29/04/2026 2:00 PM - CQE Practice Sessions
29/04/2026 2:00 PM - CQE Practice Sessions
Austen
29/04/2026 2:01 PM
Slide 1 better motivation for why capabiltiies are increasing?
Why is it still an open problem in our domain?
Slide 6: Problem: Learning to escape local minima. Ensure this is still what we are focusing on.
Maybe explain better on slide 6 what it means to be between fully human and fully autonomous. That would frame better for the next slide.
I was able to predict case based reasoning so that's good.
How do we define similarity?
What baout going back and forth bewtween local minima?
IL and RLHF not data efficient enough.
Slide 10 confusing on data efficiency.
Is this something like "learning to immitate"?
Don't understand case-base and retrieval on slide 14.
Next slide clarifies. Although still these are tighly connected problems. Case-base relies on the type of retrieval you want to do. It is a bit of a confusing separation.
Is binary polar histogram really expressive enough?
Would like to see future work. Say that you will cover in future work.
Figure for jaccard similarity? slide 10.
The explantation of the situation didn't explain parameterizing solutions.
This is revision step I think.
How do you tell if there is simply no similarity strategy? When to ask the human for help?
Can we see number of escape examples versus success rate?
Has he explored a variety of domain adaptation approaches?
25: Present median as well as mean? What is the distribution of time
Is there a way to say "this case is similar in a general sense but includes other things, but this case is less similar, but much more specific to the actual current situation"
Can you define similarities between cases in the DB to make sure you don't re-try too similar strategies. As DB grows this would become a larger and larger problem.
Why straight line? Did you try other intervention types? Why not do something like learn to place virtual potential?
What are you thinking about for better similarities?
Heirarchical methods? Rough to fine?
At the beginning weren't you saying local because we don't know the map. How does that relate to the rest of the work?
Why did the term domain adaptation never come up? It did on prior work.
For Research Gap need to say that there is not past work in using human knowledge to escape local minima using IL. That wasn't clear. I though that there were when you introduced those methods.
Parameterization of goal was unclear to me and made it confusing. What does intervention mean earlier on.
Other comments:
Miranda says: More better figures. If somebody zones out for 30 seconds, will they be lost?
Connor says: Keeps going back to background after other stuff. Quals isn't conference. Backgorund are important. Make sure your research gap is well motivated. "We need to throw babies into space because nobody has done it before" isn't a research gap. Approach summarizes background knowledge right before methods. Keep in mind color blindness. Words on slides work better if they are short enough that you can just read them out loud.
Zach: Balance boring versus lying.
Miranda
I'm getting the intro.
Non-compliance is emerging as the main problem.
Specifically when saying technical terms, slow down. Compensatory motion.
Breathe more. In your script your can put times to breathe.
How do we know that clinicians are optimal? Why should we replicated clinician feedback?
If you are showing the examples on 11, perhaps explain them more. I don't know what is being demonstrated with these.
More exploration on what the curve means. Put it in terms of the examples you just gave.
Slide 12 before jumping into the curve, explain what you are trying to model. Probability that we will intervene. How does your modelling approarch match the desires you have for what you need?
Just skip the extra parameters on slide 12.
On 13, perhaps talk about how there are multiple methods, but we use this one because... Jason will ask why not pose estimation models.
Austen brings up a good point. There was not enough of an example. The intro didn't really explain what is going on. Talk about how a person goes in, is given instructions, and then goes home. Maybe put that slide into the story instead of having a slide that explains the problem without context.
Be careful saying "model". Perhaps say psychometric model so that we know that we are not talking about deep learning for people like Jason.
Slide 16, maybe show what the mapping between stimulus and place on the graph is? Also I wanted more motivation on why the step is bad and yours in good.
Slide 17 it's impossible to know what this means. Say that you are validating the procedure by checking how close we can get to ground truth of the statistical model. Jason will understand that. Say that we are trying to find the number of trials we need to estimate the curve to high confidence. Also needs much better motivation of why this is here. Where does this even go as well. Perhaps after 18? Maybe just say "In validation we found that 50 trials were enough to fit the model and increasing the number of trials might decrease performance due to fatigue and inattention"
Slide 18 gives context to 16 and 17. There's a better way to tell this story. Show procedure as it happens? Like run through the actual procedure. Austen was confused about what the actual experiment actually was.
Everyone was confused by slide 17. I think there are just too many parameters used. The only relevant piece of info is the number of trials.
19 Say before we show this that we know that errors within a single therapist are above 1 degree so there is no reason to go lower. Don't put this in backup, this shows you are a scientist.
20 Motivate the compensatory specific models thing earlier. Those came out of nowhere. Was this a question we had in the research questions?
20 Andy says make the lengend bigger
20 What is the dashed line? It's not labelled. Also make it a different color.
There's a statistical answer to Austen's question about why they are equally weighted. The posterior. Prior given by uncertainty like a kalman filter.
Connor questions whether the key takeaway on 20 is interesting. I would say talk more about the fact that strictness is flipped. That is much more interesting.
Katie
Also speaking a bit fast.
Ask about CLASP talks sources.
Why do we have charge ratios and then F/O. Why that change.
Why odes the FIP have a flat region and then fall consistently? What is causing it?
What does dissappearance events mean? Why that name? Will people listening just know?
Bin seems odd on 31. Why have that noise?
40 and 41, can we get statistical analysis on if the density is different? Difference from uniform test?
Could 46 be just due to definitions of events? Selecting data that falls into a smaller area of speeds?
60/61 also statistical analysis for difference between these two?
67 why is a there a blank bullet?
70 are these above and below the solar plane?