Skip to content

25/02/2026 4:53 PM - Lecture

⬅️ [25/02/2026 11:17 AM - PACES](<./25_02_2026 11_17 AM - PACES.md>) | ⬆️ [EECS 504](<./README.md>) | [11/03/2026 2:11 PM - PACES](<./11_03_2026 2_11 PM - PACES.md>) ➡️

25/02/2026 4:53 PM - Lecture

Hammersley-Clifford Theorem gives us a way to sample from MRFs?

Jason absolutely just looked for ours and didn't look at most of the people's.

Bubblenets

Highly dependent on which object. passing out of frame? Will this learn biases from the dataset?
What about learning to predict whether you will be able to predict object in one frame based on another? That way we can know whether we need more that one label.
With better segmentation now could we do this directly without needing bubbling.

Bernadette

R&D sits solidly between research and products. Can we take promising parts of research and make them into something that it is clear can be productized.

For models in robotics, can we use something like SAM3 in real time? Offboard? Onboard?

Maybe ask Bernadette about generalization in VLAs.

Increase generalization = reduced complexity of actions

Which student is working on ambiguity in maps

She is building pseudolabels of entire scene graphs and learning over them. Find out who's doing that.

  • [ ] @TODO Email Bernadette about these. Or slack her I don't know.

⬅️ [25/02/2026 11:17 AM - PACES](<./25_02_2026 11_17 AM - PACES.md>) | ⬆️ [EECS 504](<./README.md>) | [11/03/2026 2:11 PM - PACES](<./11_03_2026 2_11 PM - PACES.md>) ➡️