122 results about "Continual learning" patented technology

Classification of an input task data set by meta level continual learning includes analyzing first and second training data sets in a task space to generate first and second meta weights and a slow weight value, and comparing an input task data set to the slow weight to generate a fast weight. The first and second meta weights are parameterized with the fast weight value to update the slow weight value, whereby a value is associated with the input task data set, thereby classifying the input task data set by meta level continual learning.

The invention discloses a machine learning-based SQL injection detection method, and a database security system, and belongs to the technical field of network security. The machine learning-based SQLinjection detection method comprises the steps of extracting parameters from HTTP requests, generating a grammar tree of a sample through lexical analysis and grammatical analysis, extracting featuresof the grammar tree and a URL, and performing training by adopting a machine learning algorithm of a support vector machine; and deploying a trained classification model between a Web service and a client, classifying the HTTP requests in a production environment, when it is judged that the HTTP requests comprise SQL injection attacks, giving a warning and blocking the requests, and finally storing the requests in an SQL injection attack sample library. According to the method, the dependency on background information is low, so that the HTTP requests received by the Web service only need tobe obtained; the deployment difficulty is low, so that the classification model can be deployed between a Web server and the client to serve as a flow filter; the method has high accuracy; the methodhas a continuous learning capability; and the method has high expansibility.

The invention provides a data processing method. The method comprises the steps of inputting historical learning behavior data of different users into a neural network model to obtain a historical learning path of each user; generating a learning attribute label for each user according to the personal learning data of different users; based on the learning attribute label, screening out at least one similar user of the target user by using a collaborative filtering algorithm, and processing the historical learning path of the target user and the historical learning path of the at least one similar user to form a quasi-learning path of the target user; and performing fusion processing on the planned learning data of the target user and the quasi-learning path of the target user to obtain arecommended learning path of the target user, the recommended learning path comprising a plurality of learning tasks. According to the scheme, a set of perfect learning path is provided for each target user in a personalized recommendation mode, the learning habit of continuous learning of the target user can be cultivated, and then the class completion rate of the target user can be increased.

Continuously learning and optimizing artificial intelligence (AI) adaptive neural network (ANN) computer modeling methods and systems, designated human affect computer modeling (HACM) or affective neuron (AN), and, more particularly, to AI methods, systems and devices that can recognize, interpret, process and simulate human reactions and affects such as emotional responses to internal and external sensory stimuli, that provides real-time reinforcement learning modeling that reproduces human affects and / or reactions, wherein the human affect modeling (HACM) can be used singularly or collectively to modeling and predict complex human reactions and affects.

A method for auto-labeling a training image to be used for learning a neural network is provided for achieving high precision. The method includes steps of: an auto-labeling device (a) instructing a meta ROI detection network to generate a feature map and to acquire n current meta ROIs, on the specific training image, grouped according to each of locations of each of the objects; and (b) generating n manipulated images by cropping regions, corresponding to the n current meta ROIs, on the specific training image, instructing an object detection network to output each of n labeled manipulated images having each of bounding boxes for each of the n manipulated images, and generating a labeled specific training image by merging the n labeled manipulated images. The method can be performed by using an online learning, a continual learning, a hyperparameter learning, and a reinforcement learning with policy gradient algorithms.

An Online Machine Learning System (OMLS) including an Online Preprocessing Engine (OPrE) configured to (a) receive streaming data including an instance comprising a vector of inputs, the vector of inputs comprising a plurality of continuous or categorical features; (b) discretize features; (c) impute missing feature values; (d) normalize features; and (e) detect drift or change in features; an Online Feature Engineering Engine (OFEE) configured to produce features; and an Online Robust Feature Selection Engine (ORFSE) configured to evaluate and select features; an Online Machine Learning Engine (OMLE) configured to incorporate and utilize one or more machine learning algorithms or models utilizing features to generate a result, and capable of incorporating and utilizing multiple different machine learning algorithms or models, wherein each of the OMLE, the OPrE, the OFEE, and the ORFSE are continuously communicatively coupled to each other, and wherein the OMLS is configured to perform continuous online machine learning.

