《1 Engineering research fronts》

1 Engineering research fronts

《1.1 Development trends in the top 10 engineering research fronts》

1.1 Development trends in the top 10 engineering research fronts

The top 10 engineering research fronts reviewed by the Information and Electronic Engineering Group are summarized in Table 1.1.1, and include electronic science and technology, optical engineering and technology, instrument science and technology, information and communication engineering, computer science and technology, and control science. Among these research fronts, "radar stealth technology," "new- generation mobile communication technology," "quantum coherence measurement and decoherence control," "software robot control method," "high-resolution remote sensing scene classification and image processing technology," and "human body gesture recognition method based on deep neural network” are based on the popular topics provided by Clarivate Analytics. Furthermore, "interpretable deep learning," "networked collaborative sensing and control theory," "blockchain technology," and "silicon-based optical interconnect chip technology" are recommended by the experts.

The number of core papers related to each front, published from 2012–2017, is shown in Table 1.1.2. Considering the number of core papers published in recent years, "high- resolution remote sensing scene classification and image processing technology" is the most significant front.

(1)  Radar stealth technology

Radar stealth technology is also known as radar low observable technology or radar target feature signal control technology. It is used for controlling the scattering direction, polarization mode, and radiation intensity and mode of the incident electromagnetic wave on the target surface by its shape, material, and comprehensive object and circuit designs, thereby reducing the probability of detection and recognition by the enemy radar system. Using radar is the

《Table 1.1.1》

Table 1.1.1 Top 10 engineering research fronts in information and electronic engineering

《Table 1.1.2》

Table 1.1.2 Annual number of core papers published for each of the top 10 engineering research fronts in information and electronic engineering

most efficient method of detecting targets from a long distance; thus, radar stealth technology has been the focus of stealth technology development since its inception.

Considering electromagnetic wave manipulation in technology implementation, radar stealth technology includes mainly integrated design and evaluation of shape stealth, material stealth, and stealth-integrated design and evaluation. In shape stealth, the incident direction and scattering mode of the electromagnetic wave are adjusted by the shape design to achieve the stealth effect. It aims at reflecting the incident wave to the nonthreatening directions and reducing or blocking the strong scattering source and discontinuous structure scattering of the target surface at dihedral/trihedral angles. In material stealth, the electromagnetic waves are either attenuated or their radiation is changed by applying and mounting special functional materials on the target surface or its key parts. The active stealth technology refers to real-time generation of electromagnetic field signals that are opposite to the target’s scattering field, according to the incident electromagnetic wave and target states of the target surface, thereby realizing zero scattering by spatial cancellation and reducing the target’s radar cross section. Stealth-integrated design and evaluation must be emphasized and focused on in the application of radar stealth technology. This is one of the primary reasons that although the principle of stealth technology is open to everyone, only a few countries have mastered it.

As the detection and anti-detection technologies continue to progress, the main directions of future radar stealth technology development are: ① very low and ultra-low observables; ② wideband and omnidirectional stealth; ③ dual/multiple station radar stealth; ④ integrated synthesis and thin stealth design; and ⑤ disturbance/attenuation field reduction.

(2)  Interpretable deep learning

In recent years, deep learning methods represented by convolution depth, cyclic, and generic neural networks, and deep reinforcement learning have been applied in the fields of image classification and target detection, speech recognition and synthesis, and natural language processing; and their performance has improved dramatically. Although deep learning exhibits excellent performance in various artificial intelligence (AI) applications, its interpretability has always been its weakness. Currently, the highly discriminative ability of deep neural networks is layered by constructing multiple layers of nonlinear mapping functions. As abstraction, the black box effect is presented, and the relationship between its internal network structure and learning parameters, and its decision output is difficult to establish. Interpretable AI can break through the primary bottlenecks of deep learning, such as effective learning of small samples or weakly labeled data, human–computer interaction learning at the semantic level, and semantic debugging of neural network representation.

Currently, the study of understanding or decoupling complex neural network representation to improve its interpretability includes mainly five aspects: ① visualization of the convolutional neural network representation in an intermediate network layer; ② pairs of convolution features, mapping space, and correspondence of different semantic categories for diagnosis; ③ decoupling mixed modes of different convolutional layers; ④ establishing interpretable deep network models, such as interpreting convolution depth neural networks and capsule networks; and ⑤ semantic layer learning by human–computer interaction. Moreover, anti- machine learning detects the vulnerability of deep learning models by constructing anti-samples. Anti-machine learning builds against sample detection based on the vulnerability of deep learning models. Preliminary studies have shown that it is helpful to open the black box of a depth model by adding the confrontation sample formed by the perturbation to the meta-predictor or constructing the influential function traceability model prediction for its training data, and to a certain extent its predicted behavior or decision boundaries are interpreted.

