Learning through Overcoming
Prof. Du ZHANG
(Macau University of Science and Technology)
Perpetual learning, also known as continuous, life-long, or never-ending learning, is an active research direction in machine learning. It is concerned with how to develop computing systems that can automatically, consistently and continuously improve their performance at tasks over time. Specifically, a perpetual learning agent, a long-lived problem-solving system whose knowledge could be of visualized forms, can be defined through its STEP dimensions.
Given S as a set of learning stimuli, T as a set of tasks, E as type of experience, and P as performance metric, a computing system perpetually learns with regard to (S, T, E, P) if the system automatically, consistently and continuously improves its performance P at T, following S and E over time.
There are many types of learning stimuli. In this talk, we focus on treating inconsistent phenomena (uncertainties, anomalies, surprises, conflicts, outliers or peculiarities that manifest themselves at various granularities of knowledge content) as learning stimuli, and adopting inconsistency-specific learning algorithms to refine and/or augment existing problem-solving knowledge through overcoming (explaining, or circumventing) the inconsistencies. Inconsistencies can serve as effective stimuli to learning because they often signify the inadequacies, gaps, deficiencies, or uncovered boundary conditions in an agent's (visualized) knowledge.
We describe an inconsistency-induced learning framework called i2Learning for such perpetual learning agents whose life span is an alternating and open-ended sequence of task-performing episodes and learning episodes. Each learning episode is triggered by an inconsistent stimulus, and learning is essentially embodied in an agent's capability to find ways to overcome inconsistent phenomena stemming from deficient (visualized) knowledge. Through refining its problem-solving knowledge, each learning episode leads to an improved problem-solving performance P for subsequent task-performing episodes.
Du Zhang is Professor and Dean of the Faculty of Information Technology, Macau University of Science and Technology, Macau, China. He received his Ph.D. and M.S. degrees, both in computer science, from the University of Illinois and Nanjing University, China, respectively. Previously he was a Professor and Chair of the Computer Science Department at California State University, Sacramento. He has research affiliations with numerous universities in the USA, UK, Hong Kong, China, Czech Republic, and Mexico.
Professor Zhang's current research interests include machine learning (inconsistency-induced perpetual learning), knowledge-based systems, big data analytics, and software engineering. He has over 200 publications in these and other areas. He has served in various roles on numerous international conferences, and is editor or editorial board member for several journals in the areas of artificial intelligence, software engineering and knowledge engineering, big data, and applied mathematics. Professor Zhang is a senior member of both the IEEE and ACM, a board member of the Society of Information Reuse and Integration, and a member of Upsilon Pi Epsilon and Phi Beta Delta.