In a learning device, a feature extraction means extracts image features from an input image. A class discrimination means discriminate a class of the input image based on the image features, and generates a class discriminative result. A class discriminative loss calculation means calculates a class discriminative loss based on the class discriminative result. A normal/abnormal discrimination means discriminates whether the class is a normal class or an abnormal class, based on the image features, and generates a normal/abnormal discriminative result. The AUC loss calculation means calculates an AUC loss based on the normal/abnormal result. A first learning means updates parameters of the feature extraction means, a class discrimination means, and the normal/abnormal discrimination means, based on the class discriminative loss and the AUC loss.

In a learning device, a feature extraction means extracts image features from an input image. A class discrimination means discriminate a class of the input image based on the image features, and generates a class discriminative result. A class discriminative loss calculation means calculates a class discriminative loss based on the class discriminative result. A normal/abnormal discrimination means discriminates whether the class is a normal class or an abnormal class, based on the image features, and generates a normal/abnormal discriminative result. The AUC loss calculation means calculates an AUC loss based on the normal/abnormal result. A first learning means updates parameters of the feature extraction means, a class discrimination means, and the normal/abnormal discrimination means, based on the class discriminative loss and the AUC loss.

A facial structure estimating device 10 includes an acquiring unit 11 and a controller 13. The acquiring unit 11 acquires a facial image. The controller 13 functions as an estimator 16 that estimates a facial structure from a facial image. The controller 13 tracks a starting feature point constituting a facial structure using a tracking algorithm in a facial image of a frame subsequent to a facial image used to estimate the facial structure. The controller 13 obtains a resulting feature point by tracking a tracked feature point using an algorithm in an original frame facial image. The controller 13 selects a learning facial image for which the interval between resulting and starting feature points is less than or equal to a threshold. The controller 13 trains the estimator using the facial image selected for learning and the facial structure estimated by the estimator 16 based on the facial image.

A facial structure estimating device 10 includes an acquiring unit 11 and a controller 13. The acquiring unit 11 acquires a facial image. The controller 13 functions as an estimator 16 that estimates a facial structure from a facial image. The controller 13 tracks a starting feature point constituting a facial structure using a tracking algorithm in a facial image of a frame subsequent to a facial image used to estimate the facial structure. The controller 13 obtains a resulting feature point by tracking a tracked feature point using an algorithm in an original frame facial image. The controller 13 selects a learning facial image for which the interval between resulting and starting feature points is less than or equal to a threshold. The controller 13 trains the estimator using the facial image selected for learning and the facial structure estimated by the estimator 16 based on the facial image.

This application discloses a neural network model training method, an image processing method, and an apparatus in the field of artificial intelligence. The method includes: inputting training data to a neural network model for feature extraction, and obtaining a first weight gradient of the neural network model based on an extracted feature; obtaining a candidate weight parameter, where a partial derivative of a function value of a target loss function to the candidate weight parameter is 0, the function value of the target loss function is determined based on a function value of a second loss function corresponding to a first prediction label, and the function value of the second loss function corresponding to the first prediction label indicates a difference between the candidate weight parameter and a weight parameter of the neural network model and a difference between a weight variation and the first weight gradient.

This application discloses a neural network model training method, an image processing method, and an apparatus in the field of artificial intelligence. The method includes: inputting training data to a neural network model for feature extraction, and obtaining a first weight gradient of the neural network model based on an extracted feature; obtaining a candidate weight parameter, where a partial derivative of a function value of a target loss function to the candidate weight parameter is 0, the function value of the target loss function is determined based on a function value of a second loss function corresponding to a first prediction label, and the function value of the second loss function corresponding to the first prediction label indicates a difference between the candidate weight parameter and a weight parameter of the neural network model and a difference between a weight variation and the first weight gradient.

A detection method includes obtaining a to-be-migrated model. The to-be-migrated model includes a memory feature set, and the memory feature set represents a feature vector set associated with an application scene corresponding to the to-be-migrated model. The method further includes performing a metric calculation on at least one piece of sample data of a target scene and the memory feature set to obtain at least one metric calculation result, and updating the memory feature set according to the at least one metric calculation result to obtain a target memory feature set. The target memory feature set represents a feature vector set associated with the target scene. The method further includes obtaining a target detection model by replacing the memory feature set of the to-be-migrated model with the target memory feature set.

A system and method for for determining a degree of point cloud data degradation of a LiDAR sensor of a Self-Driving Car (SDC) using a machine-learning algorithm (MLA) are provided. The method comprises: determining, based on a training point cloud generated by the LiDAR sensor representative of surroundings of the SDC, a plurality of LiDAR features; determining, for each training object in the surroundings, based on statistical data of coverage of training objects with LiDAR points, a plurality of enrichment features; receiving a respective label indicative of a degradation degree of the training point cloud; generating, based on the plurality of LiDAR features, the plurality of enrichment features, and the respective label, a given feature vector of a plurality of feature vectors; training, based on the plurality of feature vectors, the MLA to determine an in-use degree of degradation of in-use sensed data further generated by the LiDAR sensor.

A method of updating a trained cardinality estimation model includes receiving a cardinality estimation model with cardinality labels and detecting a drift in underlying data or predicates of the cardinality estimation model. The type of the detected drift is determined and new test queries that mimic test queries for the detected drift are synthesized. A portion of the synthesized test queries is selected to reduce annotation cost and used to update the cardinality estimation model.

A method of updating a trained cardinality estimation model includes receiving a cardinality estimation model with cardinality labels and detecting a drift in underlying data or predicates of the cardinality estimation model. The type of the detected drift is determined and new test queries that mimic test queries for the detected drift are synthesized. A portion of the synthesized test queries is selected to reduce annotation cost and used to update the cardinality estimation model.