Combining rule-based symbolic reasoning with data-driven learning processes can enhance the interpretability of intelligent learning processes. For example, inductive learning combines neural perceptrons with logical reasoning so that the learning process can directly use domain knowledge, thereby enhancing the learning process. Both attention and memory play an important role in the process of human cognitive reasoning, especially for knowledge acquisition, understanding, and reasoning of sequence data, such as text, speech, and video. Constructing deep neural reasoning mechanisms with the support of attention mechanism and memory structure to enhance the interpretability of intelligent learning and reasoning is an important research field of interpretable deep learning.

(3)  New-generation mobile communication technology

To cope with the explosive growth of mobile data traffic, simultaneous access of massive devices, and emergence of various services in the future, the new-generation mobile communication technology needs to provide ms-level and end-to-end connectivity for hundreds of billions of devices, provide peak Gbit/s transmission rate, and support ultra- high mobility, traffic density, connectivity density, and other diverse application scenarios. The new generation of mobile communication technology will integrate various wireless access modes to achieve flexible network deployment, operation, and maintenance; comprehensively enhance the spectrum, energy, and cost efficiencies; and promote the sustainable development of the mobile communication industry. The new generation of communication technology involves the design, development, and production of the new generation of mobile chips and related core technologies. Furthermore, it involves the development of the entire information industry. It is directly related to industrial transformation and upgrading, and has a direct, significant, and simultaneous impact on the manufacturing industry and services. Further, development of a communication technology involves global standards, various patents, and huge network construction costs. Thus, for every country, the development of a new mobile communication technology brings great economic and social benefits.

(4)  Networked collaborative sensing and control theory

A networked collaborative sensing and control system usually consists of a set of sensors and controllers, in which the sensors get information from the environment to be monitored, while the controllers try to change the environment via control signals, and the sensors and controllers are connected via communication networks. A networked collaborative sensing and control system has the functions of sensing and changing the physical world, and is applied in various areas, such as disaster relief, intelligent building, factory automation, underground safety control, and biochemical attack detection.

Compared with conventional centralized sensing and control methods, the networked distributed cooperative sensing and control methods have the following advantages. ① Decentralization of sensing and control functions improves the robustness of the system to the failure of sensing and control nodes. ② In large-scale distributed systems, achieving full network synchronization is difficult, and the nodes may work asynchronously. Conventional centralized sensing and control methods cannot deal with asynchronous information. However, in a network-based distributed structure, the asynchronous sensing and control problem can be solved via the hierarchical information processing approach. ③ The computational efficiency of a system can be greatly improved by the decentralization of sensing and control functions. Through information interaction among subsystems, computational complexity can be reduced while achieving good performance with centralized sensing and control.

In recent years, the research on the theory and method of distributed sensing and control based on information interaction has received great attention. However, many theoretical and application-related problems are still open, including distributed sensing and control based on the cooperative (such as controllers and sensors) and noncooperative modes of nodes. These include the asynchronous multi-rate distributed sensing and control of each subsystem in an asynchronous working state, extensibility of distributed cooperative sensing and control (allowing the arbitrary access and exit of any node), and the influence of network topology and vulnerability on distributed sensing and control. It is believed that breakthroughs in the future will effectively promote the development of intelligent manufacturing, networked autonomous systems, smart grids, and other fields under industrial Internet.

(5)  Blockchain technology

Blockchain technology, which includes P2P technology, cryptography, and consensus algorithms, aims to provide a mechanism for information and value transfer in untrusted environments and is a cornerstone for the development of the future Internet. It has the features of unchangeable data, collective system maintenance, and information disclosure. Blockchain, as a versatile technique, accelerates distributed technology applications from digital currencies into other areas and integrates innovations from all industries. It has the following three advantages. ① The consensus algorithm ensures that the data on the blockchain is secure and difficult to tamper with. ② Each node of a blockchain has a piece of data; thus, it is heterogeneous and reliable. ③ With smart contract attributes, it can automatically perform decentralization of applications.

From a single application to multiple fields, the blockchain technology still has a long way to go. It will continue to evolve in the future, and technical processes, such as consensus algorithms, sharded services, processing methods, and organizational forms will continue to change. However, there are a few challenges as well. First, performance and scalability cannot meet requirements. The transaction throughput and storage bandwidth are far from meeting the actual requirements of applications. Currently, Bitcoin’s transaction throughput rate is seven transactions per second and that of Ethereum is 14 transactions per second. How to improve transaction throughput without affecting the overall security of the system will be worth studying. Furthermore, by compressing the block time and increasing the block size, the sharding technology can effectively improve the transaction throughput rate. Second, data privacy and access control require improvement. In the existing public blockchains, each participant can obtain a complete data backup. All data are transparent to the participants, and it is impossible for the participants to obtain only specific information. How to ensure transaction privacy without affecting the public blockchain execution efficiency is still a challenge. Currently, research in this area includes the coin-conserving, zero-knowledge proof, and ring signature mechanisms to protect a user's transaction privacy data. Third, the governance mechanism requires improvement. Public blockchain communities have explored various upgrade mechanisms, such as “hard fork” and “soft fork”; however, the remaining problems cannot be ignored. Public blockchains cannot be “closed,” and its bug fixes are extremely troublesome; thus, it can be fatal if there is a security breach.

