Recent Advances in Learning Theory

Abstract

In the era facilitated by the Internet of Things, ubiquitous communications as well as cloud services, sensing means, and human-computer interfaces are becoming all-pervasive and online. This makes it more possible for us than ever before to study engineering problems, human activities, and social behaviors through machine learning analysis of the big data produced in the ubiquitous environment. Looking at the recent history and new trends, machine learning has made attractive progress in wide areas of applications, from natural language to nonverbal communication, from engineering application to humanities, arts, and social studies, and from the real world to cyber space. In 1997, Dietterich summarized the development of machine learning in four directions: ensembles of classifiers, methods for scaling up supervised learning algorithm, reinforcement learning, and learning of complex stochastic models [1]. Latterly, Duch addressed machine learning as the foundations of computational intelligence comprehensively [2]. Since the beginning of the 21st century, research in machine learning has made progress in all of the four directions and has become focused on new challenges in learning from big data that cover a variety of application areas. One barrier that needs to be broken through is how to avoid the “curse of dimensionality” and ensure the generalization ability in the learning process. Owing to the efforts such as work of Koller and Friedman on Probabilistic Graphical Models [3] and the Compressive Sensing (CS) theory [4], the tentative path has been lightened. As the forethought by Wang [5], “predicting” the changes based on generalization ability and “describing” the knowledge discovered from huge data will be the two major tasks of machine learning in the future. In today's society, machine learning has been in an extensive demand in the areas associated with human's psychology and behaviors, such as ubiquitous learning, e-commerce, online customer service, behavioral finance analysis, and government emergency management. We believe that machine learning could be the most promising, sometimes even the only, way to accomplish the complex computation on human psychology and behaviors in the ubiquitous environment. However, in doing so, the machine learning research needs to pay attention to the following new aspects which may be beyond the ability of computer science and technology and requires more novel interdisciplinary ideas and methods: (1) a systematic model such as the social neuroscience mechanism [6] which can describe the neural activities and dominant process of human psychology and behaviors and helps the machine to understand its globally structural features and therefore reduce the computational cost by learning from the limited samples of a big data set, (2) the comprehensive context awareness in physical, cyber, and psychosocial spaces, as well as the information fusion processing and computing ability, which has been called Cyber Psychosocial and Physical (CPP) Computation by Dai [7], and (3) smart learning that enables the machine to cope with both rational intelligence and emotional intelligence in the learning process. The aim of this special issue is to bring together researchers working in different areas for exchanging and sharing with each other their progress in the new tendency of modern learning theory and applications. We were very pleased to see various new research ideas and many innovative contributions in the submitted manuscripts, which cover a wide range of recent advances in learning theory, techniques, and applications. What follows is a brief editorial review of the published papers in this special issue from four perspectives: Neuroscience in Learning Theory, Machine Learning in Psychological Computation and Behavior Analysis, Machine Learning in Public Management and Business Service, and Machine Learning in Knowledge Discovery and Human-Computer Engineering.