202 results about "Cosine Distance" patented technology

The invention relates to the technical field of big language models, and discloses a big language model training method based on knowledge graph enhancement, comprising the following steps: S1, constructing a multi-source heterogeneous knowledge graph; s2, coding the mixed attention heterogeneity map; s3, bidirectionally mapping a pre-training task; and S4, position specific gating fusion. According to the big language model training method and system based on knowledge graph enhancement, a same proton graph is established for a structured triple and text entity description, nodes are connected across graph edges to form a heterogeneous graph, associated edges are established through entity linking and syntactic analysis, multi-source knowledge is modeled in a unified mode, and the problem of low fusion efficiency is solved; mixed attention coding adopts a layering mechanism, a semantic level calculates weights according to type compatibility, a node level calculates similarity aggregation features through cosine distance and path length, entity vectors are generated through pooling, map structures and semantics are explicitly learned, reasoning accuracy is improved, and the problem of knowledge understanding superficial layer is solved.

The invention discloses a micro-seismic multichannel data processing method and system based on a DBSCAN algorithm, and belongs to the technical field of micro-seismic monitoring. According to the invention, through feature vector construction and optimization distance calculation, the clustering reliability is enhanced; a multi-dimensional feature vector Xi containing P-wave refinement arrival time ti, an amplitude Ai and space coordinates (xi, yi, zi) is constructed, distance calculation is performed by combining a normalized cosine distance with wave velocity constraint, a micro-seismic wave propagation rule is fully utilized, the discrimination capability of DBSCAN clustering on physical consistency is enhanced, and the robustness of a screening result is improved, so that noise interference can be effectively processed in a complex environment, and the accuracy and the robustness of the screening result are improved. And the stability and the reliability of a micro-seismic positioning result under different geological conditions are ensured.

The invention provides an online multi-target multi-camera vehicle tracking method based on deep learning. The method mainly comprises the following steps: capturing vehicle video images from a plurality of cameras; identifying and positioning a vehicle instance by using a YOLOv11 algorithm; 2048-dimensional appearance features of the vehicle are extracted through a ResNet101IBN convolutional neural network; applying a single-camera multi-target tracking algorithm to generate a target track under each camera; and through a hierarchical clustering algorithm, in combination with feature cosine distance and Dunn index optimization, clustering is performed on a target trajectory, and cross-camera target matching is completed. The method iteratively executes the steps, and meets the multi-target vehicle tracking requirement in a continuous video stream. By integrating a lightweight vehicle detector, efficient appearance feature extraction, accurate single-camera tracking and an advanced cross-camera association strategy, the real-time performance, accuracy and robustness of vehicle tracking are remarkably enhanced, and the method is suitable for complex multi-camera monitoring environments such as urban monitoring, traffic management and automatic driving assistance systems.

The invention discloses an adaptive clustering federated learning method based on a prototype network in a heterogeneous scene, and the method comprises the steps: obtaining the class prototype information of different classes in each client through information extraction, constructing the user prototype representation of the client, employing a cosine distance as the similarity measurement of different clients, and carrying out the iterative clustering based on the similarity measurement. The client evaluates whether data distribution changes or not after a certain round of training, samples new data of the client to obtain a query set, generates prototype representation of each category in the query set, calculates a loss function of the client in the query set, and if the loss function value is too large, the client quits current cluster training, and if the loss function value is too large, the client does not quit current cluster training. Adding federated learning training as a new federated learning participant; when a new client joins in federated learning training, the central server is responsible for calculating the probability that the new client belongs to each cluster and further determining which cluster the new client joins or forms a cluster by itself, so that the expandability of clustering federated learning is increased, and the prediction performance of a personalized model is improved.

