主题演讲人

Prof. John Page

Research and PhD supervision of the University of New South Wales
Editor in Chief of the International Journal of Intelligent Unmanned Systems

Title: Mentoring; A solution to the AI ethics dilemma?
Abstract: The human race is facing a huge challenge in the next decade or so, one that some experts believe could even lead to our extinction. While this may be an extreme view, the work of organisations like The Future of Life Institute, Cambridge University's Centre for the Study of Existential Risk and Oxfords University's Future of Humanity Institute leave no doubt we face a dangerious future. The problem arises from the fact that as machines have become smarter, they have had to interact with us on a more sophisticated level, meaning that it is important for them, and us, to make ethical choices. This has led to bodies like the Institute of Electrical and Electronic Engineering (IEEE) to devote a great deal of time to the management of ethical design. Some complacency still exists, however, because some researchers in the field do not see a problem until the singularity is reached. This is defined as a point when technologically-created cognitive capacity exceeds human capacity. In terms of damage that AI can do to human society this has no bearing. Even narrow AI, which we already have, is changing human perceptions, behaviour and interactions and this will only grow as the capability grows. The problem is that these impacts have grown in an ad-hoc fashion with little oversight or even understanding. This is clearly seen with social media where uses are constantly surprised at its reach and the unexpected, often even to the companies controlling this technology, effects of applying big data algorithms to the content collected. This becomes even more alarming when learning programs are introduced with the specific aim of directing the behaviour of the participants. This leads to the major issue who or what decides the direction these systems should lead us in. There are three obvious alternatives, the system providers or their AI, a universal accepted legal and ethical base or the owner with the device complying to individual desires either through direct input or mimicry. It is the latter of these three approaches that this paper addresses with the device learning from the owner through there interaction. This can be described as mentoring as the AI learns acceptable practice in the same way a human professional learns by shared experience with a more experienced colleague. This strategy provides two major advantages it maintains diversity, vital to the human wellbeing and provide a clear target for sanctions in the event of the machine acting in a socially unacceptable manner. If the fault is inherent in the manufacture of the device the manufacturer is to blame. If. However, the fault lies in the way it operates the trainer is to blame. This has parallels in how we currently deal with consumer machines such as automobiles. This approach is still in an early stage of investigation but it does offer possibly the best solution to Hawkins' optimistic timewise warning "Computers will overtake humans with AI in the next hundred years. When that happens, we need to make sure their goals align with ours".

Prof. Nick Freris

School of Computer Science and Technology at University of Science and Technology of China

Title: Learning in Cyberphysical Systems
Abstract: Cyberphysical Systems (CPS) are very large networks of “smart” devices (possessing sensing, communication, and computation capabilities), that control physical entities. Notable examples enlist Smart Cities, Smart Grids, Intelligent Transportation, Sensor Networks, and Swarm Robotics. This keynote talk will comprise two parts:
In the first part, we will present two methods for real-time learning in CPS: a) Sparse Kernel Density Estimation (S-KDE), with application in online estimation of travel time densities in transportation systems, and b) Sparse Matrix Decomposition (S-MD), applied to online detection and localization of forced oscillations in smart grids.
The second part will highlight the fundamental balance between data transformation and data utility in machine learning. In Big Data applications, a key challenge lies in the fact that the data are hardly ever available in their original form, e.g., due to compression, anonymization, encryption, or right protection. We will showcase methods for exact learning from inexact data, with provable fidelity guarantees, in specific: a) Optimal distance estimation of compressed data series b) Nearest Neighbor preserving watermarking c) Cluster preserving compression, and d) Distributed consensus on encrypted data.

Prof. Jyh-Cheng Chen

Department of Biomedical Imaging and Radiological Sciences, National Yang-Ming University, Taipei
School of Medical Imaging, Xuzhou Medical University, Xuzhou

Title: Limited-angle low-dose CT image denoising using wide residual network
Abstract: Dose reduction of the computed tomography (CT) has become a serious issue in the recent radiological studies. In dental digital tomosynthesis (DTS), reconstruction from limited-angle scanning would lead to significant noise and artifacts. In this study, we constructed and validated an image denoising method for limited-angle low-dose CT or DTS images. For the training process, normal-dose DTS (NDDTS) and low-dose DTS (LDDTS) images of human teeth were acquired. We collected the real data with angular coverage of scanning from -60 to 60 degrees, with a sampling interval of one degree as limited-angle data. We also segmented each slice into small patches for training with modified wide residual network (WRN) for image denoising task.
For the streak artifacts reduction, noise reduction, visualization of the tooth structure, our denoising LDDTS images showed significantly better image quality than those of NDDTS images in terms of signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR), and artifact index as quantitative results. In this work, we trained a modified WRN as an image denoising method for limited-angle LDDTS images. The performance evaluation of the results by visual inspection as well as quantitive measurements shows that our proposed method is comparable to other main stream networks on image denoising.

Asso. Prof. Zhibin Lin

Civil and Environmental Engineering, North Dakota State University, USA

Title: Machine learning, data analytics and information fusion for structural health monitoring
Abstract: Crtical civil structures, including large-scale oil/gas pipelines and highway bridges, are often exposed to aging, loading, and other harsh environment, while these environmental factors in turn could degrade structures over time. Their health state and structural conditions are key for structural safety control and decision-making management. Despite great efforts in physics-based methods for structural health monitoring (SHM), the high level of variability due to noise and other interferences, and the uncertainties associated with data collection, structural performance and in-service operational environments post great challenges in finding information to assist decision making. The machine learning approaches in recent years have been gaining increasing attentions due to their merits to extracting information from statistical representation of events and thus enabling making decision. In this study, the deep learning-based data analytics (Convolutional neural networks and Bayesian belief network) were used to extract structural information and probabilistically determine structural conditions. Different to conventional shallow learning that highly relies on the quality of the hand-crafted features, the deep learning is an end-to-end method to encode the information and interpret vast amount of data with minimizing or no features. Several cases studies were used to demonstrate the effectiveness of the methods for structure under viabilities and uncertainties due to operation, damage and noise interferences. Numerical results revealed that the deep learning exhibits considerably enhanced accuracy for structural diagnostics, as compared to the supervised shallow learning. With predetermined training set, the deep learning could accurately determine the structural health state in terms of damage level with and without noise interference, which could dramatically help decision making for further structural retrofit or not.

Assi. Prof. Bin Zhang

Department of Mechanical Engineering, Kanagawa University, Japan

Title: Development of an AI based Teaching Assisting System
Abstract: Nowadays, people are facing the issues of declining birthrates and an aging society. Teaching resources are far from enough and the teachers usually have great amount of works, especially in countryside or small towns. Aim to reduce the work of teachers, we developed a teaching assisting system to automatically track and analyze the behaviors of the students based on artificial intelligence (AI) technology. Specifically, by observing the behaviors of students without getting tired with the sensor system instead of the teachers, the performances and growing processes of all the students can be analyzed, referring the excellent knowledge and experiences of professional teachers. The mental states of each person can also be estimated from their behaviors. By feeding back this kind of information to the teachers, the teachers can adjust the teaching contents according to the personalities, interests and levels of the students. Here, the mental state estimation model is trained in advance by using deep learning method based on a large amount of observation information such as body movements and facial expressions that can be directly measured and mental status estimation results based on human experiences. In addition, by developing information provision tools, smooth communication and trust relationships between the students and the teachers can be established.