(6)   Quantum coherence measurement and decoherence control

Quantum coherence and quantum phase transition play an important role in quantum information. Although systems have always been coherent in theory in the evolution process, an actual system is not strictly closed. Thus, it will inevitably become entangled with the environment and cause decoherence. In quantum mechanics, the quantum coherence of open quantum systems is gradually lost over time due to quantum entanglement with the external environment. This effect is called “quantum decoherence.” Quantum decoherence is the consequence of quantum entanglement between quantum systems and the environment. The interference phenomenon caused by quantum coherence will disappear because of quantum decoherence. Quantum decoherence changes the quantum behavior of a system into a classical behavior, and this transition is called quantum-to-classical transition. Hans Zehe, a German physicist, first proposed this concept of quantum decoherence in 1970. Since 1980, quantum decoherence has become a popular research topic.

Quantum decoherence usually occurs very quickly; therefore, making superimposed objects in macroscopic or mesoscopic forms is difficult. To experimentally verify the quantum decoherence effect, witness the smooth boundary between quantum and classical behaviors, test and improve the theoretical model of quantum decoherence, and find out any difference from quantum mechanical evolution, we must perform the following challenging tasks: ① preparing several quantum superposition states of macroscopic or mesoscopic states that can be resolved; ② designing a set of methods to confirm quantum superposition; ③ having the duration of quantum decoherence long enough to observe it correctly; and ④ designing a set of methods to supervise quantum decoherence.

The influence of decoherence on quantum information science can be roughly divided into two parts: quantum computing and quantum communication. We know that in quantum information science, the state of quantum systems contains information. Quantum decoherence causes loss of some or all information of the system; thus, it will cause calculation errors during quantum computation. In quantum communication, the information in the transmission channel is prone to disturbance and interference, and the receiver at the end of the channel can possibly receive noise and error information; thus, it requires assistance from a debugging system, such as an encoder.

(7)  Software robot control method

The research on control of soft robots is essentially to solve the inverse kinematics problems. Unlike conventional rigid- bodied robots that are driven by electric motors, soft robots generate motions using soft materials similar to natural muscles, which arm the soft robots with animal-like agility and flexibility, increase adaptability in complex working environments, and reduce hazards related to human–machine interactions. Theoretically, a soft robot has an infinite number of degrees of freedom, making it possible to produce extremely complicated motions, such as stretching, bending, and twisting. Thus, motion control in soft robots is faced with enormous challenges. Currently, soft materials employed for artificial muscles include pneumatic artificial muscles, shape memory alloys, electroactive polymers, and so on. Although these materials exhibit intriguing attributes, such as high energy density, large deformation, and low weight, they are usually highly nonlinear with strong viscoelasticity. Moreover, considering their inherent compliance, soft robots are likely to work in environments with remarkably high uncertainties. Therefore, fine adaptability and strong robustness are preferred for the control of soft robots. Generally, control of soft robots can be achieved using two types of approaches: ① model-based and ② learning-based control approaches.

The former approach develops a dynamic model of the system through the first principle or data-driven methods and achieves motion control using conventional feedback control schemes. However, the drawback of this method is that it partly depends on the model; a kinetic/dynamic model of soft robots will result in a lot of uncertainties during their interactions with the environments. To overcome such limitations, the latter approach, inspired by the similarities between soft actuators and natural muscles, draws lessons from the motor control of natural muscles and compensates for the model uncertainties through on-line learning, which is processed in an artificial neural network. However, many explorations are required for the optimization of the network structure and the tuning of the parameters.

(8)   High-resolution remote sensing scene classification and image processing technology

High-resolution remote sensing imagery is a large spatial data related to emergency and disaster mitigation applications, which influence national economy and livelihood of the people. Remote sensing scene classification belongs to the research category of overall image understanding. Scene classification should analyze, judge, interpret, and label image scenes. It is a mapping process for learning and discovering images and scene semantic content tags. Scenes usually contain multiple targets, and scene classification is also a key issue in image understanding. Common scene classification algorithms include scene classification of local and middle- level semantics, and that of semantic topic models. According to the hierarchy, scene classification can be divided into two main methods: low-level and middle-level feature descriptions. However, low-level feature descriptions tend to have poor generalization performance and are difficult to process during image classification outside the training set; thus, most of the current scene classification algorithms focus on scene classification based on middle-level semantic modeling.