The invention discloses an AI-based BOM material information automatic coding method, which comprises the following steps of: unifying data structures of BOMs of products at different stages, storing the data structures as a single BOM data set, screening single data with a single bit number, comparing part of data with a standard database, cleaning and then comparing when the data are not consistent with the standard database; a knowledge graph, vector conversion, cosine distance and the like can be used for further processing so as to convert single bill of material data into standard data; for data which cannot be matched, multiple strategies can be adopted to select matching entries; the invention further relates to format recognition of the bill of materials and selection of processing tools so as to ensure data integrity. When data of a single bill of material is denoised, key parameters are extracted in combination with multiple data sources, association rules are mined through a data mining algorithm, and denoising operation is executed. The method can improve the accuracy and standardization degree of material data, greatly improves the coding efficiency, and achieves the intelligentization and refinement of material management.

The invention discloses a communication interference signal open set identification method, system and device and a medium. The method comprises the following steps: carrying out short-time Fourier transform on a communication interference signal to obtain a time-frequency graph; performing multi-scale feature representation on the time-frequency diagram by using a Res2Net-based multi-scale feature extraction network; constructing a projection layer to carry out hyperspherical embedding on the multi-scale features, carrying out modeling by utilizing vMF distribution, and designing a compactness and dispersity loss function; multi-scale features after modeling are input into an OpenMax layer for open set recognition, output of a projection layer is used as an activation vector to improve OpenMax, Euclidean distance is replaced with cosine distance, input sample unknown is calculated based on an extreme value theory, category probability distribution is adjusted, and some abnormal samples can be classified into unknown categories; the Res2Net more flexibly captures multi-scale information than the convolutional network; the separation of inter-class features and the compactness of intra-class features are effectively enhanced through the hyperspherical vMF distribution; the OpenMax layer endows the network with unknown category identification capability; the method has better performance and wider universality; the system, the equipment and the medium are used for realizing the communication interference signal open set identification method.

The invention relates to an interpretable model robust training method based on topological regular terms, and belongs to the field of artificial intelligence safety. The method comprises the following steps: firstly, performing semantic preserving disturbance on an original sample to obtain a disturbance sample; secondly, inputting the original sample and the disturbance sample into a target model, and respectively calculating gradient values to generate corresponding interpretation images; secondly, calculating gradient differences before and after sample disturbance based on a cosine distance and an Euclidean distance, and extracting topological features of interpretation images before and after sample disturbance by using a persistent coherence method to quantify topological differences; and finally, taking the gradient difference and the topology difference as regular terms, forming total loss by the regular terms and the cross entropy loss, and dynamically adjusting the weight of the regular terms according to the proportion of each difference in the total loss. Aiming at the problems that the anti-interference performance of an existing method is influenced by only utilizing a gradient difference feature training model and a fixed regular term weight is difficult to adapt to generalization of a multi-type disturbance reduction model, the invention proposes that the robustness of model explanation is effectively improved by utilizing topological features of an explaining image.

The invention discloses an individual electroencephalogram conversion system, method and device fusing vision and electroencephalogram, and particularly relates to the technical field of brain-computer interfaces, and the method comprises the steps: obtaining and preprocessing electroencephalogram data of a subject and a corresponding stimulation image; respectively encoding the electroencephalogram data and the stimulation image of the subject by using an image encoding module and an electroencephalogram superficial layer feature extraction module; image encoder parameters are fixed, and electroencephalogram superficial layer feature extraction module parameters are updated through a multi-modal fusion strategy and self-adaptive contrast learning to obtain a pre-trained electroencephalogram superficial layer feature extraction module; electroencephalogram data between the subject pairs are obtained and preprocessed; initializing electroencephalogram converter model parameters, and loading pre-trained electroencephalogram superficial layer feature extraction module parameters; inputting source subject electroencephalogram data into an electroencephalogram converter to generate target subject electroencephalogram data, calculating time domain cosine distance loss, time domain mean square error loss and frequency spectrum constraint loss with real data, and updating electroencephalogram converter parameters to obtain a final electroencephalogram converter model.

