchical MIL (Multi-Example Learning) WSI (Whole Slide Image) classification method based on pseudo-bag hybrid enhancement, which can solve the problems of high calculation complexity and insufficient model generalization ability in the existing WSI classification. According to the method, firstly, WSI is divided into a plurality of pseudo bags through phenotypic clustering and layered sampling, then enhanced data is generated based on a pseudo bag mixing strategy, and finally learning and classification are carried out by adopting a multi-level architecture. A pseudo bag mixing strategy is combined with a multi-level framework, and through cross-sample semantic alignment, label smoothing and level compatibility, the problems of small samples, high noise and calculation complexity in WSI data can be effectively solved, and the accuracy of a result is improved.

The application discloses a long video retrieval method and device based on multi-scale multi-example similarity learning. The method acquires video and text preliminary features; uses coarse-to-fine coding mode to extract information of different time granularities from video segment scale and frame scale; based on video representation of two scales, uses segment scale similarity learning branch to filter out video segments most relevant to the text and obtain segment scale similarity; uses frame scale similarity learning branch to aggregate video features guided by the filtered most relevant video segments to obtain more detailed video information, and after similarity calculation with the text, frame scale similarity is obtained; a common space learning algorithm is used to learn multi-scale similarity between long videos and texts, and a model is trained in an end-to-end manner to realize text-to-long video retrieval. The application uses the idea of multi-scale multi-example learning, and can effectively solve the text-to-long video retrieval task.

This invention discloses a pathological image classification method based on self-motivated multi-instance learning. This method promotes better and more reliable decision-making by exploring the potential relationships between instances and the reciprocal relationship between feature representation and label prediction. The invention captures and aggregates tumor feature representations by introducing label-related category priors provided by pseudo-packet prediction to achieve high-accuracy pathological image label prediction. Conversely, the network is optimized based on the prediction results to further improve feature representation and thus improve pseudo-packet label prediction, forming a self-motivated learning mechanism. This invention introduces a multi-level feature fusion strategy to explore knowledge of current instances and global historical instances, while constructing a time comparison module to improve the robustness of feature representation and alleviate representation bias and overfitting problems. Furthermore, the self-motivated feature fusion module utilizes the mutual refinement mechanism between pseudo-packet prediction and feature representation to enhance the accuracy and reliability of pathological image classification.

The application discloses a low-resolution face recognition system based on deep learning and monitoring equipment, and relates to the field of deep learning and monitoring equipment fusion. The system mainly comprises: collecting face images under monitoring videos to construct a data set, including face images under clear videos and face images under low resolution. The method is a multi-example learning method, adopts a self-supervised contrast learning mode to compare the feature differences between positive and negative samples, adopts a multi-scale patch embedding to facilitate improvement of model performance, fuses multi-scale information, adds a Dropout layer in a multi-layer perceptron of a Swin Transformer, increases the generalization ability of data, prevents a certain neuron of the Swin Transformer from dominating the final result, so as to neglect the results of other neurons, and facilitates classification and prediction of the whole image.

A method, apparatus and system to train a MIL text classification model for classifying a text content bag includes determining a first classification estimate for text content instances of the text content bags using bag-level information, determining a second classification estimate for the text content instances of the text content bags using the first classification estimates by applying a contrastive learning technique, determining a pseudo classification label for each of the text content instances of the text content bags using the second classification estimates, determining a combined loss including a first loss associated with a bag constraint loss determined from a bag index of each text content instance, a second loss associated with the contrastive learning technique, and a third loss associated with the determination of the pseudo classification label, and guiding the training of the MIL classifier using the combined loss.