The invention provides a video event detection method of continuous learning. The detection method comprises an initial learning stage and an incremental learning stage, wherein in the initial learning stage, video data with a tag is prepared, sparse self coding is used for learning the video data with the tag, and a priori model is trained; in the incremental learning stage, the priori model which is trained is used for classifying new video data, a probability score and a gradient distance are calculated, and automatic tag adding or manual tag adding is selected for the new video data by active learning according to a calculation result. The video event detection method is combined with deep learning and the active learning to automatically select the most proper characteristics and utilize video streaming data to gradually improve a traditional model; when the new video data is obtained, unsupervised learning is used for extracting characteristics, then, the active learning is used for reducing the work of manual classification as much as possible, the model is perfected gradually, and finally, a purpose of continuous learning is achieved.

The invention relates to the technical field of form generation, and discloses an HS code matching method and system based on intelligent analysis and recognition, an HS code display method and systembased on intelligent analysis and recognition and a storage medium, and the method comprises the steps: obtaining a to-be-judged object; correcting an imaging problem; detecting a text in the to-be-judged object; recognizing the text content; extracting required fields and / or elements from the text recognition result to generate to-be-judged object description information; according to the obtained description information of the to-be-judged object and pre-trained graph data, judging the category of the to-be-judged object, and performing entity linking with the graph data; training the pre-trained atlas data in combination with a semantic library to generate a model according to provided HS coded document data, and carrying out continuous learning and optimization by external data feedback and an AI algorithm. The intelligent knowledge search engine capable of meeting the customs declaration pre-classification service field is built to meet the requirement for quickly obtaining knowledge, and can accurately correspond to various columns in combination with character recognition.

The embodiment of the invention provides a continuous learning framework and a continuous learning method for a deep neural network, and the framework comprises a condition generator network which isused for generating generation data of the same category as training data of a current task, and distributing a specific parameter subspace for a current task during training; a discriminator networkwhich is used for supervising the generation process of the generated data to enable the generated data to gradually approach the training data of the old task, and taking the approximate generated data as equivalent training data of the old task; and a classifier network which comprises an independent classifier network and an auxiliary classifier network of the discriminator network, and is usedfor selectively keeping parameters of the coded old task by using a weight consolidation mechanism, and continuously updating and jointly training the current task by using the training data of the current task and the equivalent training data of the old task. According to the embodiment of the invention, disastrous forgetting of old tasks in the continuous learning process can be effectively relieved, and the continuous learning ability is improved.

The invention provides a visual and auditory perception integrated multitask collaborative identification method and system and belongs to the technical field of multi-source heterogeneous data processing and identification. The system comprises a universal feature extraction, a collaborative feature learning module and a suitable scene feedback, assessment and identification module. Based on timesynchronization matching mechanism of multi-source heterogeneous data, universal features of the multi-source heterogeneous data are extracted; a long-time dependent memory model is established, anduniversal features serving as priori knowledges are continuously learned by cooperating with a collaborative attention mechanism based on external dependence; environmental perception parameters in the multi-source heterogeneous data are extracted, a progressive network depth collaborative reinforcement identification mechanism is established, and multitask identification is achieved according tothe learning features and task demands of the memory model. The system combines with a suitable scene computing theory based on environmental perception, judges the weight of each identified task through depth reinforcement feedback, self-adaptively adjusts the priority of each task according to environmental change and achieves the effect of simultaneously outputting multiple visual and auditoryperception identification results.

The invention provides a data processing method and apparatus in a graph neural network, an electronic device and a storage medium. The method comprises the steps of obtaining an initial structure ofa first graph neural network and to-be-processed data in a data stream; triggering a decoupling process and decoupling the relational triples in the first graph neural network into a plurality of embedded components; determining graph embedding corresponding to different embedding components based on the different embedding components matched with the relational triples in the first graph neural network; triggering an updating process to update the relational triples in the initial structure of the first graph neural network to form a second graph neural network; therefore, dynamic data is processed by utilizing the graph neural network, continuous learning of the graph neural network can be achieved in different data to form a new graph neural network, and knowledge acquired by the initial graph neural network is reserved, so that the richness and the perspectiveness of data processing can be improved, and the use experience of a user is improved.

An automated system and method for retaining images in a smart phone are disclosed. The system may then determine a no-reference quality score of the image using a PIQUE module. The PIQUE module utilizes block level features of the image to determine the no-reference quality score. The system may present the image and the no-reference quality score to the user and accept a feedback towards quality of the image. The system may utilize a supervised learning model for continually learning a user's perception of quality of the image, the no-reference quality score determined by the PIQUE module, and the user feedback. Based on the learning, the supervised learning model may adapt the no-reference quality score and successively the image may either be retained or isolated for deletion, based on the adapted quality score and a predefined threshold range.