The invention discloses an electric power equipment safety long text knowledge retrieval method based on sparse and dense joint enhancement, which comprises the following steps of: encoding an electric power equipment safety long text and retrieval query by utilizing a pre-training language model for performing supervision and fine tuning on an electric power equipment safety field label data set, and generating a semantic word vector; a weighted sparse vector fusing semantic information and a dense semantic vector enhanced through nonlinear transformation are generated through fine-grained semantic enhancement branches; respectively calculating sparse similarity based on a common item weight product and dense similarity adopting a double-distance weighted fusion strategy combining a cosine distance and an Euclidean distance; and carrying out weighted summation on the two similarity scores through the learnable weight to obtain a final similarity, and realizing accurate sorting retrieval of the long text knowledge. The method gets rid of dependence on global features of sentence vectors, can accurately capture local key information in a long text, and has high retrieval precision and semantic robustness.

The invention provides a machine equipment abnormal sound detection method, a medium, equipment and a product, and relates to the technical field of abnormal sound detection, and the method comprises the steps: carrying out the data enhancement processing of an original audio during the operation of machine equipment, and dividing the original audio into a source domain sample and a target domain sample; the waveform features of the enhanced audio are converted into FFT spectrogram features, amplitude spectrogram features and logarithmic Mel spectrogram features; constructing three networks, extracting three embedded features, and then connecting; based on the splicing features of the source domain samples, obtaining a plurality of clustering centers of each category of samples in the source domain through clustering, and taking the minimum value of cosine distances between the source domain samples and all the clustering centers as the similarity score of the source domain; and based on the splicing features of the target domain samples, taking the minimum value of the cosine distance between the target domain samples and the sample mean value as the similarity score of the target domain, and taking the smaller value of the two similarity scores as an abnormal score. The method can effectively cope with the situation of domain generalization.

The invention discloses a patient report outcome management method and system, and relates to the technical field of medical information. The method comprises the following steps: firstly, dividing role permissions, and defining an information viewing and operating range; key operation behaviors are collected and coded according to a unified format to generate a role behavior chain; extracting a content degradation factor from the behavior chain, determining a weight through a random forest algorithm, and carrying out weighted calculation on a content degradation index; a term vector is generated by using a biomedical model, and a medical compliance index is calculated through a cosine distance; content optimization indexes are calculated by fusing the two indexes, and fields needing to be optimized are marked by comparing threshold values; when the field content is the content optimization item, optimizing the original field content; and comparing data of the experimental group and the control group through A / B test, calculating to obtain an optimization effect index, and automatically replacing the optimization effect index with optimization content if the optimization effect index reaches the standard to form a final patient report. According to the optimization effect index, whether the optimized field content is replaced with the original field content or not is determined, and a final patient report is obtained.

A method for calculating domain relevance scores for responses generated by LLMs is disclosed. The method includes receiving a response generated by LLM corresponding to user query. The user query is associated with a domain. The method further includes splitting the response into a plurality of response chunks using a splitting technique. The method further includes generating a plurality of response vector embeddings based on the plurality of response chunks using at least one sentence transformer. The method further includes computing a plurality of cosine distances between the plurality of response vector embeddings and a corresponding plurality of training data vector embeddings, wherein the plurality of training data vector embeddings corresponds to domain-specific training data of the LLM. The method further includes calculating a domain relevance score corresponding to the response, based on a sum of the plurality of cosine distances and a number of the plurality of chunks.

The invention relates to the technical field of remote sensing earth observation, in particular to a satellite video multi-target tracking method and system based on space-time state propagation and progressive trajectory association, and the method comprises the steps: obtaining a target detection result and a state transition result of adjacent frames in a video sequence frame image through a target tracking model; the target detection result comprises a target detection frame position representation and a target category representation, the state transition result comprises a coordinate transition result and a foreground probability, calculating a spatial position matching degree by using an intersection-to-union ratio, and calculating an adjacent frame target appearance feature similarity by using a cosine distance; and combining the spatial position matching degree and the appearance feature similarity to calculate a target space-time matching affinity, associating each target tracking frame in each video sequence frame based on the target space-time matching affinity, and adding the tracking frames to the corresponding target trajectory fragments to generate each target tracking trajectory. The core intelligence information can be accurately and efficiently extracted from the satellite video data, and the utilization efficiency of the satellite video data is improved.

