67 results about "Paired samples" patented technology

The embodiment of the invention provides a government affair digital human dynamic interaction method and system based on a multi-mode large model. The method is applied to the technical field of government affair intelligent services, and comprises the following steps: acquiring a policy announcement text, and performing cleaning and structuring processing to obtain a structured policy data set; and extracting old and new policy data, performing difference comparison, marking key change fields, and generating policy change data. Abstracting and element extraction are carried out on the change data to form structured semantic fragments, and the structured semantic fragments are incrementally embedded into the policy knowledge graph. And generating a question and answer pair sample based on the updated knowledge graph, and carrying out self-supervised fine tuning training on the multi-modal large model. A user inputs multi-modal data, and the model generates a government affair response and feeds back the government affair response. According to the scheme, the multi-modal large model can continuously keep the latest policy knowledge; the model is enabled to generate accurate government affair response with consistent context while understanding multi-modal input such as text, voice and image, and timeliness, accuracy and interactive experience of policy interpretation are improved.

The invention relates to a Han-Cross industry text classification method and system based on prompt learning and adaptive loss weighting, and belongs to the technical field of natural language processing. The method comprises the following steps: designing and constructing a universal prompt template; the method comprises the following steps: recombining a Han-Cross cross-border industry text classification data set, namely converting an original single sample into paired samples; in a few-sample and multi-language scene, related vocabularies are adopted as external knowledge resources, and vocabularies most related to the mapping labels are retrieved from the related vocabularies; expanding the vocabulary mapper by introducing synonyms and associated vocabularies; adopting a dynamic mixed loss function and applying the dynamic mixed loss function to a pre-training language model to optimize a few-sample classification task; and classifying Chinese and Vietnamese cross-border industry texts by using the optimized pre-training language model. The method shows a remarkable effect in Chinese and Vietnamese industry text classification tasks, and is particularly suitable for a few-sample learning scene with data scarcity and language imbalance.

The invention discloses a steel structure quality defect tracing and analysis method based on deep learning, and the method comprises the following steps: collecting image data, sensor data and construction log information of a steel structure member, and generating a tracing identifier; performing alignment based on the traceability identifier to generate an alignment data sequence; based on the aligned data sequence, outputting a defect segmentation result by using an SE (3) isovariant graph neural network; performing continuous coherence topology analysis on a defect segmentation result, and outputting topologically continuous defect areas and severity scores; extracting process parameters of the defect area, calculating statistical dependency by utilizing an independence criterion, and screening out a paired sample set; based on the sample set, performing stability screening on the causal edges to form a causal graph for output; calculating a causal contribution score output by the causal graph, and outputting a liability sorting list; and filing the responsibility sorting list, and visually outputting a defect traceability analysis atlas and report at the same time. According to the invention, steel structure quality defect tracing and analysis are realized.

The invention provides a field NL2SQL data automatic synthesis method and system based on a large language model, and the method comprises the steps: building an SQL template library through collecting a public data set and real business query data, carrying out the grading according to the SQL complexity, and forming an SQL template set covering different difficulties; specific SQL queries are generated based on an SQL template library and a large language model, verification is carried out through multiple mechanisms such as grammar check, execution verification and result dimension consistency check, and correctness and performability of the SQL queries are ensured. Generating a plurality of candidate natural language questions according to the verified SQL query and the large language model, calculating a semantic consistency score of each candidate question and the SQL query through a cross consistency evaluation mechanism, screening out the question with the most consistent semantics as a final question-answer pair sample, and meanwhile, eliminating low-quality samples by setting a consistency threshold value, so as to obtain a final question-answer pair sample; and the data accuracy and consistency are further improved.

The invention provides a visual-text cross-modal panda behavior recognition method based on an attention mechanism, and relates to the technical field of attention mechanisms, and the method comprises the steps: inputting a multi-modal data set into an initial model, extracting panda behavior features, introducing a customized cross-modal attention mechanism to achieve feature alignment, and enhancing the visual and text feature interaction depth; the bottleneck of insufficient semantic fusion is broken through; constructing video frame sequence-text description paired samples based on features, carrying out bidirectional matching learning in a unified embedding space through a cross-modal representation network, optimizing parameters through symmetric cross entropy loss, curing a model in combination with a verification and early stop mechanism, capturing video time sequence information, and solving the problem of incomplete behavior dynamic representation; and finally, the visual part of the target panda behavior data is optimized, a preprocessing strategy is adjusted according to quality parameters, an optimization result assists an attention mechanism to focus on key features, the current situation that preprocessing does not have a unified standard and cannot be fed back is improved, and finally the target behavior category is accurately recognized.

