TU Berlin

Quality and Usability Lab2019_03_25_Thórisson

Inhalt des Dokuments

zur Navigation

Autonomous Learning of Situated Dialog with Cumulative Learning

LOCATION:  TEL, Room Room 209 (2nd floor), Ernst-Reuter-Platz 7, 10587 Berlin  

Date/Time: 25.03.2019, 14:15-15:00 

SPEAKERKristinn R. Thórisson (School of Computer Science, Reykjavik U. & the Icelandic Institute for Intelligent Machines)


An important part of human intelligence is the ability to use language. Humans learn this in a society of language users, which is probably the most effective way to learn a language from the ground up, assuming a learner has the capacity for cumulative learning. I present a framework which demonstrates cumulative – continuous, incremental, life-long, non-destructive – learning. Our auto-catalytic, endogenous, reflective architecture (AERA) learns from experience. Our AERA-based S1 agent learns situated communication by observing two humans interacting in a realtime mock TV interview, using gesture and situated spoken dialog. At the outset S1 is only given a small set of knowledge – a seed – no information is provided to it about how to resolve anaphora, use co-verbal gestures, take turns, form grammatically correct sentences, or indeed, how to relate words, grammar, and question-answer pairs with the goals or content of the dialog. S1 learns through on-line observation, demonstrating unequivocal and correct interpretation and generation of all of the above: The fluent use of pragmatics, semantics, and syntax of natural unscripted dialog spoken by the human subjects using a vocabulary of 100 words, on the topic of materials recycling (aluminum cans, glass bottles, plastic, and wood). Using a novel reasoning process based around auto-generated causal-relational models of the task-environment AERA becomes able to predict, interact with, and understand complex temporal relationships, patterns, and contingencies in the environment, starting from only a small bootstrapping seed. The resulting behavior, and the knowledge thus acquired, is highly predictable and reliable.


Dr. Kristinn R. Thórisson, Professor of Computer Science at Reykjavik University (RU), has been researching artificial intelligence for 30 years, in academia and industry. His research centers on realtime interactive intelligence, constructivist learning and broad-spectrum cognitive systems. At MIT he pioneered new ideas in the area of real-time virtual robots with unified multi-modal perception and action control. Between 1999 and 2005 he worked in New York as Vice President of Engineering at Soliloquy and co-founder of semantic-Web company Radar Networks Inc., which was funded by Paul Allen’s Vulcan Ventures, and launched the first semantic-Web portal. In 2007 he developed a cognitive architecture for the Honda’s ASIMO robot; more recently he directed the research on a new kind of AI that can learn complex tasks by observation through self-programming and led a six-member European research project on the development of a new type of artificial intelligence that can learn complex work autonomously. Dr. Thórisson has taught advanced AI courses at Columbia University, KTH and Reykjavik University, and consulted for NASA and British Telecom, among others. He has authored numerous scientific papers and sits on the editorial board of the Journal of Artificial General Intelligence and the LNCS Transactions on Computational Collective Intelligence. He is the founder and Director of the non-profit research center Icelandic Institute for Intelligent Machines in Reykjavik, is co-founder of RU’s AI lab CADIA, which he co-directed 2005-2010. Dr. Thórisson is a two-time recipient of the Kurzweil Award.



Schnellnavigation zur Seite über Nummerneingabe