The invention provides an unmanned aerial vehicle tea tree disease detection method based on morphological perception, and belongs to the technical field of agricultural information and computer vision. The method comprises the steps of firstly collecting tea garden images and constructing a data set; then, a target detection network embedded with a differentiable morphological sensor module is constructed, and the differentiable morphological sensor module extracts multi-scale shape features by using differentiable morphological operation; an adversarial learning mechanism is introduced in training, and the distinguishing ability of the model on disease and health areas is enhanced through a discriminator; after the training is completed, mining a difficult case sample based on the cosine distance between the morphological characteristics and the disease prototype vector, and carrying out supplementary training; and finally, integrating a plurality of models with optimal performance, and generating a final detection result through weighted reasoning. The method effectively strengthens the perception capability of the model for the subtle morphological characteristics of the diseases, solves the problems of low disease detection precision and insufficient difficult sample learning in a complex tea garden background, and remarkably improves the detection accuracy and robustness.

A scenic spot garbage identification method and device based on causal features, a medium and a product relate to the field of data identification, and the method comprises the steps of obtaining a plurality of scenic spot monitoring images of a target park scenic spot, and processing the images to obtain a plurality of standardized images; extracting a garbage region and a background region to obtain a garbage sample set and a background sample set; performing image fusion to generate a garbage image training set; constructing a twin network comprising two coding branches, and setting a gradient backhaul prevention operation for a second coding branch of the twin network to obtain an asymmetric twin network; respectively inputting two images of the junk image pair to two coding branches of the asymmetric twin network, carrying out model training by taking a cosine distance of output vectors of the two coding branches as a training standard, and determining a feature extraction model based on a first coding branch after training is completed; and inputting a to-be-detected monitoring image to the feature extraction model to obtain a garbage recognition result. According to the invention, scenic spot garbage can be accurately identified.

The invention discloses a high-voltage circuit breaker state monitoring and risk prediction method, and the method comprises the steps: carrying out the dynamic calibration of a health degree model based on an electromechanical-thermoelectric-material coupling differential equation set and unscented Kalman filtering through synchronous collection of mechanical, electrical, environmental and insulation characteristic quantities; and adaptive dynamic threshold optimization is realized by combining kernel density clustering and a deep Q network, identity verification is completed by using operation fingerprint cosine distance comparison, and finally a precise risk early warning instruction is generated. According to the method, the problems of poor multi-parameter synchronism, threshold staticization, identity verification deficiency and the like in traditional monitoring are solved, the fault recognition accuracy and the early warning coverage rate are remarkably improved, and the operation and maintenance cost and the power grid power failure risk are reduced.

The invention discloses a daily load curve clustering method based on a mixed distance improved FCM algorithm, and belongs to the technical field of data processing, and the method comprises the steps: collecting daily load data of a user at a sampling frequency of 5 minutes, carrying out the corresponding data preprocessing, analyzing the fluctuation characteristics of a daily load curve, and carrying out the clustering of the daily load curve; euclidean distance representing load curve power amplitude information and cosine distance representing load curve power fluctuation direction information are constructed respectively, and then an entropy weight method is adopted to construct a hybrid similarity distance (HSD) fusing the Euclidean distance and the cosine distance as a similarity criterion between daily power load curves; replacing the Euclidean distance in the FCM algorithm with the constructed HSD, then calculating a corresponding membership degree and a clustering center, iteratively updating the membership degree and the clustering center by using a target function based on an intra-class error weighted quadratic sum, and completing daily load curve clustering according to a preset maximum iteration number; and finally, determining the optimal clustering number and the corresponding optimal daily power load curve clustering result by using the sum of intra-class errors (SAE), thereby effectively solving the prominent problems of poor effect, low efficiency and the like of the traditional power load curve clustering method.

