SCHEDULE
Tuesday, June 18 (second workshop day)
Room: Arch 304
UTC-7 / GMT-7 (Seattle)
09:00 am Welcome Remarks: Timo Sämann
09:10 Invited Talk*: Cameron Tuckerman-Lee
09:40 Invited Talk*: Reinhard Stolle
10:10 Long Orals** (IDs 8, 11, 22)
10:40 Coffee Break
11:00 Invited Talk*: Been Kim
11:30 Invited Talk*: Stefan Roth
12:00 Long Orals** (IDs 21, 17)
12:20 Lunch Break
01:30 Invited Talk*: Bharath Hariharan
02:00 Short Orals*** (IDs 14, 25, 2, 5, 12, 20, 26, 27, 29, 30, 31, 35)
02:30 Coffee Break
02:45 Poster Session ****(All IDs, Find your poster ID here: Accepted Papers) (Poster IDs: #82 - #99)
04:00 Invited Talk*: Ludwig Schmidt
04:30 Best Paper Award (All IDs)
05:00 Closing
*Invited Talks: 30 min inclusive Q&A
**Paper Long Orals: 9 min (Q&A at the poster)
***Paper Short Orals: 2 min (Q&A at the poster)
****Poster size is 84” x 42”: Printing Information
Note: Paper IDs can be found here: Accepted Papers
Cameron Tuckerman-Lee
The Road to Embodied AI
Reinhard Stolle (Co-Author: Karsten Roscher)
Building safe systems with AI – a call to action
Since I have worked in both research and industry, and in AI and safety, I want to help strengthen the bridge between the AI community and the safety community. I will go through some examples of problems that I have encountered in practice and share some approaches
Kim Been
Alignment and interpretability - how we might get it right
Stefan Roth
Accelerating and Evaluating Visual Attributions
Bharath Hariharan
Recognition in an open world
Deployed vision systems (robotic or otherwise) must function in an open world, where they will encounter objects that were not part of their training vocabulary. What should the system do in this case? With current systems the most we can expect is some form of out-of-domain detection. But this is not enough for safety: consider what might happen when a trained self-driving system encounters a snow plow for the first time. In this talk, I will argue that the system should go much further: it should automatically uncover novel objects and adapt the trained system to them, without a human in the loop. A key ingredient in this goal is *discovering* objects and training detectors for them autonomously from unlabeled data (e.g., by simply driving around). I will describe some of the progress we have made on this problem of open-world, unsupervised object discovery.
