Talk: feedback from Scabby on 2016 08 26 - mobeets/nullSpaceControl GitHub Wiki

To do:

How would you show certain results are statistically significant? (e.g., the unity-line plots comparing cloud and hab)
Uncontrolled should be uniform sampling in spike space, not in factor space
In the 6x6 histogram view, how would this look if you just showed the intuitive data in this space? Pre-learning pert data?
Why does minimal firing hyp show tuning?

Slides to do:

Slow down on the point that we're finding cursor velocity angle on each timestep
Introduce output-null histogram with "activity level" or "firing rate" or something as xlabel--make sure people understand this slide before showing full 6x6 grid of histograms
Ticks, labels, and units on all axes
Minimal intervention: maybe show the new blob
More detail on covariance error for Rob in PNC talk

Suggestions:

Only present learning hyps for PNC talk
Or combine all 5 hyps in together, with simpler hypotheses somewhat brushed over
Animate transitions in learning hyps to show how the cloud changes

Good questions:

Uncontrolled: Why uniform? Why not Gaussian?
Why use IME if results don't change?
How does error in hypotheses vary as trials go on?
How similar are the two mappings (e.g., in principle angle), both with IME and without
What's the difference here between support and distribution when you're talking about predicting output-null activity?
Pick one target and view its results--do they look similar? Pick the hardest target and easiest target--how do they compare?
How does IME error fall-off with trials compare to performance measures? (i.e., do they look similar?)

Future directions: