All Sessions：

In the recent years, mobile technologies and Internet of Things (IoT) have become popular in many fields. Therefore, AI is widely applied in daily life by using various data collection approaches. Sport is a typical field applying AI to enhance performance and reduce injury. AI is impacting nearly every major sport event such as Olympic Games and professional sport like NBA and MLB. It is clear from the direction of this trend that coaches and athletes are demanding more AI technology to help them developing smarter training method. In this respect, AI in sport won’t be much different than its applications in other fields. Sensor-based system or video-based system are 2 applications of AI in sports that hold great promise for future growth. AI can make the most benefits out of those data collected in sport field or training site, if the algorithm is appropriate to determine the right key factors for athletes and coaches. It’s important to note that most of the applications of AI in sport are still in a pilot phase. In order to reach the need to go beyond just tracking data to converting it to meaningful insights that actually help athletes meet their performance goals. Combined AI technology and sport science domain knowledge is essence for future development of AI in sports.

When we see "internet of things", let's make it an internet of beings. When we see"virtual reality", let's make it a shared reality. When we see "machine learning", let's make it collaborative learning. When we see"user experience", let's make it about human experience. When we hear "the singularity is near", let us remember: the Plurality is here.

1. An overview of the booming investment and industrial trends of AI+ healthcare in China 2. The collaboration framework between private sector and public policy makers in Chinese AI+ healthcare ecosystem 3. An investor’s perspective: Lessons, opportunities, and challenges for AI applications in healthcare

We envision a world-wide voice web where everybody, including the illiterate speaking in rare languages, can easily use voice to ask for information, transact business, control IoTs, and automate compositional web-based tasks. At the core of WWvW is the technical challenge of natural language understanding. Our approach is to combine deep learning with formal programming languages. We use a neural contextual semantic parser to map dialogues to their executable formal semantics, expressed in our new ThingTalk programming language. The technology is demonstrated with an open-source privacy-protecting virtual assistant that can control hundreds of IoT devices and perform other popular skills. Our open-source Genie Toolkit embodying this methodology lets non AI experts create multilingual dialogue agents cost-effectively. We invite participation in building the wwVw through contributions to the Genie Toolkit and our open, crowdsourced voice skill repository, Thingpedia.

We are in a new computing era of domain-specific accelerators, where Google's TPU is a visible example. Building such accelerators calls for broader codesign, not just traditional codesign at the hardware/software interface, but vertically integrated codesign that reaches up to applications and down to materials science and device physics. I'll talk about the balance between science and engineering, about how codesign works in TPUs, and I'll pose some materials challenges looking forward.

Over the last half decade, there have been tremendous advances in new model architectures for machine learning that have led to more powerful and more accessible AI solutions. However, these models have become more data hungry than ever, and specifically rely on massive quantities of labeled training data to learn from. This massive change has led to a shift from model-centric AI, where the primary activity of an ML developer is iterating on models (e.g. feature engineering, architecture and algorithm design, etc.) to data-centric AI where labeling, slicing, augmenting, and managing the data at the center of AI is the key activity and leverage point. In this talk, I'll describe how data-centric AI must be programmatic to succeed, and will review our work on Snorkel Flow, a system for data-centric AI.

AI technology has been widely used in various industries, whether it is smart home or smart transportation, there are a large number of successful application developments. COVID-19 has also accelerated the development of smart medicine. At the same time, artificial intelligence has become a competitive force among major countries, and machine learning models have become larger and larger. The cost of machine learning training continues to increase, and talent training and recruitment are also competing with these countries. In this speech, I will discuss the recent impact and development trends of artificial intelligence technology on the industry, including research, application, geopolitical competition, and talent cultivation.

In recent years, exciting technological breakthroughs in AI research and technology have been witnessed across different domains, ranging from computer vision, natural language processing, robotics, medical diagnosis, game playing, to scientific discovery, among others. While we are expecting an AI-driven revolution to transform our industry and bring something good to our daily lives, it is also necessary to inspect its potential negative influence and possible failure modes that could incur undesirable ethical, societal, social, technological, and environmental impacts. In this talk, I will share the journey toward making AI trustworthy as an industrial practitioner and as an AI researcher, including the following four main topics: fairness, explainability, robustness, and transparency. For each topic, I will provide motivation, case studies, and technical tools for infusing trust into AI technology and regulation. Finally, I will share my vision on how practicing trustworthy AI can create a new ecosystem that ensures responsible, reliable, and safe use of our AI technology and how trustworthy AI can create new business and research opportunities.

