Program

Workshop Day - August 16, One Full Day

Notes for presenters:

• The schedule below is tentative and subject to minor changes.

• Full papers are allotted up to 20 minutes which includes presentation + discussion. Use the time as you wish; the MC will encourage you to take questions after 17 minutes.

• Extended abstracts are allotted up to 10 minutes which includes presentation + discussion. The MC will encourage you to take questions after 8 minutes.

• You are encouraged to use your own laptop for the presentation. HDMI/USB-C inputs will be available.

• If you prefer not to bring your laptop, a laptop is available; please bring your slides on a USB drive or email them to Ben Armstrong (research@benarmstrong.ca)

• If you are unable to attend or have other concerns please email Ben Armstrong!

Invited Speaker - Dr. Nicholas Teh

Title: Learning Something Acceptable: Preference Elicitation for Safe AI Deployment

Abstract: Safe AI deployment often requires satisfying multiple stakeholders with different (and sometimes incompatible) requirements. However, eliciting complete preferences or numerical utilities may be unrealistic; stakeholders may only be able to answer whether a concrete deployment plan is acceptable. This talk develops a theory of learning from simple binary feedback. I will discuss when we can efficiently find a deployment acceptable to everyone, how interaction and query budgets limit what can be learned, and what to do when unanimous acceptability is impossible. This perspective connects preference elicitation, learning theory, and optimisation.