Editorial: Special Topic on Data Intelligence in New AI Era

Editorial on Special Topic: 

Data Intelligence in New AI Era

Xu Chengzhong1, QIAO Yu2

（1. University of Macau (China)
2. Shenzhen Institutes of Advanced Technology (SIAT), the Chinese Academy of Science,(China) )

        The new artificial intelligence (AI) era heavily depends on three converging forces: the advance of AI algorithms, the availability of big data, and the popularity of high performance computing plat forms. Data-driven intelligence, or data intelligence, is a new form of AI technologies that leverages the power of big data and advanced learning algorithm. It is becoming an extremely ac tive research area with broad area of applications such as com puter vision, speech recognition, natural language processing, medial and healthy, intelligent transportation system, multime dia system, communication, and social network.

        With the huge volume of data available in various domains, big data brings opportunities to boost the performance of AI system with advanced machine learning especially deep learn ing techniques. It has been widely verified that deep neural net works achieve significantly better performance than previous shallow models and even surpass human performance in cer tain specific tasks or datasets. One well-known example is Ima geNet Large Scale Visual Recognition Challenge (ILSVRC) which aims to classify or detect objects in images from 1 000 categories. The state-of-the-art deep convolutional neural net works like squeeze-and-excitation networks (SE-Net) have achieved error rates lower than 3%, which is better than human performance (error rate 5.1%). These networks usually include a deep architecture with a huge number of parameters, which are optimized with one million training datasets. In nature lan guage processing, recent language networks like BERT or XL Net leverage more than 100 GB text for training and achieve re-markable performance on wide tasks like SQuAD, GLUE, and RACE. All these successes heavily rely on the large scale training data. In addition to the amount of data, the label or annotation of data is also important in supervised learning. Although unsupervised learning is desirable in many applications, supervised learning usually exhibits better performance. In the next, it is important to design effective learning algorithms in unsupervised, semi-supervised, or weakly-supervised setup. On the other hand, it also presents unprecedented challenges to manage and exploit big data for a variety of applications. Learning with big data is not easy, which always needs powerful models and efficient training algorithms. Take deep networks as an example. One may need to carefully design network architectures, training losses and strategies, and effectively exploit high performance computing platforms.

        This special issue seeks original articles describing development, relevant trends, challenges, and current practices in the field of big data, artificial intelligence and their applications. After careful reviews, four papers have been selected for publication in this special issue.

        The first paper is titled “A Lightweight Sentiment Analysis Method” . It proposes a data driven approach to perform sentiment analysis of film’s critics from the Douban website and visualize the results with a word cloud.

The second paper is a survey paper with the title of “Big Data-Driven Residents’ Travel Mode Choice: A Research Overview” . This paper surveys the studies of residents’travel mode identification, influencing factors acquisition and choice model construction using data driven approaches.

        Face detection is a fundamental yet important problem in computer vision. The third paper“Face Detection, Alignment, Quality Assessment and Attribute Analysis with Multi-Task Hybrid Convolutional Neural Networks”introduces multi-task hybrid convolutional neural networks for face detection, alignment, quality assessment and attribute estimation.

        The last paper “RAN Centric Data Collection for New Radio” is from the communication area, which exploits self-organizing networks and minimization of driver tests to support de-ployment of new radio (NR) system and conduct performance optimization.

Finally, we would like to thank all the authors, the external reviewers for their contributions and efforts to organize this spe-cial issue in this esteemed journal.