The embodiment of the invention discloses a question and answer agent training method and device, equipment and a storage medium, and belongs to the technical field of artificial intelligence. The method comprises the steps that a question generator is trained based on a time sequence knowledge graph and a first question and answer data set corresponding to the time sequence knowledge graph, the time sequence knowledge graph comprises at least two time sequence fact groups, and each time sequence fact group comprises fact information with time sequence features; the time sequence knowledge graph and question generation cue words are input into a question generator, a second question and answer data set is output through the question generator, and the question generation cue words are used for indicating generation of question and answer pair samples for at least one time sequence fact group; and based on the second question and answer data set, training a question and answer agent, the question and answer agent being used for outputting an answer corresponding to the time sequence question based on the input time sequence question. By adopting the scheme provided by the embodiment of the invention, the diversity of the question and answer pair samples can be improved, so that the training effect of the question and answer agent is optimized.

Example solutions for fine-tuning a language model include: generating a dataset that includes a plurality of paired samples, each paired sample of the plurality of paired samples includes (i) a factual question and a true outcome for that factual question and (ii) a counterfactual question and a true outcome for that counterfactual question; submitting a factual query to an answer model, the factual query including the factual question and the true outcome of the factual question, the answer model generating a factual answer in response to the factual query; submitting a counterfactual query to the answer model, the counterfactual query including the counterfactual question and the true outcome of the counterfactual question, the answer model generating a counterfactual answer in response to the counterfactual query; and performing fine-tuning on a target model using at least the factual question paired with factual answer and the counterfactual question paired with counterfactual answer.

The invention discloses a bridge damage detection method and system, and the method comprises the steps: collecting a bridge vibration signal through an acceleration sensor, carrying out the band-pass filtering preprocessing, converting a time domain signal into a frequency domain vibration signal through fast Fourier transform, and extracting a feature frequency through gradient change detection and threshold screening; a pairing difference value sequence of a monitoring characteristic frequency and a healthy baseline frequency is constructed, and a pairing sample T test is applied to perform statistical significance analysis on a difference value, so that the abnormal frequency change caused by structural rigidity degradation is effectively identified. According to the method, complex modal analysis is replaced by simplified frequency domain features, environmental interference influences are eliminated in combination with a pairing inspection mechanism, the micro-damage detection sensitivity and the false alarm resistance are remarkably improved, and the method is suitable for engineering scenes of long-term bridge monitoring and rapid damage screening.

The invention discloses a generated text detection method and system based on lightweight rewriting conversion and a medium, and the method comprises the steps: carrying out the AI removal conversion of a to-be-detected text through a forward rewriting model, fusing the to-be-detected text and a generated human-like equivalent text, inputting the fused text into a classifier, and outputting a classification result; the training process of the forward rewriting model comprises the following steps: constructing a pairing sample set by utilizing an artificially written text and a machine generated text, and finely tuning the forward rewriting model by minimizing cross entropy loss, the machine generated text is a text which is obtained by generating and filtering redundant samples through at least one large language model in a reverse manner through the large language model; according to the generated text detection method and system and the medium, access to any external, closed source or exclusive large language model API is not needed completely, the cost and delay are reduced, the system architecture and the deployment process are greatly simplified, and the problems that a watermark technology and a similarity regeneration method are limited by a closed source model and are difficult to actually deploy are solved.

The invention provides a semi-supervised deep learning image restoration enhancement method based on dual-network cooperation. The method comprises the following steps: constructing a dual-task branch image restoration network, training the dual-task branch image restoration network by adopting a teacher-student semi-supervised training mode, and performing feature extraction and restoration enhancement processing on a to-be-restored image by a restoration branch in the trained dual-task branch image restoration network, outputting the restored image, the low-order semantic features and the high-order semantic features, fusing the low-order semantic features and the high-order semantic features by an evaluation branch, and obtaining a structural similarity index SSIM score matrix by using a self-attention mechanism; and according to the SSIM score matrix, determining the confidence coefficient of pseudo label screening in the training process of the teacher-student semi-supervised training mode of the dual-task branch image restoration network. According to the method, values of a large number of low-quality samples and a small number of high-low-quality paired samples are fully mined, high-quality pseudo labels are generated, and accurate restoration and enhancement of trackside images are realized.

