[Objective] Artificial Intelligence (AI) has become a hot technology topic in the society. This paper aims to provide updated information of AI developments and advances to government, research community, industry and public. [Methods] This paper mainly reviews AI news and scientific publications in 2019 to summarize the overall technology progresses of AI. [Results] The AI community has become more rational and pragmatic relatively in 2019. On one hand, AI technologies are getting mature with more and more successful applications in the era of Internet of Things, Big Data, Super-computing, and Deep Learning. On the other hand, there are more and more promising and new directions in AI researches such as reinforcement learning, automatic machine learning, brain-machine interface, brain-like intelligence, and interpretable artificial intelligence. [Limitations] This paper mainly introduces the current development status of AI with limited efforts in prediction of future AI directions. [Conclusions] Great progress has been achieved in the areas of neuromorphic chips, brain-machine interface, human-computer game, intelligent robot, natural language understanding, and computer vision in 2019. But it is still a challenging task towards general AI. Technology breakthroughs are needed to achieve a new generation of AI.
[Objective] Document images carry important information of texts which are extensive in daily life. Extracting texts from images and converting to digital texts to be processed by computers is the main objective of document image recognition (also called as character recognition or OCR). Since 1950s, the field of document recognition has seen tremendous advances in research and applications. This paper provides an overview of document image recognition, facilitating research innovations and engineering applications. [Methods] In this article, I first introduce the applications needs of document recognition, then review the main advances of research in this field, analyze the strengths and weaknesses of the methods, and finally, prospect the future development. [Results] Numerous methods of statistical recognition, feature extraction, structural analysis, character segmentation, character string recognition and layout analysis were proposed during 1950s-2000s. [Conclusions] In recent years, deep learning methods (deep neural networks, DNNs) dominate the field, and have promoted the performance of text detection and recognition significantly. However, insufficiencies are still evident in complex layout analysis, character recognition reliability and generalization.
[Objective] This paper firstly introduces the start-of-art technical framework and main challenges of Automatic Speech Recognition (ASR) systems, then provides reference for further research in the field of ASR. [Methods] Firstly, the newest framework of end-to-end speech recognition is introduced, including the Connectionist Temporal Classification(CTC) and attention based framework. Secondly, four challenging problems in ASR applications are presented, including the recognition of noisy and distant field speech, the recognition of code-switching, the recognition of domain related terms, and minority language speech recognition with limited resources. [Results] For the problem of robustness of end-to-end ASR system, an improved enhancement method and filtering attention mechanism is proposed. The start-of-art methods and future development directions are discussed regarding to the challenging problems of ASR systems. [Conclusions] There is a major challenge for the commercialization of the end-to-end ASR systems, and the research on four challenging problems plays a key role in the application of ASR systems.
[Objective] This paper reviews the recent progress of deep learning researches and applications in medical image analysis. [Coverage] Relevant papers were first retrieved by keyword search and then by citation screening. [Methods] Deep learning based on convolutional neural networks is briefly introduced. Then, we review the diagnostic performance of deep learning on medical images in recent years with respect to different types of diseases, such as stroke, pulmonary nodules and bone age estimation. [Results] Deep learning for medical image interpretation has demonstrated advantages in many aspects, including accuracy, speed, stability and scalability. Meanwhile, existing problems may hinder clinical adoption of deep learning, such as dependence on a large amount of labelled data, inconsistent labeling standards, poor generalizability and interpretability of deep learning methods. [Limitations] There may be omissions of the retrieved literature, and it is impossible to compare the performance of the same deep learning model across different diseases. [Conclusions] Powerful artificial intelligence can improve the efficiency and accuracy of image interpretations for radiologists, but artificial intelligence is not perfect. Before being widely adopted in medical image interpretation, deep learning methods need more verification in real applications.
[Objective] Different from traditional underwater vehicles, the squid-like underwater robot has the advantages of low noise and good maneuverability. In this paper, the system design of the squid-like underwater robot was studied, including the mechanical structure design and drive system design. [Methods] The dynamic model of the squid-like underwater robot, the kinematic model of the undulating fins and the mapping model from propulsive force/torque to traveling wave parameter were established. [Results] The prototype of the squid-like underwater robot was developed. Then, multiple motion patterns such as marching, receding, turning, submerging and surfacing were achieved and tested. Besides, depth control experiment was performed. [Conclusion] Experimental results show that the squid-like underwater robot propelled by undulatory fins has good mobility.
[Objective] By providing the multi-participation mechanism of gathering, processing, openness, sharing, publication, citing and reusing scientific data, the block-chain-based scientific data identification ensures the reliability and stability of correlated data with accesses-controllable information storage and authentication. Thus it can provide a trusted and valuable reference for scientific data research. [Coverage] Besides the key technical methods of block chain construction and the applications based on identification knowledge, the article also focuses on the international and domestic researches on identification protocols, system instances, citation statistic methods of scientific data as well as the researches on cross disciplinary areas correlated to identification and block chain. [Methods] This paper presents an identification block chain platform in the field of scientific data citation statistic. Based on the overall architecture of the platform, we expound the key technologies and platform tools for constructing identification block chain, and give specific application cases based on three participants which are identification center, scientific data center and publications. [Results] The identification block chain platform constructs a trusted certification service for the key point life-cycle management for scientific data, while providing a trusted, multi-source, stable and cross-field intelligent correlated service. [Limitations] The integrated innovation of identification and block chain technology is constantly developing. However, the quantity of peers in the identification block affects platform service ability in the range of applications. In the future, we hope to carry out more services by block-chain crossing. [Conclusions] The identification block chain for scientific data, with its transparent and trusted authentication mechanism, flexible accesses control and strong correlated query ability, can organize the trusted information storage of the massive data from cross-fields and multi-source in a better way, which provides auxiliary functions for future technology evaluation of scientific data.