The middle-layer feature is a type of aggregation and integration of low-level features, and its essence is to use statistical distribution to establish a relationship between features and categories. Generally, the global low-level features cannot reflect local objects. Considering local low- level feature description, multi-local feature fusion and integrated learning can improve the recognition rate of scene classification. The scene classification based on the bag-of-visual-words (BoVW) model is a widely used middle- level semantic algorithm. Considering that spatial symbiosis and context of the vocabulary will help to improve the interpretation of the semantics of the scene structure, the BoVW model does not need to analyze the specific target composition of the scene, establish visual words according to the statistical characteristics of the low-level features of the scene, and then express the image scene information by using the visual word distribution of the image. However, the number of words to be set in the BoVW model is not known. The generated objects tend to have a greater correlation with the training samples that is an important factor affecting the robustness of the algorithm.

Statistical learning theory is a relatively mature research algorithm. In multi-sensor fusion remote sensing image scene classification, the use of support vector machines based on structural risk minimization principle and random forest algorithm using bootstrap resampling has been reported. To achieve scene classification of airborne radar and multispectral images, classification is often conducted using random forests combined with a classical priori model and an unsupervised image segmentation tool, Markov random field (MRF). Recently, semantic topic models such as pLSA and LDA that originated from text classification research have achieved good experimental results in scene classification of spaceborne landslides and airborne aerial photography.

The middle-level semantic scene classification alleviates the problem of semantic gap to a certain extent; however, there are no good solutions to the change of scene scale, the difference of sensor shooting angle and time, and the change of semantic object combination.

(9)  Silicon-based optical interconnect chip technology

Silicon is not only an electronic material but also a photonic material. Using silicon as the substrate, photonic devices can be fabricated using existing integrated circuit processes. Silicon-based optical interconnect chip technology uses silicon and silicon-based substrate materials as an optical media to fabricate corresponding photonic devices and optoelectronic devices (including silicon-based optical transceivers, silicon-based optical modulators, and silicon- based optical waveguides) through integrated circuit processes. Further, they use these devices to process and manipulate photons to achieve optical interconnection between systems, motherboards, chips, CPUs, and CPU cores. Inter-chip and on-chip optical interconnections, compared with electrical interconnection technology, has the fundamental advantages of ultra-high bandwidth, high speed, low power consumption, low distortion, low crosstalk, and no electromagnetic interference. The main research directions of silicon-based photonic integration for optical interconnection are as follows: ① light sources for emitting light waves as an information carrier; ② optical waveguides for transmitting optical signals; ③ modulator for generating calculation unit (the electrical signal is loaded on the optical wave carrier); ④ optical signal receivers for receiving the optical signal and converting it into an electrical signal for feedback to the computing unit; and ⑤ system integration.

Although silicon has many unique advantages as an optical interconnect material, as a single material it cannot accomplish all the functions of an optical interconnect device. For example, optical transmission and detection are a contradiction. Light transmission requires the material be transparent to photons, while light detection requires the material be opaque to light and absorb photons. Silicon is an indirect bandgap semiconductor that cannot achieve stimulated radiation easily; thus, fabricating it as a source material is difficult. Furthermore, silicon has some technical bottlenecks to overcome in optical interconnect applications. For example, the silicon-based photonic processors developed in 2015 and 2018 by the University of Berkeley and other institutes, as reported by Nature, used external light sources. Therefore, an important development trend is to find other materials that are compatible with silicon-based CMOS processes to compensate for the deficiencies of silicon in the optical interconnect chip technology.

(10)  Human body gesture recognition method based on deep neural network

Human motion recognition based on smart phones and wearable devices, such as wristbands and watches, has become a mainstream recognition method. Traditional machine learning methods, such as support vector machines, Bayesian networks, time- and frequency-domain analysis, and other machine learning methods, require extraction of features based on professional human motion domain knowledge. Neural network-based methods are still few, and neural networks and deep learning still require artificial extraction features. Feature extraction is a key step in machine learning and deep learning. Similarly, for human motion recognition, feature extraction of sensor data is extremely important.

Human motion recognition is an important research topic, especially in the popularity of smartphones and smart wearable devices nowadays. For human motion recognition, machine-learning methods include mainly traditional support vector machine, decision tree, KNN, naive Bayes, neural network, and deep learning. Model-dependent training data sources include a single acceleration sensor, or a combination of gyroscopes, magnetic fields, and even sound information. The position of the sensor in motion recognition is divided mainly into a fixed position (a plurality of sensors are generally fixed positions) and a nonfixed position. Feature extraction is conducted mainly in the time domain, and in fewer cases, in the frequency domain.