The invention provides a multi-label group activity learning and classification method based on equidistant depth embedding, and mainly relates to a multi-label group activity identification problem. The method comprises the following steps: firstly, converting an original low-dimensional label vector into a high-dimensional equidistant embedding vector through an equidistant depth embedding network; then, taking a high-dimensional equidistant embedded vector as a new label, taking minimization of cosine similarity loss as a target, and training a group activity recognition model; and finally, calculating the cosine distance between the high-dimensional vector output by the group activity identification network and each high-dimensional vector label, and taking the high-dimensional label vector with the shortest cosine distance as a classification result. Due to the fact that the original low-dimensional label vectors and the high-dimensional label vectors have the one-to-one correspondence relation, after the output high-dimensional vectors are obtained, the final activity type can be obtained according to the output high-dimensional vectors. The problem of non-equidistant label vector and label dependency in the current group activity identification task is effectively solved.

A network response retrieval enhancement generation method fused with a large language model, an electronic device and a computer readable storage medium belong to the field of text data processing, are used for automatically generating network responses, and are technically characterized in that an inquiry appeal text and a prompt are input into the large language model, and inquiry appeals and types are output by the large language model; inputting similar cases and prompts of the appeal text into the large language model, and outputting a response normal form to the political appeal of the type in the similar cases by the large language model; a prompt is input into the large language model, the response normal form is generalized through the large language model, and a first framework suitable for appeal response is generated; the method for obtaining the similar cases of the appeal text comprises the following steps: calculating an appeal vector after keyword embedding is carried out on the appeal text; determining a target sub-case library according to a cosine distance between the appeal vector and each sub-case library center based on appeal content clustering; and determining similar cases according to cosine distances between the appeal vectors and the appeal vectors of the cases in the target sub-case library.

The invention discloses a spatial data management method and system based on artificial intelligence, and relates to the technical field of spatial data management, and the method comprises the steps: defining the number of wavelet decomposition layers, calculating a self-adaptive threshold value, screening edges of an incidence matrix, generating a sparse incidence matrix, constructing a projection matrix, and projecting a feature matrix to a feature vector space through employing a matrix multiplication method. Generating a comprehensive feature matrix; and using a logistic function to define a nonlinear stream function, forming state vectors, calculating a mean value of the state vectors, fusing the mean value with the node feature matrix to form a dynamic feature matrix, and constructing a graph convolutional network model to predict the fault probability of the nodes. According to the method, fine-grained dynamic features of spatial data are accurately mined through combination of multi-scale wavelet decomposition and adaptive threshold screening, feature energy screening and a cosine distance matrix are introduced, the discrimination ability of feature selection and dimension reduction is improved, and the spatial data are extracted by using the graph convolutional network and combining a sparse embedding mechanism. And the generalization ability of the model in node fault prediction is obviously enhanced.

The invention provides a hyper-parameter optimization method based on meta-learning and an improved genetic algorithm for improving the precision of a neural network on the basis of determining a data set and not modifying the network in a real scene, and the method comprises the steps: extracting the sample features of each sample of the data set through a data sample feature extractor; using a K-means clustering algorithm to extract a clustering center as a feature centroid to obtain data set element features, calculating element feature cosine distances in pairs, and obtaining a data set element feature similarity matrix; calculating an optimal hyper-parameter combination by using an optimization genetic algorithm added with reverse learning and an adaptive crossover mutation rate; the execution time of the algorithm is reduced through an automatic early stop strategy. According to the method, the optimal hyper-parameter of the neural network can be searched under the condition of less time overhead, and the precision of the neural network is improved.