[Objective] This paper analyzes the research background of big data intelligent computing and associated challenging problems, then introduces multi-granularity cognitive computing, a novel model for big data intelligent computing in the view of cognitive computing. [Methods] Big data intelligent computing is shown to be a way to utilize the value of big data. The data computing mechanism of most traditional big data intelligent computing models is found inconsistent with the cognition mechanism of human brain. This paper introduces the multi-granularity cognitive computing model, which is a model for big data intelligent computing and unifies the “global precedence” law of human brain’s cognition mechanism (from coarse granularity to fine granularity) and the information processing mechanism in computer systems (from fine granularity to coarse granularity). The framework of data-driven granular cognitive computing (DGCC) is introduced. Furthermore, some application examples in intelligent manufacturing process industry are introduced. [Results] It is found that three kinds of scientific problems need to be studied for establishing the data-driven granular cognitive computing model and the integration of knowledge driven and data driven computing mechanisms. Those are multi-granularity space description, multi-granularity joint problem solving and human-computer cognitive mechanism integration. [Conclusion] Through its preliminary exploration on intelligent manufacturing in process industry, it is shown that multi-granularity cognitive computing is an effective new model to solve the problem of data-knowledge fusion in intelligent decision-making based on big data.
[Background] The rapid development of artificial intelligence technology depends on large-scale computing and data resources. Massive computing capacity is an effective guarantee for fast training of deep learning models. Standardized data is an important basis for artificial intelligence algorithms to carry out training processes and accuracy improvements. [Objective]The artificial intelligence computing and data service platform can effectively integrate computing, data and software resources into a unified virtual infrastructure. The platform can provide an integrated working environment for researchers. Besides, it also supports full life-cycle of model design, training and inference. [Methods] Under the circumstance of the convergence of artificial intelligence and high performance computing, we discuss typical usage scenarios, key features, and non-functional requirements related to artificial intelligence platforms. The platform for Artificial Intelligence Computing and Data Application Services at the Chinese Academy of Sciences is introduced. Then the architecture and services of the platform are discussed to address the issues in design and implement of the platform. [Results] By means of Web Services, command lines and online debug tools, the platform can support the rapid creation of artificial intelligence models in an easy-to-use way, process massive training jobs, and enable data access and transfer. [Conclusions] The platform for artificial intelligence and data services can provide an easy-to-use integrated work environment, and further advance the development of artificial intelligence in multiple research areas and disciplines.
[Objective] The rapid development of intelligent acoustic systems has remarkably promoted the progress of human life. While the convenient communication is being build between human and machine, it has also exposed various security issues, which can greatly threaten users’ privacy, property security, and even personal safety. This paper makes a comprehensive introduction to the research status of intelligent acoustic system security and provides references for the related research work in China. [Coverage] The references are selected from domestic and international journals or conferences in the field of artificial intelligence system security, with a total of more than 50 articles. [Methods] We firstly establish the security model of the intelligent acoustic system and analyze the purposes and abilities of attackers. Then, we systematically summarize the existing acoustic system security studies into five topics, which are input components, pre-processing components, machine learning models, output components and actual implementation frameworks. Finally, we discuss future research trends. Particularly, we point out potential research directions of pre-processing and output components, which may need more attention. [Results] The researches about intelligent acoustic system security have achieved considerable results, but is still at the initial stage in all, which means there are more directions to be explored, and problems to be solved. [Limitations] Acoustic system security has developed rapidly in recent years. This article is limited by the timeliness of access to data, some included researches may have new progress. [Conclusions] The researches on speech system security have high practical value. However, the existing researches have limitations in research focus points and practicality. There are still a lot of blank areas awaiting for scholars to further explore and study.
[Objective] Machine Learning (ML) is a booming field for many complicated problems which were previously unable to solve effectively with conventional methods. In order to deliver big science findings, a modern accelerators such as the under construction low emittance synchrotron radiation based light source, High Energy Photon Source (HEPS), located in suburban Beijing require very high precession control systems to handle thousands of individual devices to work coherently and smartly to perform at its highest running level. This paper introduces some initial work for adopting ML in the accelerator field. [Method] For such a large accelerator, conventional control approach may be insufficient for handling the complexity of operations. This paper outlines that ML techniques may help the accelerator in many aspects and, more importantly, discusses how to prepare the data for ML. Also, a software architecture which is suitable for ML applications applied to accelerator field is introduced. [Results] A global database structure which can cover nearly all accelerator data has been designed and under implementation. In addition, initial works of ML for both accelerator operation and design are shown. [Conclusion] The ML techniques for accelerator start well with some positive results. In the meantime, collaborations between organizations are formed to share work load and speed up development. As the software infrastructure being developed and more good quality data being collected, the ML for accelerators should produce much better results.
[Objective] This paper investigates the frontiers and trends of artificial intelligence research between 2006 and 2019. [Coverage] 28862 articles published between 2006 and 2019 in Class A and Class B journals recommend by China Computer Federation were studied. [Methods] By means of burst term detection on a large number of articles and content analysis on a few articles, this paper investigates the frontiers and trends of artificial intelligence research. [Results] 20 major frontier research topics between 2006 and 2019 were detected, which could be summarized in three development stages. The emergence of frontier research topics has been accelerating in recent 5 years. [Limitations] The coverage of content analysis is not broad enough; the data and information only considers journal research articles; the tool and algorithm employed still have some limitations. [Conclusions] The emergence of frontier research topics of artificial intelligence are accelerated towards more sophisticated problems, enhancing diversity of research topic.