The invention provides a continuous learning method and system for biomacromolecular particle recognition of a cryoelectron microscope. The method comprises the following steps: acquiring to-be-recognized cryoelectron microscope image data; and inputting the to-be-identified cryoelectron microscope image data into a trained network parameter model to obtain an identification result of the biomacromolecular particles, wherein the trained network parameter model is obtained by training a neural network according to example cryoelectron microscope sample image data and cryoelectron microscope sample image data obtained in a historical training process, and the example cryoelectron microscope sample image data and the cryoelectron microscope sample image data are labeled with biological macromolecular particle coordinate tags. The method does not need to store a large amount of old data sets, reduces the single training time, and can continuously enhance the feature recognition range and recognition precision of particle selection.

The invention discloses a space-ground integrated carrier-class satellite mobile communication system architecture and an implementation method. The system is composed of a plurality of satellites in a space section, an operation and maintenance management and control center and a plurality of gateway stations in a ground section, and a plurality of user terminals deployed on land, on the sea and in the air in a user section. A plurality of satellites in a space segment carry out splicing coverage on a ground target coverage range, a plurality of gateway stations are dispersedly deployed according to needs and support cross hot standby of key equipment, and an operation and maintenance management and control center carries out satellite-ground integrated operation maintenance and management control on the satellites, the gateway stations and user terminals; and for various complex use scenes, a network is flexibly established according to different application plan modes to provide high-quality comprehensive guarantee for users, and continuous learning updating of application plans is supported. According to the invention, resource comprehensive networking of multiple satellites is supported, and the expandability is good; mutual backup of multiple gateway station equipment levels is supported, and the reliability is high; and continuous updating of multiple application modes is supported, and the adaptability to complex scenes is good.

A method for training a deep learning network based on artificial intelligence is provided. The method includes steps of: a learning device (a) inputting unlabeled data into an active learning network to acquire sub unlabeled data and inputting the sub unlabeled data into an auto labeling network to generate new labeled data; (b) allowing a continual learning network to sample the new labeled data and existing labeled data to generate a mini-batch, and train the existing learning network using the mini-batch to acquire a trained learning network, wherein part of the mini-batch are selected by referring to specific existing losses; and (c) (i) allowing an explainable analysis network to generate insightful results on validation data and transmit the insightful results to a human engineer to transmit an analysis of the trained learning network and (ii) modifying at least one of the active learning network and the continual learning network.

The invention discloses a multi-working-condition process monitoring method with continuous learning capability and improved PCA, and relates to the field of industrial monitoring and fault diagnosis.The method comprises the following steps: sequentially collecting process data of an industrial system under normal working conditions to form a training set; training the initial working condition by utilizing principal component analysis, and calculating an initial projection matrix; constructing an optimization function according to an elastic weight consolidation method and a principal component analysis principle, and training subsequent working conditions to obtain an optimal projection matrix; constructing monitoring statistics and calculating a threshold value; collecting process dataunder the real-time working condition of the system as a test sample, calculating statistics of the sample by utilizing the current training model, comparing the statistics with a threshold value, and judging whether a fault occurs or not. The weight matrix is determined by combining the system principle and priori knowledge, the interpretability of the method is improved, the algorithm is simple, the calculated amount is small, implementation is easy, and the method can be widely applied to the fields of chemical engineering, processing and manufacturing, large thermal power plants and the like.

The embodiment of the invention relates to an intelligent device, and discloses a learning control method and a tutoring device based on the tutoring device. The method comprises the following steps:when the current working state of the tutoring device is a learning state, the tutoring device acquires the continuous learning time length in the learning state; if the duration of continuous learning is longer than or equal to the first preset duration, the tutor switches the current working state from the learning state to the leisure state, and outputs any knowledge points corresponding to anypreset extracurricular knowledge category and the detailed contents corresponding to any knowledge points for the user to learn; the tutoring equipment detects the duration of leisure time in the leisure state; if the continuous leisure time is longer than or equal to the second preset time, the tutor device switches the current working state from the leisure state to the learning state. By implementing the embodiment of the invention, the learning efficiency of the user can be improved.

A method for performing continual learning on a classifier, in a client, capable of classifying images by using a continual learning server is provided. The method includes steps of: a continual learning server (a) inputting first hard images from a first classifier of a client into an Adversarial Autoencoder, to allow an encoder to output latent vectors from the first hard images, allow a decoder to output reconstructed images from the latent vectors, and allow a discriminator and a second classifier to output attribute and classification information to determine second hard images to be stored in a first training data set, and generating augmented images to be stored in a second training data set by adjusting the latent vectors of the reconstructed images determined not as the second hard images; (b) continual learning a third classifier corresponding to the first classifier; and (c) transmitting updated parameters to the client.

