CSIS logoCenter for Secure Information Systems

Securing the World's Cyber Infrastructure

Aerial View of the George Mason Fairfax Campus

CSIS Seminar

Towards Trustworthy Decision-Making and AI: Explainability and Safety

Speaker:   Bo Liu, Auburn University
When:   March 30, 2021, 11:00 am - 12:00 pm
Where:   Zoom

Abstract

In this talk, I will discuss my recent work towards trustworthy artificial intelligence, particularly trustworthy decision-making. Many companies are now building self-driving vehicles and medical robots, and the development of advanced autonomous systems is already a billion-dollar industry. These new technologies offer oversight, advanced automation, and autonomous instruments, and they are adaptable to changing situations, knowledge, and constraints. However, introducing new technologies into our technical and social infrastructures has profound implications and requires establishing confidence in their behavior to avoid potential harm. Therefore, the effectiveness and broader acceptability of autonomous smart systems rely on these systems' ability to explain their decisions. Building trust in artificial intelligence (AI) systems is a critical requirement in human-robot interaction and essential for realizing the full spectrum of AI's societal and industrial benefits. This talk identifies two critical factors for establishing the trustworthiness of autonomous systems: explainability and safety. First, to achieve human-level interpretability, I propose new algorithms leveraging symbolic AI and data-driven ML to enable real-world applications. Particularly, I investigate an explainable and data-efficient hierarchical sequential decision-making framework based on symbolic planning and deep reinforcement learning, termed Symbolic Deep Reinforcement Learning (SDRL, IJCAI'2018, AAAI'2019, ICLP'2019). This approach achieves state-of-the-art results on the most challenging Atari games, Montezuma's Revenge, and outperforms other methods by a large margin. Second, to enhance safety and risk-awareness in decision-making, I propose the Mean-Variance Policy search (MVP, NeurIPS'2018, JAIR'2018, ICML'2020, AAAI'2021) algorithm family. Instead of merely maximizing the expected mean of cumulative rewards in sequential decision-making, the MVP algorithm enables a trade-off between the mean and variance by utilizing the Legendre-Fenchel duality. Unlike conventional mean-variance optimization, which often has multi-timescale stepsizes to tune, this algorithm is single-time-scale and can thus scale up easily. Third, I will discuss a “magic” meta-algorithm framework that can “robustify” any off-the-shelf risk-oblivious decision-making algorithms. This line of work has a wide range of practical applications such as control, robotics, e-commerce, autonomous driving, and medical treatment. Join Zoom Meeting https://gmu.zoom.us/j/99051517102?pwd=YTNEc094K3VkazZiRmJVTXpIc1crQT09 Meeting ID: 990 5151 7102 Passcode: 915413

Speaker Bio

Bo Liu is a tenure-track assistant professor in the Dept. of Computer Science and Software Engineering at Auburn University. He obtained his Ph.D. from Autonomous Learning Lab at the University of Massachusetts Amherst, 2015, co-led by Drs. Sridhar Mahadevan and Andrew Barto. His primary research area covers decision-making under uncertainty, human-aided machine learning, symbolic AI, trustworthiness and interpretability in machine learning, and their numerous applications to BIGDATA, autonomous driving, and healthcare informatics. In his current research, he has more than 30 publications on several notable venues, such as NIPS/NeurIPS, ICML, UAI, AAAI, IJCAI, AAMAS, JAIR, IEEE-TNN, etc. His research is funded by NSF, Amazon, Tencent (China), Adobe, and ETRI (South Korea). He is the recipient of the Conference on Uncertainty in Artificial Intelligence (UAI)'2015 Facebook best student paper award and the Amazon research award in 2018. His research results have been covered by many prestigious venues, including the classical textbook "Reinforcement Learning: An Introduction" (2nd edition), NIPS'2015/IJCAI'2016/AAAI'2019 tutorials. He is an Associate Editor of IEEE Transactions on Neural Networks and Learning Systems (IEEE-TNN), a senior member of IEEE, and a member of AAAI, ACM, and INFORMS.