《1.2 Interpretations for three key engineering research fronts》

1.2 Interpretations for three key engineering research fronts

1.2.1 Radar stealth technology

Radar stealth (invisibility) technology, also known as radar low observability technology, is professionally called as radar target signature control technology. It is used for manipulating the scattering direction, polarization, intensity, and pattern of the electromagnetic wave illuminating the target surface via synthetic designing of shape, material, circuit and other aspects, thereby reducing the probability of detection and recognition by the enemy radar system. Using radar is the most efficient method to detect remote targets; thus, radar stealth has always been the focus of stealth technology.

Considering electromagnetic wave manipulation in technology implementation, radar stealth technology includes mainly shape stealth, material stealth, active stealth, and stealth-integrated design and evaluation.

Shape stealth was the first to receive attention and development in the radar stealth technology. It regulates the incident direction and the scattering mode of electromagnetic waves to achieve stealth. The focus is to reflect the incident waves to nonthreatening directions, and to reduce or block the strong scattering sources, such as dihedral angles/trihedral angles of the target surface as well as the discontinuous structure scattering. Considering a stealth aircraft as an example, the critical technologies to be solved in the shape stealth design are the integrated design of stealth shape and aerodynamic layout, the real-time electromagnetic calculation technology of the electric large size target, and the integrated fast forming technology.

Material stealth is used to change the attenuation or radiation of the electromagnetic waves by applying and mounting special functional materials on the target surface or its key parts (such as electronic system antennas and some strong scattering points). According to the electromagnetic regulation method of materials, the stealth materials currently in research and development include mainly: absorbing type stealth materials that introduce the radar waves into the material and then dissipate their energy; surface type stealth materials represented by metamaterials that change the scattering pattern and the polarization mode of electromagnetic wave by sub-wavelength periodic geometric circuit structure design. The key technologies to be addressed in the development of material stealth technology include electromagnetic/thermal/force integrated design and analysis, micro-nano processing, and metamaterial technology. The surface electromagnetic control materials represented by metamaterials are the focus of future stealth materials development.

The active stealth technology refers to real-time generation of electromagnetic field signals opposite to the target scattering field according to the incident electromagnetic wave state (incident direction and polarization) and target state (attitude velocity), thereby achieving zero scattering through space cancellation and RCS reduction of the target. The active stealth technology is still in the stage of exploration and development. The key technologies to be solved include real- time electromagnetic spectrum sensing and measurement, real-time generation and precise control of electromagnetic field signals, information metamaterials, and so on. Intelligent skinning for active stealth is an important direction for the development of future stealth technology.

Stealth integrated design and evaluation is a technical field that must be emphasized. This is one of the important reasons why theoretical principle is basically open but the stealth technology is mastered by only a few countries. The development of multi-sensor battlefield sensing technology, stealth technology must also emphasize the integrated control technology of active and passive features such as electromagnetic, optical, infrared, and acoustic ones. Therefore, the key technologies that must be solved and developed include multi-field coupling parallel computing, simulation testing and evaluation of stealth performance, as well as related design tools and instrumentation techniques.

Currently, radar stealth technology has been successfully applied to multi-type weapons and equipment after more than 50 years of development. Taking stealth aircraft as an example, it evolves from the earliest representative F117 to the current F22, F35, and B2. China and Russia have also become two of the few countries that have mastered the design techniques of stealth aircraft. As the escalation of the game between detection and anti-detection technology, the main directions of future radar stealth technology include:

(1)   Development in a very low observable and ultra-low observable direction;

(2)   Development in the broadband and omnidirectional stealth direction;

(3)  Development from mono-static radar stealth to dual/multi- static radar stealth direction;

(4)  Stealth design in the integrated and lightweight direction;

(5) Development from scattering field reduction to disturbance field/attenuation field reduction.

The countries or regions with the greatest output of core papers, institutions with the greatest output of core papers, countries or regions with the greatest output of citing papers, and institutions with the greatest output of citing papers on“radar stealth technology” are shown in Tables 1.2.1–1.2.4. The collaboration networks among major countries or regions as well as among major institutions are shown in Figure 1.2.1 and Figure 1.2.2, respectively.

