Summary: weekly mtg 20160526 (Byron, Steve, Matt, me) - mobeets/nullSpaceControl GitHub Wiki
Prep
-
Choosing histogram bin size: Not sure how to choose this. Making it large enough always changes ordering, makes "true" the worst, and leads to Unconstrained being the best--though by this point the histogram is clearly meaningless. See summary plots (those in the last comment) here.
-
Removing factor bound constraint: For one or two sessions I can match performance by replacing the factor bounds with conditional null space bounds, but for most I can get close but not all the way. See summary plots here.
-
IME c-v scores: For Blk1, improvement is less than it were before, but still an improvement. See summary plots here; more in
plots/imeCv
.
Meeting
histograms
- no bins = zero error
- infinite bins = timepoint by timepoint error?
- unconstrained may be lowest because it has a larger sample size
- add in a c-v histogram, not c-v hypothesis
- find way of picking bin size (maybe to do with sampling from c-v estimating the true latents)
- has to get better and then get worse, so pick that max
- use 90% to fit
- maybe separately for each null space dimension? maybe not
ime
- split trials randomly
- error bars
- verify cursor and target size
- (for more info: Patrick's other paper in supplemental info)