Dual Neural Network Architecture for Personalisation in HCI - Jan Magyar, MSc.

Jan Magyar, MSc.
Pre-Defense of PhD thesis in AI
May 7,2021

14.30 - Slovak Time zone,
06.30 am Los Angeles time zone
8.30pm - Beijing Time zone
9.30pm - Tokyo Time zone

Broadcast will be recorded

Dual Neural Network Architecture for Adaptive and Personalized Behavior in Homogenous Environments
(PhD thesis)

----------------------------------------------------------------------------------------

As digital systems pervade almost all domains of everyday human life, it is becoming desirable for them to offer dynamic behavior. Such change is required in two ways: first, the system should be able to adapt its general behavior based on previous experiences to arrive at better solutions, and it should also personalize to its users’ needs and preferences.

The submitted thesis describes new reinforcement learning methods and approaches that support system adaptability. The author presents a two-part neural network architecture that enables intelligent agents to acquire a general behavior and personalize it directly form interactions. This approach supports generalization of experiences with homogenous environments, while a single trained agent provides optimal policies personalized to each environment.
The thesis further proposes a reinforcement learning-based system architecture for better system adaptability. This system is tested in the context of education where the goal of the agent is to adapt the learning process so that it maximizes human learners’ gains. While comparable solutions currently exist for knowledge acquisition and retrieval mostly in learning vocabulary, the work looks at how a similar approach could be used in general training and skill development.

The author also describes two robotic systems showcasing adaptive behavior using traditional reinforcement learning methods. One of these systems is a conversational robot that showed more social behavior by adapting to user's preferences. The other system includes a tutoring robot playing a simple game with users. Using reinforcement learning, the robot posed questions to users so that users acquired new knowledge while still maintaining a positive experience of the interaction. The example of both robotic systems is used to point out some shortcomings of traditional reinforcement learning methods, which were addressed in later research.

---------------------------------------------------------------------

Bio :

Ján Magyar, MSc earned his Master of Science degree in Computer Science in 2017 at the Technical University of Košice. His master’s thesis focused on the visualization of textual data in scientific publication corpora. He started his PhD studies in 2017 at the Department of Cybernetics and Artificial Intelligence at the same university. His research focuses on applying reinforcement learning in adaptive systems. In 2018 he spent three months in Japan in the Hiroshi Ishiguro Laboratories at the Advanced Telecommunications Research Institute International on a research stay, where he helped to develop an autonomous robotic system that used reinforcement learning for more social and adaptive system behavior. He is the main author or co-author of multiple papers that have been presented at international IEEE conferences. He designed and is teaching an introductory course in programming in Python, and he served as teaching assistant in several further courses on artificial neural networks and deep learning.

Join Zoom Meeting
https://us02web.zoom.us/j/84005040569

Meeting ID: 840 0504 0569
Passcode: HCI