The mission of my research is to answer the question: howcan we provide people with cyber-physical systems (CPSs)they can bet their lives on? Many CPSs are safety-criticalsuch as self-driving cars and critical infrastructures. In thistalk, I will present how my research tackles this problemtowards the goal to enhance the security and safety ofsafety critical CPSs whilst deployed in dynamic, uncertain,and adversarial environments. I will discuss two applicationcases: one is safety assured planning and control for selfdriving cars, the other one is intelligent intrusion detectionfor a water treatment system. For the first application caseI rigorously verify the safety properties of learning-enabledcomponents in autonomous systems and combine formalmethods and optimal control to design safety-guaranteedplanning and control. For the second application case, ldeveloped interpretable symbolic machine learning calledautomata learning to learn hybrid dynamics from dataduring my Ph.D. study. The learned model for a real watertreatment system and intrusion detection results areunderstandable and verifiable for system operators.

Automatically understanding the body posefrom camera inputs promotes many real-lifeapplicationssuchactivityashumanrecognition, autonomous driving,assistantrobotics and sport analysis. This highlydemanding task has seen extraordinaryprogress over the recent years. The success canbe credited to two main factors: effectiveappearance modeling by deep neural networksand the accessibility of large-scale annotateddatasets. However, the current systems are notflawless that still many challenging issues areleft to be alleviated especially when peopleare in complex articulations or several instancesstay close, occluding each other. We argue thatincorporating prior knowledge like the inherentstructure of our body into the network design isequally essential. To this end, in this work, westudy how to design efficient algorithms tojointly optimize the parameters of deep featureextractors and also the probabilistic inferencemodels which encode priors.

As indicated in a famous adage "A picture isworth a thousand words, images usually conveya lot of information. It is desirable to separateinput images into multiple layers via imageperception algorithms for people to understandthem easily. In this talk, I will present our imageperception works, which benefit outdoor visionsystems, multimedia, and healthcare. First, I willpresent our innovative methods to addressadverse weather image restoration for outdoorvision systems. Then, I will talk about our works onthe multimedia era, which separates input photosinto layers with textures, shadows, lanes, saliency.and so on. Lastly, l will also describe our Al-basedhealthcare algorithms.

Using robot arms, or a collection of them, to performvarious tasks is becoming increasingly popular inboth industry and our daily life. Recent advances inmachine learning provide us with an opportunity toemploy innovative learning to reach autonomouscontrol. Most existing neural network based robotcontrol methods are designed based on statisticalcriteria with reasonable average performance, butlacks amechanismtoensureworst-caseperformance. This talk will introduce our research onconstructing neural networks based on knowledge ofthe robot model and apply it to solve robot armmanipulation problems. This approach inherits theadvantage of neural networks in adaption andsatisfies guaranteed stability. It provides a tractablesolution using neural networks to address safetycritical tasks with certified performance

Building intelligent visual systems is essential for the nextgeneration of artificial intelligence systems. It is a fundamentaltool for many disciplines and beneficial to various potentialapplicationssuch asautonomousdriving,robotics.surveillance,augmented reality, to name a few. An accurateand efficient intelligent visual system has a deep understandingof the scene, objects, and humans. lt can automaticallyunderstand the surrounding scenes. In general, 2D images and3D point clouds are the two most common data representationsin our daily life. Designing powerful image understanding andpoint cloud processing systems are two pillars of visualintelligence,enabling the artificial intelligence systems toand interact with the current status of theunderstandenvironment automatically. In this talk, I will first present ourefforts in designing modern neural systems for 2D imageunderstanding,including high-accuracy and high-efficiencysemantic parsing structures, and unified panoptic parsingarchitecture. Then, we go one step further to design neuralsystems for processing complex 3D scenes, includingsemantic-level and instance-level understanding. Further, weshow our latest works for unified 2D-3D reasoning frameworkswhich are fully based on self-attention mechanisms. In the end.the challenges, up-to-date progress, and promising futuredirections for building advanced intelligent visual systems will bediscussed.

Rapid advances in artificial intelligence (Al) have made itincreasingly applicable to support human work (eg. healthcare)However, the achievement of only accurate predictions of Alsystems is not sufficient for their deployment in practice. lf notcarefully designed with stakeholders, Al systems can exacerbateuser experience, and easily be abandoned. Instead, it is criticalthat these systems are designed to leverage the best of humanability, but also assist to overcome human limitations.In this talk, will introduce my work on creating two interactivesystems that augments a machine learning model with humanfeedback in the context of physical stroke rehabilitation therapy1) human-Al collaborative decision making on rehabilitationassessment for therapists and 2) human-robot collaborativerehabilitation therapy for post-stroke survivors. Then, I will shareinsights from the design, development, and evaluation ofcollaborative systems on rehabilitation with therapists andpost-stroke survivors. Finally, I will discuss emerging and futuredirections for my research, exploring core challenges of creatingeffective human-Al/robot collaborative systems.