The present application provides a method of processing a video, a method of querying a video, and a method of training a video processing model. A specific implementation solution of the method of processing the video includes: extracting, for a video to be processed, a plurality of video features under a plurality of receptive fields; extracting a local feature of the video to be processed according to a video feature under a target receptive field in the plurality of receptive fields; obtaining a global feature of the video to be processed according to a video feature under a largest receptive field in the plurality of receptive fields; and merging the local feature and the global feature to obtain a target feature of the video to be processed.

Methods and apparatuses are described for intelligent text classification with limited or no training data. A server computing device receives one or more of structured text or unstructured text corresponding to compliance text data from a database. The server computing device executes a trained few-shot natural language inference (NLI) classification model on one or more sentences in the received compliance text data to identify whether the one or more sentences comprise a compliance violation. The server computing device transmits the results of the model execution to a remote computing device.

Methods and systems for training a model include collecting unlabeled training data during operation of a device. A model is adapted to operational conditions of the device using the unlabeled training data. The model includes a shared encoder that is trained on labeled training data from multiple devices and further includes a device-specific decoder that is trained on labeled training data corresponding to the device.

A learning device includes: a data acquisition unit that acquires learning target data that is full-size data of a learning target; a data generation unit that divides the learning target data to generate multiple pieces of first divided data that is divided data of the learning target data, and adds, to each piece of the first divided data, first identification information for identifying a region of the first divided data in the learning target data; and a model generation unit that generates a learned model for determining an anomaly in the first divided data using first correspondence information that is a set of the first divided data and the first identification information corresponding to the first divided data.

Provided are a voice packet recommendation method and apparatus, a device and a storage medium, relating to intelligent search technologies. The solution includes constructing a first video training sample according to first user behavior data of a first sample user in a video recommendation scenario and first video data associated with the first user behavior data; constructing a user training sample according to sample search data of the first sample user and historical interaction data about a first sample voice packet; pretraining a neural network model according to the first video training sample and the user training sample; and retraining the pretrained neural network model by using a sample video and sample tag data which are associated with a second sample voice packet to obtain a voice packet recommendation model. With the solution, the neural network model can be trained in the case of cold start so that the neural network model can recommend a voice packet automatically in the case of cold start.

An improved neural network is disclosed that supports rapid retraining using human feedback. A weighted ensemble of Extreme Learning Machines (ELMs) is used to implement a model. The ensemble of ELMs may be trained in parallel with a variation in individual parameters gridding a parameter set selected to achieve consistent accurate model results when the model is trained and subsequently retrained when user feedback data become available. An exemplary application is the scoring of resumes.

A method for data sequence prediction and resource allocation includes determining, by a memory system, a plurality of resource parameters associated with operation of the memory system and determining respective time intervals associated with usage patterns corresponding to the memory system, the respective time intervals being associated with one or more sets of the plurality of resource parameters. The method further includes determining, using the plurality of resource parameters, one or more weights for hidden layers of a neural network for the respective time intervals associated with the usage patterns and allocating computing resources within the memory system for use in execution of workloads based on the determined one or more weights for hidden layers of the neural network.

The present invention presents a universal reconfigurable video stream processing system where a digital twin is applied to 3D marker cloud mapping of a set of parameters, related to the current state of the monitored person (object). The invention includes two reconfigurable units, with at least one of these units being universally adjusted for any input-output mapping application with fixed input size, fixed output size and numerical values ordered by their meaning. Each reconfigurable unit includes at least one machine learning based mathematical model with a high number of parameters and non-linear functions performing as a universal approximator and ensuring high flexibility during training process. Each unit of the presented system, which includes a machine learning based mathematical model should be trained in advance of system execution with input-output mapping examples, where the range of the input values in the training example set should cover the range of the input values that will be used during system execution.

Components are extracted from user data being read from a reader of a hard disk drive. The components collectively indicate both a magnitude and direction of a read offset of the reader over a track. The components are input to a machine-learning processor during operation of the hard disk drive, causing the machine-learning processor to produce an output. A read offset of the reader is estimated during the operation of the hard drive head based on the output of the machine learning processor. While reading the user data, a radial position of the reader over the track is adjusted via an actuator based on the estimated read offset.

A social recommendation method based on a multi-feature heterogeneous graph neural network is provided and includes: extracting attribute information of users and topics to code; processing user coding information and topic coding information through a multi-layer perceptron to obtain initial feature vectors of the users and the topics; establishing a heterogeneous graph by taking the users and the topics as nodes; inputting the heterogeneous graph into a heterogeneous graph neural network to perform information transmission in combination with an attention mechanism, and updating the feature vectors; and performing similarity calculation on the feature vectors of the users, and selecting the user and the topic with the highest similarity with the feature vector of the user for recommendation. Social information can be mined more comprehensively, features of users and interested topics of the users can be deeply fused, and recommendation accuracy and user experience can be improved.