1.2.2 Interpretable deep learning

(1)  Core of artificial intelligence: learning and reasoning

The intrinsic feature of AI lies in its abilities of never-ending or self-taught learning from data and experiences, intuitively reasoning, and self-adaptation. Reasoning, the basic form of thinking and simulation, is a process of obtaining new judgments or conclusions from one or several given judgments or premises. Reasoning could be categorized into deductive reasoning, inductive reasoning, analogy reasoning, presumptive or abduction reasoning, causality reasoning, synthesis reasoning, etc. Early studies towards reasoning started in areas of logicism and knowledge engineering. Logicism resorts to the formalization method to represent the objective world, e.g., using the first-order logic or predicate logic to conduct reasoning. Recently developed knowledge- graph reasoning, memory-driven reasoning, multi-agent reasoning, and cross-media synthesis reasoning receive more and more attention of researchers.

Intelligent learning follows three main paradigms: ① learning with formalization methods that first represents rules by symbolic logic and then conduct reasoning. ② Statistical learning that could be considered as learning from data or the so-called data-driven learning. First, we can construct supervised learning process on large-scale labeled data samples, such as deep learning methods. Moreover, using data distributions or prior knowledge, Bayesian learning can be effectively performed on small-size data samples. ③ Learning based on cybernetics—i.e., self-improving from experiences—that could

《Table 1.2.1》

Table 1.2.1 Countries or regions with the greatest output of core papers on the “radar stealth technology”

《Table 1.2.2》

Table 1.2.2 Institutions with the greatest output of core papers on the “radar stealth technology”

《Figure 1.2.1》

Figure 1.2.1  Collaboration network among major countries  or regions in the engineering research front of “radar stealth technology”

be considered as the question-guided or feedback-guided learning, such as reinforcement learning.

(2)  Deep learning: black box versus interpretability

Deep learning methods, such as deep convolutional neural networks (CNNs), recurrent neural networks (RNNs), generative adversarial network (GAN), and deep reinforcement learning (DRL), have recently achieved outstanding predictive performance in a wide range of applications, including visual object recognition, speech recognition and synthesis, and natural language processing. There has been an explosion of interest in interpreting the representations learned by these models.

Currently, deep neural networks obtain high discrimination

《Figure 1.2.2》

Figure 1.2.2 Collaboration network among major institutions in the engineering research front of “radar stealth technology”

《Table 1.2.3》

Table 1.2.3 Countries or regions with the greatest output of citing papers on the “radar stealth technology”

《Table 1.2.4》

Table 1.2.4 Institutions with the greatest output of citing papers on the “radar stealth technology”

power at the cost of low interpretability of their black-box representations. For example, the end-to-end learning strategy makes CNN representations a black-box. Except for the final network output, it is difficult for people to understand the logic of CNN predictions hidden inside the network. Though dramatic success in deep learning has led to a torrent of AI applications and systems that can perceive, learn, decide, and act on their own, the effectiveness of these systems is limited by their current inability to explain their rationale, characterize their strengths and weaknesses, predict their applicability on new task, and even convey an understanding of how they will behave in the future. A high-model interpretability may help people to break several bottlenecks of deep learning, such as learning from very few annotations, learning via human– computer interaction at the semantic level, and semantically debugging network representations.

(3)  Interpretable deep learning: data-driven and knowledge- guided

As discussed above, deep neural networks’ low interpretability lies in their black-box representations. It is difficult for people to understand the relations between decision outputs and logic of predictions and parameters hidden inside the network. Current studies on understanding neural- network representations with interpretable or disentangled representations fall mainly into five directions: ① visualization of the CNN representations in intermediate networks layers, ② diagnosis of the CNN representations, ③ disentanglement of the mixture of patterns encoded in each filter of CNNs, ④ building explainable models, such as interpretable CNNs and capsule networks, and ⑤ semantic-level middle-to- end learning via human–computer interaction. Moreover, adversarial machine learning tries to explore deep learning methods’ vulnerability and its robustness to adversarial attacks. Recent studies show that learning meta-predictor via imposing explainable rules to the process of constructing adversarial examples could help open up the black-box of deep neural networks and make explanations to the decision boundary of their prediction models.

As a data-driven learning paradigm, incorporating explicit reasoning rules into the deep learning models is difficult and limits their interpretability. Symbolic AI refers to a set of methods based on high-level symbolic problem representations and its goal is to dene intelligent systems in an explicit way that is understandable by humans and thus is interpretable. Recent developed abductive learning connects symbolic reasoning with neural perception. Owing to the expressive power of first-order logic, abductive learning is capable of directly exploiting general domain knowledge, so as to enhance the interpretability of the learning process. Moreover, inspired by the intelligence of human brain, perception or cognition requires brain not only to handle the data of target tasks but also to activate related information stored in the brain. Therefore, attention and memory play an important role in the process of cognitive reasoning, especially for the knowledge extraction, understanding and reasoning from sequential media data of texts, audios and videos. Based on these motivations, learning from external knowledge could be combined with the data-driven learning paradigm if we appropriately incorporate the attention mechanism and memory structures into end-to-end deep learning. Representative methods include neural tuning machine, memory networks, adaptive computation time, neural GPU, neural random-access machines, as well as random accessing to external memories via reinforcement training, and can be taken as new intelligent learning frameworks of combining data-driven and knowledge-guided paradigms. Thus, constructing deep neural reasoning on top of attention mechanism or memory structures should be a promising direction to improve the interpretability of AI’s learning and reasoning abilities.

