Abstract: Techniques are disclosed relating to classifying transactions using post-transaction information. Training architecture may be used to train a first classifier module using first data for a set of transactions as training data input, where the first data includes both pre-transaction information and post-transaction information for transactions in the set of transactions. During training of the first classifier module, in disclosed techniques, correct classifications for the transaction in the set of transactions are known. The training architecture, in disclosed techniques, generates respective weights for multiple transactions in the set of transactions based on classification outputs of the trained first classifier for the multiple transactions. In disclosed techniques, the training architecture trains a second classifier module, based on the generated weights, using second data for the set of transactions as training data input.

Abstract: Embodiments of this application disclose an artificial intelligence-based (AI-based) animation character control method. When one animation character has a corresponding face customization base, and one animation character has no corresponding face customization base, the animation character having the face customization base may be used as a driving character, and the animation character having no face customization base may be used as a driven character.

Abstract: Techniques are disclosed relating to improving machine learning classification using both labeled and unlabeled data, including electronic transactions. A computing system may train a machine learning module using a first set of transactions (of any classifiable data) with label information that indicates designated classifications for those transactions and a second set of transactions without label information. This can allow for improved classification error rates, particularly when additional labeled data may not be present (e.g., if a transaction was disallowed, it may not be later labeled as fraudulent or not). The training process may include generating first error data based on classification results for the first set of transactions, generating second error data based on reconstruction results for both the first and second sets of transactions, and updating the machine learning module based on the first and second error data.

Abstract: This application disclose an artificial intelligence (AI) based animation character drive method. A first expression base of a first animation character corresponding to a speaker is determined by acquiring media data including a facial expression change when the speaker says a speech, and the first expression base may reflect different expressions of the first animation character. After target text information is obtained, an acoustic feature and a target expression parameter corresponding to the target text information are determined according to the target text information, the foregoing acquired media data, and the first expression base. A second animation character having a second expression base may be driven according to the acoustic feature and the target expression parameter, so that the second animation character may simulate the speaker's sound and facial expression when saying the target text information, thereby improving experience of interaction between the user and the animation character.

Abstract: Methods and systems for efficiently compressing information comprising a plurality of data points along a particular dimension are presented. In some embodiments, a model may be generated using a semi-parametric modeling technique or a non-parametric modeling technique to represent the plurality of data points. The model may include a set of parameters that is less in size than the plurality of data points. Once the model is generated, the set of parameters may be stored and subsequently used to represent the information, with a significant reduction in storage space over the original data. In response to a request to analyze the information, the set of parameters may be analyzed to produce an outcome. Since the set of parameters have less cardinality than the plurality of data points in the original information, the efficiency of the analysis tool is enhanced.

Abstract: Embodiments of this application provide an object tracking method performed by a computer device. The method includes, when a target object is lost in a second image frame in a first subsequent image frames, determining, according to a first local feature and in second subsequent image frames starting with the second image frame, a third image frame in which the target object reappears after the target object is lost during the tracking; determining a location of a target object region in the third image frame including the target object; and continuing to track the target object in image frames according to the location of the target object region in the third image frame. Through the object tracking method, a lost object can be detected and repositioned by using an extracted first local feature of the target object, thereby effectively resolving the problem in the existing technical solution.

Abstract: This application provides a method for training a pose recognition model performed at a computer device. The method includes: inputting a sample image labeled with human body key points into a feature map model included in a pose recognition model, to output a feature map of the sample image; inputting the feature map into a two-dimensional (2D) model included in the pose recognition model, to output 2D key point parameters used for representing a 2D human body pose; input a target human body feature map cropped from the feature map and the 2D key point parameter into a three-dimensional (3D) model included in the pose recognition model, to output 3D pose parameters used for representing a 3D human body pose; constructing a target loss function based on the 2D key point parameters and the 3D pose parameters; and updating the pose recognition model based on the target loss function.

Abstract: This application discloses a method for training a gaze tracking model, including: obtaining a training sample set; processing the eye sample images in the training sample set by using an initial gaze tracking model to obtain a predicted gaze vector of each eye sample image; determining a model loss according to a cosine distance between the predicted gaze vector and the labeled gaze vector for each eye sample image; and iteratively adjusting one or more reference parameters of the initial gaze tracking model until the model loss meets a convergence condition, to obtain a target gaze tracking model. According to the solution provided in this application, a gaze tracking procedure is simplified, a difference between a predicted value and a labeled value can be better represented by using the cosine distance as a model loss to train a model, to improve prediction accuracy of the gaze tracking model.

Abstract: An electronic device obtains an image that includes a face. The device performs feature extraction on the image, to obtain facial expression information corresponding to the face and facial feature information corresponding to the facial expression, wherein the facial feature information indicates an extent of the facial expression. The device determines facial emotion information according to the facial expression information. The device also determines facial feature expression information according to a target feature value corresponding to the facial emotion and the facial feature information. This expression recognition techniques disclosed herein can implement multi-task learning and reduce an amount of data required for model training, and can obtain both an emotion recognition result and a local expression recognition result, thereby improving efficiency and real-time performance of expression recognition and improving user experience.

Abstract: A video editing method, apparatus and storage medium are provided. The method includes obtaining an object, the object including one or more images; determining a content element of the object for video editing, the content element having a content type identifier; determining a material set identifier corresponding to the content type identifier according to a first behavior tree logic; determining a video editing material set corresponding to the material set identifier; and obtaining an edited video according to the content element and the video editing material set.