The application relates to a heterogeneous feature clustering method based on a privacy image, characterized in that a diffusion generation process is converted from a pixel domain to a feature domain through a featureized noise and matrix matching technology, a to-be-trained diffusion generator model with a conditional denoising UNet architecture as a core is constructed, and training is conducted in combination with a sample-level consistency constraint and a distribution-level alignment constraint, then a feature pair with mutually aligned semantics is constructed by using a pseudo image generated by the trained diffusion generator model; subsequently, a variational autoencoder model is trained, cross-domain feature mapping is realized through a double-layer alignment loss, and finally a unified domain feature dataset is formed and clustering is conducted; the method has the advantages that cross-domain feature unification is completed under the premise that no source domain data and no paired samples are available, original image leakage is effectively avoided, a high-quality unified feature set can be directly output, and a safe and efficient solution is provided for cross-domain visual analysis under a privacy limited scene.

One or more embodiments of the invention provide a model iterative training method, a sample expansion method, equipment, a medium and a product. The model iterative training method comprises the steps of obtaining a target data set, and splitting the target data set into a training set and a verification set; the target data set comprises a plurality of questions and replies conforming to preset safety specifications; training the to-be-trained language model based on the training set to obtain a trained target language model; generating to-be-evaluated replies of the questions in the verification set by using the target language model; using the trained reply evaluation model to evaluate whether the reply to be evaluated accords with a preset safety specification, and outputting analysis information used for describing violation content in the reply to be evaluated when a conclusion that the reply to be evaluated does not accord with the preset safety specification is obtained; generating a question and answer pair sample capable of avoiding illegal content described by analysis information by utilizing a trained sample generation model; and updating the target data set based on the question and answer pair sample to be used for next iterative training for the target language model.

The present application provides a kind of thermal exposure risk time partition determination method, device, equipment and storage medium, comprising: obtaining the time series thermal exposure risk spatial distribution layer of target city in target period under predetermined time scale, hierarchical clustering analysis is carried out based on thermal exposure risk spatial pattern similarity, obtain the time partition under predetermined time scale;Based on the time partition under predetermined time scale, the average city thermal exposure risk spatial distribution layer of time partition under target time scale is obtained and paired sample difference test is carried out for the spatio-temporal partition statistics of thermal exposure risk, to determine the reliability of the time partition under predetermined time scale according to test result.The present application considers the spatial pattern of thermal exposure risk, thereby realizing the quantitative division of thermal exposure risk time partition under different predetermined time scales, and the obtained time partition is matched with the travel time of urban residents, which can effectively guide the partition control of urban thermal exposure risk.

The invention discloses an intelligent evaluation method for operation of an energy-saving facility based on artificial intelligence. The method comprises the following steps: step 1, collecting multi-source time sequence data of the energy-saving facility; 2, extracting a supervision label and binding the supervision label with the paired sample; step 3, inputting the paired samples into an improved TSMixer model, generating feature mixing weight parameters and bias parameters in a feature mixing layer through a HyperNetwork sub-module, and embedding an SE-block channel attention structure to obtain intermediate feature representation; 4, the step 3 is executed repeatedly; 5, constructing a dynamic feature fusion network; 6, inputting the comprehensive feature set into the LightGBM model, and outputting a result sequence; and 7, comparing the result sequence with the corresponding reference data in the reference library, and outputting an energy-saving facility operation intelligent evaluation result. The method has the advantages of comprehensive evaluation, efficient processing and intelligent result.

The invention relates to a heterogeneous feature clustering method based on a privacy image, which is characterized in that a diffusion generation process is converted from a pixel domain to a feature domain through a characteristic noise and moment matching technology, and a to-be-trained diffusion generator model taking a conditional denoising UNet architecture as a core is constructed; training by combining the sample-level consistency constraint and the distribution-level alignment constraint, and constructing feature pairs with mutually aligned semantics by using a pseudo image generated by the trained diffusion generator model; then training a variational auto-encoder model, realizing cross-domain feature mapping through double-level alignment loss, and finally forming a unified domain feature data set and performing clustering; the method has the advantages that cross-domain feature unification is completed on the premise of passive domain data and no paired samples, original image leakage is effectively avoided, a high-quality unified feature set can be directly output, and a safe and efficient solution is provided for cross-domain visual analysis in a privacy limited scene.