The invention relates to a continuous learning method of a sequence recommendation model based on sample playback, which relates to the technical field of sequence recommendation, and comprises the following steps: 1, constructing a sequence recommendation model, and training the sequence recommendation model by using initial data; 2, sampling a small part of representative example samples based on a sample selection strategy of article category balance; 3, calculating and storing soft labels of the sampled example samples so as to participate in calculation of a distillation loss function part in next model updating; 4, providing an accurate recommendation service for a user by using the sequence recommendation model, and collecting new data obtained in a new period at the same time; and 5, updating parameters of the sequence recommendation model by using new data obtained in a new period and previously stored sampled examples The problem of disastrous forgetting faced by using a neural network sequence recommendation model in a continuous learning scene is effectively solved.

The invention provides a preprandial insulin dosage learning optimization decision-making system assisted by expert experience. According to the system, artificial intelligence and expert experience methods are combined, information contained in patient blood sugar monitoring and insulin infusion data is mined at the same time, a safe and effective preprandial insulin dosage can still be determined based on historical data under the condition of few samples, the postprandial blood sugar management is improved, and meanwhile, the system is endowed with abilities of continuous adaptive learning and decision-making performance improvement. Therefore, in order to improve postprandial blood glucose management and utilize a small amount of patient historical data, an expert decision-making assisted preprandial insulin dose individualized learning decision-making system is designed, a model prediction evaluation method is introduced into the system, system decision errors under the condition of few samples are effectively avoided, and the blood glucose metabolism rule of the patient is continuously learned; and meanwhile, an iterative updating thought is further introduced, a postprandial blood sugar management target is determined in a self-adaptive mode, and safe and effective preprandial insulin doses can be rapidly determined under the condition that few samples are used for diabetic patients under different illness conditions.

The invention discloses a power grid dispatching knowledge graph data optimization method and system. The method comprises the following steps: firstly carrying out automatic mining on high-quality phrases of a field through a deep learning method, and completing automatic recognition and equivalent disambiguation of a dispatched entity; then, completing dispatched entity global relationship extraction according to deep learning technology, so as to complete entity relationship recognition and verification, and achieve the purpose of establishing an initial power grid dispatching knowledge graph; on the basis that the two steps are completed, using a natural language learning knowledge fusion technology, and conducting incremental training on newly-added scheduling plan data based on timestamps; meanwhile, introducing life cycle management of knowledge graph knowledge content in the completion process of the steps; finally, completing a continuous learning dynamic knowledge graph under the cooperation of the steps. According to the invention, the high precision of a power grid dispatching optimization decision knowledge graph is ensured, and the consumption of computing resources and time during updating training is reduced while dynamic updating of incremental knowledge is ensured.

The invention discloses an online learning method and system for an unmanned helicopter model, and belongs to the technical field of unmanned aerial vehicles, and the method comprises the steps: building an unmanned helicopter model with uncertain items according to a fixed world reference coordinate system and a body-following coordinate system built through taking the mass center of an unmannedhelicopter as an original point; training the two residual convolution sub-networks by using historical flight data of the unmanned helicopter to obtain two trained sub-networks; constructing a deep continuous learning network model by using the two trained sub-networks and a dynamic Hadamard product layer; and learning uncertain items in the unmanned helicopter model by using the deep continuouslearning network model to realize online dynamic learning of the unmanned helicopter model. According to the method, the historical data flow online dynamic learning model of the unmanned helicopter can be utilized to realize online multi-step prediction of various stunt action states, and the method has good generalization ability and good robustness.

The present disclosure relates to a method and device for providing questions for learning. According to the present disclosure, a method by which a server provides questions for question includes step a of displaying learning content on a first region of a screen displayed on a terminal of a learner; step b of fixedly displaying an AI object on a second region of the screen; step c of determining whether a learning progress of the learner satisfies a predetermined object activation condition; and step d of displaying an AI message corresponding to the object activation condition on a third region of the screen adjacent to the AI object while the AI object is activated, when the learning progress satisfies the object activation condition. According to the present disclosure, a user interface may be adaptively changed according to a question solving situation of a learner, thereby inducing continuous learning of a user.

