WHO: Haitham Bou Ammar
WHEN: October 26th, 2016, 17:00-18:30
WHERE: AUB, Nicely hall #412


Transferring knowledge from prior experience to a new problem is a key characteristic of human intelligence, enabling us to continually build upon and refine our knowledge. This process allows humans to rapidly learn new tasks, often with very little training. Over a lifetime of experience (i.e., lifelong learning), it allows us to develop a wide variety of complex abilities across many domains. Facing an unprecedented growth in data, the need for versatile agents which successfully learn across a multitude of domains is ever-pressing. These agents face a trade-off between learning optimally on each task and efficiently across domains. Inspired by human intelligence, I believe that lifelong transfer is key. Not only will lifelong learning allow agents to master tasks, but will also support knowledge transfer reducing sample and computational complexities when learning on new domains. It will also equip agents with the ability of continually building upon and refining knowledge over their lifetime. In this talk, I will introduce lifelong learning as a novel paradigm in machine learning allowing for agents to reuse old knowledge as a building block in new tasks. I will formalize lifelong machine learning as an online multi-task learning one and show how current methods for knowledge reuse fall-out as special cases from this broader setting. I will also reflect upon some of the theoretical and empirical results attained demonstrating the effectiveness of this new paradigm on controlling a variety of complex tasks.

Haitham Bou Ammar is an Assistant Professor at the American University of Beirut (AUB ). Prior to joining AUB, Haitham was a post-doctoral researcher in the Operational Research and Financial Engineering (ORFE) Department at Princeton University. Prior to Princeton, he was a post-doctoral researcher in the Department of Computer and Information Science at the University of Pennsylvania, and a member of the GRASP (General Robotics Automation, Sensing, Perception) lab. His primary research interests lie in the fields of machine learning, artificial intelligence, data mining, optimization, and control theory with applications to robotics. In particular, his research focuses on developing versatile systems that can learn multiple tasks over a lifetime of experience in complex environments, and transfer knowledge to rapidly acquire new capabilities. Haitham received his Ph.D. in Artificial Intelligence from Maastricht University in the Netherlands, with a thesis focusing on transfer for reinforcement learning tasks. His dissertation developed methods for autonomous transfer between control systems for efficient and effective learning.

Photo credit: http://multimediahomeschool.com