(4)  Related academic endeavors and projects

Interpretable machine learning has received more and more focuses from academic community. There has been a tutorial of “Interpretable Machine Learning: The fuss, the concrete and the questions” at the International Conference on Machine Learning, 2017 (ICML 2017), organized by researchers from Harvard University and Google Brain. In 2017 conference on Neural Information Processing Systems, researchers from MIT, MPI, Microsoft, Cornell, UW-Madison, Johns Hopkins, JPL, DeepMind, UCSD, and NYU gathered together and organized an Interpretable ML Symposium. The symposium solicited more than 30 papers from areas of deep learning, kernel or probabilistic methods, automatic scientific discovery, safe AI and AI ethics, causality, human–computer interaction, quantifying or visualizing interpretability, symbolic regression, etc. This symposium was designed to broadly engage the machine learning community to discuss how to enhance the interpretability of machine learning.

The defense advanced research projects agency (DARPA) launched the program of “explainable artificial intelligence (XAI)” on Oct. 10, 2017, whose goal is to create a suite of machine learning techniques that produce more explainable models while maintaining a high level of learning performance and enable human users to understand, appropriately trust, and effectively manage the emerging generation of artificial intelligent partners. XAI is one of a handful of current DARPA programs expected to enable “third-wave AI systems,” where machines understand the context and environment in which they operate, and over time build underlying explanatory models that allow them to characterize real world phenomena.

Very recently, on July 20, 2018, DARPA announced its Artificial Intelligence Exploration (AIE) program, a key component of the agency’s broader AI investment strategy. AIE continues DARPA’s five-decade streak of pioneering groundbreaking research and development in AI. Past DARPA investments facilitated the advancement of “first wave” (rule-based) and “second wave” (statistical-learning-based) AI technologies. DARPA-funded projects enabled some of the first successes in AI, such as expert systems and search, and more recently the agency has advanced machine learning algorithms and hardware. DARPA is now interested in researching and developing “third wave” AI theory and applications that address the limitations of first and second wave technologies by making it possible for machines to contextually adapt to changing situations.

The countries or regions with the greatest output of core papers, institutions with the greatest output of core papers, countries or regions with the greatest output of citing papers, and institutions with the greatest output of citing papers on “interpretable deep learning” are shown in Tables 1.2.5–1.2.8. The collaboration network among major countries or regions and the collaboration network among major institutions are shown in Figure 1.2.3 and Figure 1.2.4, respectively.

1.2.3 New-generation mobile communication technology

Speaking of the research paradigm, the new generation of mobile communication technology will continue to advance along the three main paradigms of the past, namely measurement modeling, performance analysis, and system design. Measurement modeling refers to the measurement, characterization, and modeling of the objective physical world; performance analysis refers to given wireless communication system and transmission mechanism, and then analyzes its system performance; system design refers to design and optimizes the mobile communication network architecture given the system design indicators. Moreover, with the advancement of big data, machine learning, and AI technology, how to combine it with mobile communication to achieve interdisciplinary integration development is one of the key research directions of the next-generation mobile communication technology, and even may become a mainstream research and an important branch in the next five years. The key technologies to be solved are: ① co-integration technology of storage, computing, and communication; ② ubiquitous network architecture and ultra-dense heterogeneous network technology supporting massive terminal access; ③ the theory and implementation of ultra-reliable low-latency mobile communication system; ④ intelligent mobile communication system design for scenes and services; ⑤ theory and implementation of data- driven mobile communication system; ⑥ optimization theory and real-time implementation of ultra-large-scale mobile communication network.

《Table 1.2.5》

Table 1.2.5 Countries or regions with the greatest output of core papers on the “interpretable deep learning”

《Table 1.2.6》

Table 1.2.6 Institutions with the greatest output of core papers on the “interpretable deep learning”

《Figure 1.2.3》

Figure 1.2.3 Collaboration network among major countries in the engineering research front of “interpretable deep learning”

Currently, the development priorities of major countries and regions in the world are as follows: ① The United States is committed to the intellectual property protection of the fifth generation mobile communication (5G) core technology, and focuses on data-driven, AI-based mobile communication technology; ② Europe is promoting the use of 5G core technology, such as the core challenges and commercialization of large-scale antennas, at the same time, several countries, mainly Germany, pay special attention to the research of ultra-reliable low-latency mobile communication systems relying on industrialization 4.0; ③ NTT Docomo is the representative of Japan, focusing on a series of core issues of 5G commercialization, such as the implementation of large-scale antennas.