Artificial Intelligence shows its great potential, e.g., vision, speech and language recognition and creation. However, AI of other brain functions, such as reasoning, feeling, strategy, and knowledge learning, is still in early-stage exploration. Since human brain is a graph of 100 billion nodes and 700 trillion edges, we have been building graph computing foundations for more than a decade to explore achieving full-brain functions. Our latest version is the Graphen Ardi AI platform. What can Full-Brain AI technologies be used? For instance, a US consulting firm published a white paper in May 2020 listing Google, Graphen, Intel and Nvidia as potential companies whose foundations will power the advance of future drug development. Graphen Medical has been utilizing the Ardi platform to design vaccines based on the molecular structure of virus proteins and antibodies. By considering nearly a million of possible mutations, now or future, and more than two million of the genetically sequenced worldwide SARS-Cov-2 viruses, we are able to predict functions of mutations and find ways to fight with them. Graphen’s AI Tools for Medicine (Atom) provides protein structure, functioning, and binding prediction, and Small Molecular and Biosimilar Drug Development. It therefore powers large-scale document reasoning for Whole Genome Analysis applications. Graphen Financial is impacting industry through Virtual Finance Agent, Bank Monitoring and Cybersecurity, Non-Performing Loan Prediction, Fraud Detection, Money-Laundering Detection, etc. Graphen’s AI platform shows outstanding performance at car fixing diagnostics and solutions, reaching high accuracy of fix suggestions with small training set. It can be also used for Renewable Energy Monitoring and Customer Service Agents for any industry. We can see exciting potential of such Full-Brain AI platform. Now is just the beginning.

The world is vast, and can hardly be described by a small sample of real images and labels. Not to mention that acquiring high-quality labels for AI training is both time-consuming and expensive, and sometimes infeasible. With synthetic data, we can fully capture a small but relevant aspect of the world in perfect detail. In our case, we create large-scale store simulations and render high-quality images with pixel-perfect labels, and use them to successfully train our deep learning models. This enables AiFi to create superior AI platform to automate all stores of the future at massive scale.

Recent studies constantly demonstrate that the accuracy of deep learning algorithms in detecting disease from medical images is equivalent to health-care professionals if not over perform doctors. On the other hand, doctors taking care of the acute patient in the intensive care unit (ICU) facing greater challenges. Resuscitating a critically ill patient require the ability to evaluate all vital organ functions based on lab data, images, dynamic physiological data from bedside sensors to search for the etiology in a limited time. Take Covid-19 infection as an example, clinical presentations are affected by the variety of virus, virus load, as well as patients’ age and genetic makeup. All these variables will modulate the translation of protein (cytokines) that orchestrate the circulatory shock, acute lung, kidney and liver injure and coagulopathy. To cope with such a tremendous amount of high-speed, diverse, and unclear information, sometimes, it is beyond the human knowledge and ability to make a precise diagnosis and treatment. A precise and personized treatment requires meticulous calculation of the risk and benefit of each medication and dosage that incorporate the dysfunctional organ information and its interaction with the unique individual makeup. We desperately need an AI to continuously gather the massive and complex physiologic data to identify the distinct phenotype and unique biorhythm to formulate an optimal treatment plan. AI definitely will transform our medical practice, from Empirical to Precision Medicine, from population constructed evidenced based medicine to personalized treatments. Most of all, AI will become an important member of our team in ICU.

Stereotactic radiosurgery (SRS), a validated treatment for brain tumors, requires accurate tumor contouring. This manual segmentation process is time-consuming and prone to substantial inter-practitioner variability. Artificial intelligence (AI) with deep neural networks have increasingly been proposed for use in lesion detection and segmentation. We validated AI-assisted contouring in a clinical setting. Less-experienced clinicians gained prominent improvement on contouring accuracy but less benefit in reduction of working hours. By contrast, SRS specialists had a relatively minor advantage in accuracy, but greater timesaving with the aid of AI. Deep learning neural networks can be optimally utilized to improve accuracy and efficiency for the clinical workflow in brain tumor SRS.

The self-driving vehicle team at MMSL, ITRI completed all requested tests at Taiwan Car Lab in Tainan in July 2019 and obtained the first permission to test self-driving vehicles on public roads in Hsinchu, Taiwan in September 2019. The team also obtained the permissions to test self-driving buses in Taichung in October 2020 and has completed the first stage operation verification in November 2020. This talk will describe the experiences and lessons learned from over one year public road testing and operations of self-driving vehicles in Taiwan.

The post-epidemic era accelerates the adoption of AI in 5G-connected edge devices. The success of AI relies on big data and powerful computing capabilities. Behind the benefits of AI, it is often accompanied by personal privacy and security issues. To further enhance the user experience such as real-time response and ubiquitous availability and address the concern of privacy and security, the market demands high performance connected edge devices. For smartphone devices, mobile AI is evolving from face recognition and object detection to image enhancement. For smart TV, AI function is also evolving from scene detection and image segmentation to pixel-level super resolution. It becomes an irreversible trend in daily life both at home and workplace. In a typical application such as a conference call with virtual background, multiple computing and communication functions operate at the same time, including camera and display image processing, AI for image segmentation, video encoding/decoding, 3D graphics rendering, and low latency wireless transmission. The applications bring higher challenges to the chip design, especially in computing power, thermal, and memory bandwidth management. This presentation will address the opportunities and challenges brought by AI+5G technology on chip design.

Recently, there have been rapid advances of anomaly detection based on the deep learning framework. In this talk, I will first give a brief overview of different deep anomaly detection approaches, such as supervised, semi-supervised, weakly supervised, and unsupervised approaches with various formulations. Then, I will present our recent works on how we develop the deep anomaly detection techniques to various real-world computer vision problems, including defect detection from images, face anti-spoofing, video surveillance, and abnormal human action detection.