Prep: weekly mtg 20160518 (Byron, Steve, Matt, Aaron, me) - mobeets/nullSpaceControl GitHub Wiki
Questions
- should we continue fitting by thetaGrps?
- is optimizing tuning parameters by date cheating?
Completed
Fit unconstrained using pairs of targets
so say:
- 0,180 to fit 90,270
- 90,270 to fit 0,180
- 45,225 to fit 135,315
- 135,315 to fit 45,225
Example:
For all dates, no real change in errors! Kinda cool.
Perturbation predicting intuitive
changes:
- for means: ordering mostly preserved, except for 20120709 habitual is better than pruning
- for means: pruning/habitual becomes less better than zero/unconstrained
- for cov: unconstrained usually gets better (all but 20131205), and sometimes is actually the lowest (20120525, 20120601)
For example of cov getting better, see below (top row is normal, bottom row is fitting reverse):
Resampling when null space activity out of bounds
For example:
Histograms of marginals
IME
hypothesis errors
Cloud becomes one of the best hypotheses--other hyps stay the same (mean shift becomes less good).
IME fits
see plots/ime
for all figs. here's an example:
IME fit on latents just adds a constant offset to angular errors.
ratio of mean(abs(ang_errors)) = range of 1.25 - 1.75 difference of mean(abs(ang_errors)) = range of 0.3 - 0.45
so basically it's a 25-75% addition to the errors, but still better than true decoder