The application discloses a method for detecting glioma chromosome abnormalities based on targeted sequencing, and belongs to the technical field of biological medicine. The method first acquires the allele frequency of a to-be-detected sample at preset SNP sites (covering 1p, 1q, 19p, 19q, chromosome 7 and chromosome 10), and then calculates and determines whether specific chromosome arms or chromosomes have loss of heterozygosity. Meanwhile, the copy number of the region where each SNP site is located is calculated based on the sequencing depth, and the total copy number of the above-mentioned chromosomes is obtained by integration. Finally, the loss of heterozygosity determination result and the chromosome copy number information are comprehensively combined, so that the simultaneous identification of 1p / 19q co-deletion, gain of chromosome 7 (+7) and deletion of chromosome 10 (-10) is realized. The method does not require paired samples, can accurately quantify the copy number, avoid false positives, and only needs to detect part of the SNP sites, that is, can be combined with hot spot mutation detection, thereby saving cost and improving detection efficiency.

The invention provides a program error positioning and explanation generation method and system based on reinforcement learning, and belongs to the field of AI auxiliary programming education and natural language processing, and the method comprises the steps: S1, based on student submission records of a real teaching scene, screening out error codes and correct code pairing samples through an editing distance, and constructing a training set; s2, sampling and filtering the training set by using a pre-training language model, and constructing a sampling rejection data set to supervise and finely adjust the pre-training language model to obtain a base model; s3, constructing a multi-signal fusion reinforcement learning reward function for performing reinforcement learning training on the base model to obtain a trained error positioning and explanation generation model; and S4, inputting the question description and the error code into the trained error positioning and interpretation generation model, and generating structured feedback. According to the method, the error positioning accuracy is remarkably improved, the false drop rate is effectively reduced, and a feasible technical path is provided for intelligent programming education.

The invention discloses a shallow lake turbidity remote sensing inversion method and system based on wind drive physical constraint, and the method comprises the steps: constructing a time sequence pairing sample through combining a remote sensing image and actually measured data, and extracting wind drive characteristics such as wind speed, wind stress and strong wind duration; establishing direction consistency priori by using monotone calibration to generate an expected direction signal, and constructing a wind-driven amplitude response surface interpolation to obtain an expected change amplitude; in model training, a double-moment regression network sharing weights is adopted, and an actual measurement supervision error and wind-driven physical constraints are jointly optimized; and finally, carrying out pixel-level reasoning and verification on the multi-temporal remote sensing image to realize continuous estimation of spatial distribution and dynamic change of turbidity. According to the method, a double-moment input and shared weight network structure is introduced, the physical mechanism and deep learning advantages are fused, the time sequence continuity, the physical interpretability and the cross-space-time generalization ability are remarkably improved while high inversion precision is kept, and an efficient and reliable technical approach is provided for water environment remote sensing monitoring.

Example solutions for fine-tuning a language model include: generating a dataset that includes a plurality of paired samples, each paired sample of the plurality of paired samples includes (i) a factual question and a true outcome for that factual question and (ii) a counterfactual question and a true outcome for that counterfactual question; submitting a factual query to an answer model, the factual query including the factual question and the true outcome of the factual question, the answer model generating a factual answer in response to the factual query; submitting a counterfactual query to the answer model, the counterfactual query including the counterfactual question and the true outcome of the counterfactual question, the answer model generating a counterfactual answer in response to the counterfactual query; and performing fine-tuning on a target model using at least the factual question paired with factual answer and the counterfactual question paired with counterfactual answer.