Abstract: A face key point detection method includes determining, according to an image containing a face, a multi-channel feature map of the image; converting the multi-channel feature map into a predetermined channel quantity of target feature maps; performing a convolution operation on each target feature map in the predetermined channel quantity of target feature maps by using a convolution kernel corresponding to each target feature map; generating a feature vector corresponding to the image based on a result of the convolution operation on each target feature map; and determining key point coordinates of the face on the image according to the feature vector.

Abstract: A risk analysis system configures the decision engine to detect anomalous online activities by analyzing usage patterns associated with one or more user accounts across multiple frequencies. The risk analysis system obtains transaction log data representing transactions associated with one or more accounts, and extracts data from the transaction log data to generate time-series data along a time dimension. The time-series data may represent usage characteristics of one or more user accounts over a period of time. The risk analysis system derives pattern data representing usage patterns across multiple different frequencies based on the time-series data. The risk analysis system then configures the decision engine to detect anomalous account activities based on the derived pattern data.

Abstract: A computer system creates a plurality of indexes from a first plurality of records, wherein each index corresponds to an attribute of a plurality of attributes. The computer system detects a record of a second plurality of records, wherein the record includes a value corresponding to each of the plurality of attributes. The computer system determines a first set of values from a first index of the plurality of indexes that corresponds to a first attribute. The computer system determines a plurality of individual similarity scores for the first set of values by utilizing a similarity function. The computer system determines an overall similarity score for each record of at least a portion of the first plurality of records and based on the overall similarity scores, determines a record of the first plurality of records that corresponds to the record of the second plurality of records.

Abstract: A system performs operations that include receiving a request to process a current payment transaction between a payment provider and a user having a user account with the payment provider. The operations further include determining a state of a recurrent neural network (RNN) fraud model for the user account by accessing a cache storing a set of encoded states for a plurality of nodes included in the RNN fraud model. The RNN fraud model is executed based on data associated with the current payment transaction and the set of encoded states stored in the cache. A new set of encoded states for the plurality of nodes are encoded and stored in place of the set of encoded state previous stored in the cache.

Abstract: A computer system is configured to receive a dataset that includes a plurality of transaction request records and is divisible into a plurality of segments. Each transaction request record includes an individual score calculated by a machine learning algorithm. The computer system also receives a plurality of constraints. The computer system is configured to calculate, using a linear programming algorithm, a decision threshold score for a particular segment of the plurality of segments using the transaction request records. The computer system is configured to provides access to the calculated decision threshold score to a production computer system. The production computer system is configured to use the decision threshold score to evaluate a subsequent transaction request corresponding to the particular segment.

Abstract: A risk analysis system configures the decision engine to detect anomalous online activities by analyzing usage patterns associated with one or more user accounts across multiple frequencies. The risk analysis system obtains transaction log data representing transactions associated with one or more accounts, and extracts data from the transaction log data to generate time-series data along a time dimension. The time-series data may represent usage characteristics of one or more user accounts over a period of time. The risk analysis system derives pattern data representing usage patterns across multiple different frequencies based on the time-series data. The risk analysis system then configures the decision engine to detect anomalous account activities based on the derived pattern data.

Abstract: An object detection training method can include receiving a training sample set in a current iteration of an object detection training process over an object detection neural network. The training sample set can include first samples of a first class and second samples of a second class. A first center loss value of each of the first and second samples can be determined. The first center loss value can be a distance between a feature vector of the respective sample and a center feature vector of the first or second class which the respective sample belongs to. A second center loss value of the training sample set can be determined according to the first center loss values of the first and second samples. A first target loss value of the current iteration can be determined according to the second center loss value of the training sample set.

Abstract: A new hammering system with electromagnetic power for dynamic pile testing. The basic working principle of the hammering system is as follows: when an internal coil is energized, a magnetic force is generated to attract tightly, via a magnetic conduction panel, an adaptive weight hammer disposed in contact with the surface of the panel; when the internal coil is de-energized, demagnetization occurs, and the weight hammer falls instantaneously to impact the pile top, thereby achieving the effects of a stable weight hammer and quick attraction and falling of the hammer. A clamping scale is arranged inside an adjustment section of a guide frame. A falling height of the weight hammer may be selected arbitrarily.

Abstract: An object detection training method can include receiving a training sample set in a current iteration of an object detection training process over an object detection neural network. The training sample set can include first samples of a first class and second samples of a second class. A first center loss value of each of the first and second samples can be determined. The first center loss value can be a distance between an eigenvector of the respective sample and a center eigenvector of the first or second class which the respective sample belongs to. A second center loss value of the training sample set can be determined according to the first center loss values of the first and second samples. A first target loss value of the current iteration can be determined according to the second center loss value of the training sample set.

Abstract: Techniques are disclosed relating to classifying transactions using post-transaction information. Training architecture may be used to train a first classifier module using first data for a set of transactions as training data input, where the first data includes both pre-transaction information and post-transaction information for transactions in the set of transactions. During training of the first classifier module, in disclosed techniques, correct classifications for the transaction in the set of transactions are known. The training architecture, in disclosed techniques, generates respective weights for multiple transactions in the set of transactions based on classification outputs of the trained first classifier for the multiple transactions. In disclosed techniques, the training architecture trains a second classifier module, based on the generated weights, using second data for the set of transactions as training data input.