《Figure 1.2.4》

Figure 1.2.4 Collaboration network among major institutions in the engineering research front of “interpretable deep learning”

《Table 1.2.7》

Table 1.2.7 Countries or regions with the greatest output of citing papers on the “interpretable deep learning”

《Table 1.2.8》

Table 1.2.8 Institutions with the greatest output of citing papers on the “interpretable deep learning”

Currently, research on next-generation mobile communication technologies based on 5G focuses mainly on the following mainstream directions and branches:

(1)  Channel measurement and modeling

To meet the needs of enhanced mobile broadband (eMBB) services, 5G systems communicate in high-frequency bands, so channel measurement and modeling around high- frequency bands is one of the key research directions; high- band channel modeling based on big data is also important in the future.

(2)  Large-scale antenna

The 5G system further uses large-scale antennas to provide spectrum efficiency and support more user access. Currently, large-scale antennas, especially for high-frequency large- scale antennas, still have many problems. For the first time, large-scale commercialization of 5G is carried out, and the realization of the level of challenges is also one of the key research directions.

(3)  Massive connection

For the needs of the Internet of Things, smart cities, smart grids, wireless communication under massive connection scenarios will still be one of the key research directions.

(4)  Ultra-reliable low-latency wireless communication

Currently, there is no theory of ultra-reliable low-latency wireless communication to guide the specific design of this system, so this research direction will become one of the research priorities of the sixth-generation (6G) mobile communication in the future.

Other wireless communication technologies such as drone- based, based on energy harvesting and wireless energy transmission, especially the inter-satellite communication technology based on wireless energy transmission in space scenarios, are also the research focus of next-generation wireless communication technology.

Future mobile communication technologies will be developed in the following key directions: ① co-integration technology for storage, computing and communication; ② ubiquitous network architecture for supporting massive terminal access and ultra- dense heterogeneous network technology; ③ ultra-reliable low- latency mobile Theory and implementation of communication system; ④ intelligent mobile communication system design for scene and service; ⑤ theory and implementation of data- driven mobile communication system; ⑥ optimization theory and real-time implementation of ultra-large-scale mobile communication network; ⑦ the design and implementation of communication systems, especially the integration of terrestrial satellites with terrestrial networks.

The countries or regions with the greatest output of core papers, institutions with the greatest output of core papers, countries or regions with the greatest output of citing papers, and institutions with the greatest output of citing papers on “new-generation mobile communication technology” are shown in Tables 1.2.9–1.2.12. The collaboration network among major countries or regions and the collaboration network among major institutions are shown in Figure 1.2.5 and Figure 1.2.6, respectively.

《Table 1.2.9》

Table 1.2.9 Countries or regions with the greatest output of core papers on the “new-generation mobile communication technology”

《Table 1.2.10》

Table 1.2.10 Institutions with the greatest output of core papers on the “new-generation mobile communication technology”

《Figure 1.2.5》

Figure 1.2.5 Collaboration network among major countries in the engineering research front of “new-generation mobile communication technology”

《Figure 1.2.6》

Figure 1.2.6 Collaboration network among major institutions in the engineering research front of “new-generation mobile communication technology”

《Table 1.2.11》

Table 1.2.11 Countries or regions with the greatest output of citing papers on the “new-generation mobile communication technology”

《Table 1.2.12》

Table 1.2.12 Institutions with the greatest output of citing papers on the “new-generation mobile communication technology”

《2 Engineering development fronts》

2 Engineering development fronts

《2.1 Development trends in the top 10 engineering development fronts》

2.1 Development trends in the top 10 engineering development fronts

The top 10 project development fronts reviewed by the Information and Electronic Engineering Group are summarized in Table 2.1.1, covering electronic science and technology, optical engineering and technology, instrument science and technology, information and communication engineering, computer science and technology, and control science. The annual disclosure of core patents involved in various development fronts from 2012 to 2017 is shown in Table 2.1.2.

(1)    Unmanned aerial vehicles and autonomous driving technology

An unmanned aerial vehicle is an aircraft that relies on the system's own perception, processing and operation capabilities, and changes its own decisions in real time according to the external environment to complete the established tasks. The autonomous driving technology for vehicles is based on high-precision maps, supplemented by intra-vehicle sensing equipment to collect data, makes

《Table 2.1.1》

Table 2.1.1 Top 10 engineering development fronts in information and electronic engineering

《Table 2.1.2》

Table 2.1.2 Annual number of core patents published for the top 10 engineering development fronts in information and electronic engineering