The invention discloses a noise correlation cross-modal matching loss adjustment method and device based on meta-learning. The method comprises the following steps: constructing a pairing sample set containing a multi-modal data pair; quantifying semantic correlation of each multi-modal data pair in the pairing sample set by using a cross-modal model, and selecting an initial loss function; constructing a loss adjustment network, and balancing the noise sensitive loss function and the noise insensitive loss function to obtain an adaptive loss function; optimizing the loss adjustment network by using a meta-learning algorithm; updating model parameters of the cross-modal model on the target noise data set by using the optimized loss adjustment network; and inputting the query target and each object in the matching database into the updated cross-modal model to obtain the similarity of the query object relative to each object in the matching database. The problem that the robustness and the matching accuracy of the cross-modal model are difficult to consider at the same time in the prior art is solved, and powerful support is provided for landing of the cross-modal model in an actual scene.

This invention discloses an automatic assessment method and apparatus for deformed central cusps based on convolutional neural networks. The method includes acquiring X-ray images and intraoral photographs of the target to be detected; determining N X-ray regions in the X-ray image and N intraoral photograph regions corresponding to the N X-ray regions in the intraoral photograph; acquiring a deformed central cusp assessment model, wherein the deformed central cusp assessment model is trained using pseudo-paired samples as training samples; inputting each paired region of the target to be detected into the deformed central cusp assessment model to determine whether a deformed central cusp exists in the paired regions of the target to be detected. This invention significantly improves the assessment accuracy of deformed central cusps.

The application discloses a retrieval enhancement generation method and system based on knowledge boundary perception, an electronic device and a storage medium, and belongs to the technical field of natural language processing. The method comprises the following steps: generating a high-quality supervised track by using a teacher model, and learning the ability of gap planning and answers by instruction fine-tuning of a weak model; paired samples reflecting overconfidence and over-conservatism are constructed, and a DPO algorithm is used for confidence calibration; gap planning is generated by a student model during actual prediction, and it is accurately determined whether each knowledge point needs retrieval according to cognitive information labels, and accurate retrieval is triggered only for the knowledge points with knowledge gaps. The application can be widely applied to open domain question answering, dialogue systems and knowledge-intensive tasks by explicitly identifying knowledge boundaries, dynamically adjusting thresholds and fine-grained on-demand retrieval, while ensuring the accuracy of answers, significantly reducing the consumption of computing resources and response delay.

The invention discloses a cross-instrument Raman spectrum data alignment method based on a Cycle-GAN network, and the method comprises the following steps: S1, collecting sample data through different Raman spectrometers, and obtaining a first data set and a second data set which have systematic differences and have no paired samples; s2, carrying out denoising, baseline removal and normalization preprocessing on the data set; s3, constructing a Cycle-GAN network model containing structure-symmetric double generators and double discriminators, the generators being used for data set bidirectional mapping, and the discriminators discriminating the authenticity of spectral data; s4, using the data set to train a network model in a self-supervision form; and S5, inputting the to-be-aligned spectral data into the trained corresponding generator to obtain alignment data consistent with the spectral characteristics of the other data set. According to the method, a self-supervised bidirectional Cycle-GAN network architecture is adopted, the method has an automatic optimizing capability, does not depend on a standard spectrum, is compatible with non-paired training data, can also retain differentiated spectrum characteristics of tumor heterogeneity, and can realize accurate alignment of cross-instrument Raman spectra.

The invention discloses an instruction data set generation method and device, equipment and a storage medium. Comprising the steps of generating question and SQL pair sample data of various question and answer types based on collected small sample question and answer data; generating a template for the sample data based on the question and the SQL, wherein the template comprises card slots corresponding to various data types; generating card slot data based on a preset rule, and filling a card slot in the template according to the card slot data to obtain a plurality of expanded initial instructions; the method comprises the steps of obtaining an initial instruction, calculating a score of the initial instruction, determining proportion distribution of each problem type based on the score of the initial instruction, extracting an instruction of a preset proportion type from the initial instruction according to the proportion distribution, and generating a final instruction data set. According to the method, a large number of high-quality fine-tuning data samples can be generated by utilizing limited small sample data, so that the accuracy of the model during SQL statement generation is improved.

The application discloses a kind of celery root system to water growth new induction mode determination method, belong to plant to drought stress adaptation mechanism field of inquiry.It is set to bottle celery of certain drought stress, and it is placed on one side independent open water body, cultivate after a period of time, using ImageJ image analysis software, respectively, the difference of the total length of each celery near, far water side visible root section is measured and compared, and using paired sample t-test is carried out difference significance analysis.According to the difference result, it is clear whether celery can perceive the existence of one side independent open water body through aboveground part and induce its root system to grow in the direction of the water body, so as to further determine whether there is a new induction mode for the water growth of celery root system.The application is simple and easy to operate, accurate data, reliable conclusion, not only is the new exploration and new expansion of the concept of plant water research, but also has guiding significance for seedling cultivation and root directional regulation in horticultural production.