The invention discloses a referring multi-target tracking method based on multi-modal correlation graph alignment and a multi-level cross fusion model, and relates to a computer vision technology. The method comprises the following steps: sampling and enhancing training data containing a video frame sequence and referred sentences, and extracting visual and text features; performing early fusion through a cross attention module, processing through an encoder, and initializing query by using a semantic injection mechanism; in the decoder, the self-attention and semantic clearness intensifier adjusts feature interaction, and multi-level fusion is carried out on output of each layer; a correlation graph is constructed by adopting cosine distance, triple loss learning is carried out after samples are screened, and a fine-grained corresponding relation between region-level vision and word features is established; and finally, performing branch prediction and optimization. A precise relation is captured through a multi-modal correlation graph alignment model, semantic understanding is enhanced by using a multi-stage cross fusion model, and the adaptive learning ability of the model is improved. Experiments show that multiple indexes are remarkably improved, positioning and tracking are more accurate, and robustness is higher.

The invention provides an unsupervised learning tablet defect detection method based on a diffusion model and a related device, and belongs to the technical field of tablet defect detection. The method comprises the following steps: extracting semantic information of an original tablet image by using an auxiliary reconstruction network, and reconstructing an obtained potential space vector by using the extracted semantic information of the original tablet image to assist a diffusion model denoising network to obtain a reconstructed tablet image; performing feature extraction on the original tablet image and the reconstructed tablet image to obtain original tablet image features and reconstructed tablet image features; calculating Euclidean distances and cosine distances of different scales by using the original tablet image features and the reconstructed tablet image features; weighting the cosine distance and the Euclidean distance of different scales to obtain an abnormal score; and detecting the tablet defect according to the abnormal score to obtain a tablet defect detection result. According to the method, the problems of poor reconstruction effect and low detection accuracy are solved.

The invention relates to an arc fault monitoring method, system, medium and device for an aircraft direct current electrical system, and the method comprises the steps: processing filtered current data based on a CUSUM change point detection algorithm, and obtaining a current sudden change moment used for determining a sudden change point of current; calculating a current spectrum cosine distance between the current data of the current time window and the current data of the time window before the preset moment; the characteristic value ratio of the cosine distance to the historical value is calculated, and the characteristic value ratio of the characteristic value of the latest time window obtained along with time flow to the historical average value of the characteristic values of multiple continuous time windows before the preset moment is calculated; and when a sudden change point is detected, comparing the sudden change point of the frequency spectrum cosine distance of the plurality of time windows behind the sudden change point with a preset threshold value so as to judge that an arc fault occurs.

The invention relates to a gearbox fault diagnosis model construction method based on metric guide graph comparative learning, and belongs to the field of gearbox fault diagnosis model construction. The method comprises the following four core stages: firstly, carrying out frequency domain conversion and normalization on vibration signals of the gearbox to generate a node characteristic matrix; secondly, a cosine distance and an Euclidean distance are fused to construct a mixed distance matrix, and a fault diagnosis graph is generated based on a K-nearest neighbor algorithm; then unsupervised graph comparison pre-training is realized through graph data enhancement and a dynamic graph attention network (DGAT); and finally, weak supervision fine tuning is carried out by using a small number of marked samples to complete construction of the gearbox fault diagnosis model. According to the method, the construction of the high-precision gearbox fault diagnosis model can be realized in a scene with extremely few marked samples (1-10 samples per class), and the method is suitable for planetary gearbox health monitoring in the fields of wind turbines, helicopters, hybrid electric vehicles and the like.

