A feature representation learning system for multi-modal big data in cyberspace

Abstract

The invention relates to a feature representation learning system oriented towards multi-modal big data in the network space. The system mainly includes seven units: a multi-modal sample generation unit, three homogeneous feature extraction units, a data classification unit, a feature measurement unit, and a multi-task loss function integration unit. The multi-modal sample generation unit constructs a training sample set with four components. The three homogeneous feature extraction units are responsible for processing the first three component data of the training samples and generating threeone-dimensional feature vectors. The data classification unit learns the first three component data of the training samples by category, and constructs a classification task loss function based on the fourth component data. The feature measurement unit carries out feature measurement learning on the first three component data of the training samples, and constructs a measurement task loss function. The multi-task loss function integration unit integrates different tasks in a weighted manner, and optimizes the system parameters. Compared with the existing method, the feature representation learning system has the advantages of multiple modes, high accuracy, strong generalization ability, convenient implementation, and the like, and can be effectively applied to public opinion monitoring, Internet health care, personalized recommendation, intelligent question answering and other fields.

Description

technical field [0001] The invention relates to the field of computer application technology, in particular to a feature representation learning technology of multi-modal big data. Background technique [0002] In recent years, with the rapid development of technologies such as the Internet of Things, cloud computing, and social networks, big data in cyberspace has increasingly shown 4 "V" (Volume, Velocity, Variety, Veracity) characteristics. Google needs to process more than 500PB of data per month; Baidu processes tens of PB of data every day; Facebook has more than 1.5 billion registered users, uploads more than 2 billion photos every month, and generates more than 400TB of log data every day. According to the calculation of International Data Corporation IDC, in 2017, cyberspace will generate 2,000 EB of data, and in 2018, it will increase by 40%, reaching 2,800 EB. By 2020, it will reach 35,000 EB, exceeding the current storage capacity of disk space. [0003] With th...

AI Technical Summary

Problems solved by technology

However, we found that these existing methods are basically aimed at a certain type of single-modal big data in cyberspace, such as structured data, text data, image data or video data, without involving simultaneous processing of multiple modalities. Coexisting cyberspace big data, and when the cyberspace big data includes noise, the accuracy and generalization ability of existing methods are relatively poor

Images