The invention relates to the technical field of data mining, in particular to a comparative learning entity matching method and system based on similar samples, and the method comprises the steps: carrying out the serialized representation of to-be-matched entities in a to-be-matched entity data set; the serialized representation of the to-be-matched entity is input into an entity matching model, the entity matching model is used for obtaining the matching result of each entity pair in the to-be-matched entity data set, and the entity matching model is based on a positive entity pair sample, a negative entity pair sample and a similar entity pair sample and is trained through a comparative learning mechanism; the similar entity pair samples are entity sample data with similar entity pairs but not matched with the entity pairs, so that the model learns similarities and differences among different entities. According to the method, similar but unmatched entities serve as similar samples, more comprehensive high-quality comparison samples are provided for the comparison learning process, similarity and difference characteristics of three types of sample entity pairs of positive, negative and similar samples are learned through comparison learning, and the entity matching effect is improved.

The application discloses a CT image domain adaptive lung nodule classification method based on generative same-different learning and prototype guidance, a medium and equipment, and belongs to the technical field of image processing. Paired samples of a source domain are acquired, and a lung nodule classification model is supervisedly trained. Then, samples of a target domain are acquired, and the lung nodule classification model is alternately subjected to generative same-different learning training and prototype label guided classification training until a feature extractor and a classification head converge, so that an optimal lung nodule classification model is obtained, thereby realizing effective target domain classification capability learning. Finally, a CT image to be processed is input into the optimal lung nodule classification model for lung nodule classification. The application generates enhanced views with semantic invariance by interpolation in the feature space of a generative model, and enables the classification model to learn the ability to distinguish between samples with the same semantics and samples with different semantics in a self-supervised manner, so that the feature representation capability that can be generalized to different domains is more effectively obtained.

The application belongs to the technical field of fault detection, and discloses a bearing unknown fault detection method based on steady-state and non-steady-state feature similarity mining, which comprises phase one, a pre-training phase, which is used for training similar labels or dissimilar labels of paired sample data on a labeled set, and completing training of a similarity prediction network; phase two, a new class discovery phase, which is used for identifying and discovering new classes; the application proposes a similarity feature extraction framework based on differential diagnosis, provides a fault diagnosis method based on similarity measurement and differentiation, and adds a time-frequency attention mechanism to integrate information in two dimensions of time and frequency, improve the accuracy of feature representation, and weaken the influence of noise; a salient feature deep fusion module is used to maximize the use of information, deeply mine information correlation between features, and improve the accuracy of detection results.

The application relates to the technical field of data prediction, and discloses a marine foundation bearing capacity prediction method and system based on scour pit fractal reconstruction and deep learning, a terminal and a storage medium.The method comprises the following steps: acquiring measured three-dimensional point cloud data, performing interpolation and reconstruction, and generating a three-dimensional scour pit digital geometric model; based on the three-dimensional scour pit digital geometric model, the foundation limit bearing capacity under different scour pit morphologies is calculated, and a scour pit morphology quantitative descriptor is extracted to form a data pair sample library; the deep learning model is trained by taking the data pair sample library as a training set, and a trained deep learning prediction model is obtained; the newly detected scour pit morphology data is input into the deep learning prediction model after being reconstructed and quantitatively described, and the foundation prediction bearing capacity under the current morphology is obtained. The application realizes end-to-end rapid prediction from detection data to bearing capacity evaluation, and provides an efficient and accurate decision support tool for the safe operation and maintenance of marine structures.

This application discloses a method, system, device, and storage medium for identifying facial expression intensity. The method includes: obtaining a data sample set, the data sample set comprising several facial expression sequences; collecting paired samples from the data sample set to construct a training sample set, wherein the intensity label of each sample in the training sample set is represented by a label probability distribution, the parameters of which are determined by the distribution of the original intensity labels of all samples within an observation window containing the target sample in the facial expression sequence; and using the training sample set to train a neural network-based facial expression intensity recognition model. This invention can improve the accuracy of facial expression intensity recognition.