The invention provides a data processing method. The method comprises the steps of according to plan learning data of a target user, subjecting a universal learning path containing standard knowledgepoints to the personalized processing to determine a recommended learning path of the target user, wherein the universal learning path comprises a plurality of learning tasks; obtaining an actual learning path of the target user, and comparing the recommended learning path with the actual learning path to obtain a comparison result; and determining a learning supervision strategy for the target user according to the comparison result, wherein the learning supervision strategy is used for promoting the target user to continuously learn the recommended learning task of the recommended learning path. According to the scheme, the personalized recommendation learning path is provided for each target user according to the plan learning data of the target user, the learning habit of continuous learning of the target user can be cultivated, and thus the course completion rate of the target user can be increased. Moreover, the recommended learning path is timely and correspondingly adjusted based on the actual learning condition, so that the course completion rate of the target user can be further improved.

A method for on-device continual learning of a neural network which analyzes input data is provided for smartphones, drones, vessels, or a military purpose. The method includes steps of: a learning device, (a) uniform-sampling new data to have a first volume, instructing a boosting network to convert a k-dimension random vector into a k-dimension modified vector, instructing an original data generator network to repeat outputting synthetic previous data of a second volume corresponding to the k-dimension modified vector and previous data having been used for learning, and generating a batch for a current-learning; and (b) instructing the neural network to generate output information corresponding to the batch. The method can be used for preventing catastrophic forgetting and an invasion of privacy, and for optimizing resources such as storage and sampling processes for training images. Further the method can be performed through a learning for Generative adversarial networks (GANs).

A method for calibrating a pitch of a camera on a vehicle is provided. The method includes steps of: a calibration device (a) inputting driving images from the camera into an object detection network to detect objects and generate object detection information and into a lane detection network to detect lanes and generate lane detection information; (b) profiling the object and the lane detection information to generate object profiling information and lane profiling information, inputting the object profiling information into an object-based pitch estimation module to select a first target object and a second target object to generate a first pitch and a second pitch, and (iii) inputting vanishing point detection information and the lane profiling information into a lane-based pitch estimation module to generate a third pitch and a fourth pitch; and (c) inputting the first to the fourth pitches into a pitch-deciding module to generate a decided pitch.

The invention discloses a distributed near-end strategy optimization method based on cognitive behavior knowledge and application thereof, and the method comprises the following steps: employing cognitive behavior knowledge to establish a cognitive behavior model of an Agent, introducing the cognitive behavior model into deep reinforcement learning, and constructing a deep reinforcement learning framework based on cognitive behavior knowledge; based on the deep reinforcement learning framework, providing a distributed near-end strategy optimization algorithm based on cognitive behavior knowledge; and quantitatively designing a guide mode of the cognitive behavior model for Agent strategy updating, and achieving continuous learning of the Agent on the basis of cognitive behavior knowledge. According to the method provided by the invention, cognitive behavior knowledge can be effectively utilized, and strategy updating is carried out on the basis, so that the learning efficiency of the Agent is improved.

Systems and methods are described for automatically identifying an anatomical landmark in a medical image according to local preferences associated with a particular clinical site. A medical image for performing a medical procedure is received. An anatomical landmark is identified in the medical image using a pre-trained machine learning algorithm. Feedback relating to the identified anatomical landmark is received from a user associated with a particular clinical site. The feedback is received during a normal workflow for performing the medical procedure. The pre-trained machine learning algorithm is retrained based on the received feedback such that the retrained machine learning algorithm is trained according to local preferences associated with the particular clinical site.

The invention provides an intrusion detection system rule matching optimization method based on machine learning, and the method comprises the steps: continuously learning the historical alarm of an intrusion detection system through a machine learning period construction module, and periodically constructing a machine learning data training set; enabling the machine learning prediction module to perform real-time prediction on the selected pre-matching rule sequence after the network message enters the second detection engine stage, and outputting the hit probability of the pre-matching rule base sequence; and performing inverse sorting on the pre-matching rules according to the hit probabilities, so that the intrusion detection system preferentially traverses the pre-matching rule with the highest hit probability. The method can effectively reduce the number of invalid matching times of the intrusion detection system, improve the performance of the intrusion detection system, dynamically adjust the pre-matching rule sequence of the intrusion detection system, and ensure the stability of the efficiency of the intrusion detection system in different flow scenes. By regularly updating the machine learning data set, the stability of the intrusion detection system in a multi-flow scene can be effectively improved.