The invention discloses an intelligent substation virtual terminal automatic connection method and system based on an SCHO-GO algorithm, and the method comprises the steps: obtaining a comprehensive distance according to an editing distance and a cosine distance represented by a virtual connection vector, and constructing a virtual terminal matching model according to the comprehensive distance; optimizing a distance weight vector in the virtual terminal matching model according to a preset SCHO-GO algorithm to obtain an optimal target distance weight vector, and updating the virtual terminal matching model according to the target distance weight vector to obtain a target virtual terminal matching model; and inputting IED data to be matched into the target virtual terminal matching model, outputting by the target virtual terminal matching model to obtain a comprehensive similarity between the input virtual terminal and the output virtual terminal, and performing matching connection on the input virtual terminal and the output virtual terminal by adopting a preset matching rule according to the comprehensive similarity. The efficiency and the accuracy of automatic connection of the virtual terminal are remarkably improved, and the method is suitable for secondary design of the intelligent substation in a complex equipment naming scene.

The invention relates to the technical field of construction management, in particular to an urban and rural construction data analysis optimization method and system, and the method comprises the following steps: obtaining remote sensing images, analyzing the change characteristics of land parcels, calculating the difference sorting of block land parcels, constructing a functional path chain, analyzing the coupling behaviors of the land parcels, and recognizing abrupt change and coupling double-height regions. And generating an urban and rural construction data analysis result. According to the method, the green vegetation, the building density and the road linear feature of the land parcel are quantified through the remote sensing image data, the frequency and the duration of the mutation feature are recorded, the mutation vector sequence is formed, the sensitivity capture of the dynamic change of the land parcel can be realized in the initial stage, and the excessive dependence on the original data is avoided. And in combination with vector cosine distance comparison between plots in the block, a difference intensity sequence is established, accurate sequencing of spatial disturbance conditions in the region is realized, and the scientificity of plot screening is enhanced.

The application provides a ViT luggage re-identification method based on local data enhancement, aiming to improve the identification accuracy in automatic luggage sorting. Traditional RFID technology is high in cost and susceptible to interference, so the introduction of computer vision technology has become a new trend. The method comprises the following steps: first, a dataset containing 1500 different identity luggage is constructed, each luggage has 10 photos of different angles, and LabelMe and YOLOv5 are used for labeling and detection; second, multiple cameras are arranged on the luggage check-in and sorting pipeline to collect luggage images; then, based on the ViT model, the local data enhancement technology and the position excitation module are applied to improve the attention of the model to the structural features of the luggage and the generalization ability; then, the luggage features are calculated through the multi-layer visual conversion module, and the cosine distance is used for identity comparison; finally, the identity loss, the triplet loss and the contrast loss are used to optimize the model to improve the identification accuracy.

The present invention discloses a method, device and medium for backdoor defense of image recognition based on swarm intelligence. Based on the association relationship of multiple neural network image recognition models in a distributed cluster and ranking according to the degree of connection density, a densely connected neural network model is selected as a seed neural network model; for each two seed neural network models, an attention distillation representation is used to perform a distillation operation on the two seed neural network models to extract an attention activation map that is invalid for backdoor attacks on the seed neural network models. A l ; By designing a robust distillation loss function L Measuring the activation attention maps of two seed neural network models A l The Euclidean distance and cosine distance of the seed neural network model are based on L Calculate the gradient value and perform back propagation to update the parameters of the seed neural network model. The performance and efficiency of the swarm intelligence image recognition backdoor defense algorithm proposed in the present invention under the image recognition backdoor defense task have reached the current optimal level.

The disclosed technology relates to methods, accent conversion systems, and non-transitory computer readable media for real-time accent conversion. In some examples, a set of phonetic embedding vectors is obtained for phonetic content representing a source accent and obtained from input audio data. A trained machine learning model is applied to the set of phonetic embedding vectors to generate a set of transformed phonetic embedding vectors corresponding to phonetic characteristics of speech data in a target accent. An alignment is determined by maximizing a cosine distance between the set of phonetic embedding vectors and the set of transformed phonetic embedding vectors. The speech data is then aligned to the phonetic content based on the determined alignment to generate output audio data representing the target accent. The disclosed technology transforms phonetic characteristics of a source accent to match the target accent more closely for efficient and seamless accent conversion in real-time applications.