Title
Inverse Inverse Reinforcement Learning
Abstract
Inverse Reinforcement Learning has been proffered as an antidote to the Alignment Problem, the challenge that AI is not truly aligned with our plans and wishes. By Inverse Inverse Reinforcement Learning (IIRL) I mean to identify a complex manner in which human beings are cognitively affected by AI. IIRL might not sound as exciting as catastrophic risk from AI, but unlike the much-discussed AI catastrophe, IIRL might be ongoing and pervasive. Although the concrete form the influence takes is co-determined by a particular form of deception that is basic to most everyday applications of LLMs as well as to, e.g., personalised or care robotics, the effect exists whether or not the alignment is deceptive. By “exceptionalism about AI” I mean features that make it correct to conclude that AI is not like most phenomena that crop up with their ethical problems, but requires a more radical approach. And what I identify as Inverse Inverse Reinforcement Learning poses a challenge that might warrant exceptionalism in this sense.
Bio
Professor Hallvard Fossheim, University of Bergen (institutional page)
Speaker is presenting in person.
Practical
The Tuesdays seminars series are devoted to various topics relevant for the Integreat´s research focus. Presenters from the Integreat community and beyond have 40 minutes to present, followed by a group discussion.
Seminars are open for attendance for everybody.
For those unable to attend in person: https://uio.zoom.us/j/62535744013
Point of contacts:
- Organiser and moderator: Pooya Zakeri, University of Oslo
- Integreat Administrative Leader Maria Dikova, University of Oslo