Summary: weekly mtg 20160817 (Byron, Matt, me) - mobeets/nullSpaceControl GitHub Wiki

To discuss

  • lack of Lincoln sessions... (18 vs 4)
  • uncontrolled uniform hypothesis: uniform distribution between min(Y1*NB2), max(Y1*NB2), as long as the spike count is valid
  • minimal energy sampling hypothesis: given Y2(t,:), pick Y1*(NB2*NB2') such that the final spike count is nearest the minimal energy point, as long as the spike count is valid
  • proposal: Only present sampling-based hypotheses. Cartoons still work as a guide, but we just implement them by using actually observed latent values on a held-out set.
  • how to illustrate habitual
  • rotation hypothesis

New hyps results

Equation-based screen shot 2016-08-17 at 11 18 58 am

Sampling-based screen shot 2016-08-17 at 11 18 38 am

Discussed

sampling-based minimal energy and uncontrolled go at the very end--maybe even just in supplementary

  • like look, even if you make these more sampling-based, the story doesn't change

last bar plot figure should be more of a summary

  • but still summarize across sessions
  • normalize to tallest bar? normalize from 0 to 1?
    • that will allow you to average across data sets (with error bars)

try patrick's learning (doesn't have to match exactly, just in spirit)

  • otherwise: trial length -> acquisition time
  • this should hopefully fix the fact that lincoln's

IME

  • compare blue in pert to red in int
    • if it's not the same or lower, note by what percent it went towards red in int vs red in pert
  • also by how much the ime reduced the error compared to the red in pert
  • and some way of summarizing this for each sesssion

arrangement of results fig

  • find some way of comparing all three panels, making it clear how they're connected
  • maybe return to earlier version where you just have panels A and C, horizontally aligned, with highlighting to show connections
    • panel B could just be supplementary

Extra

Hypotheses

First approach is pure equation-based. Either small variance (minimal energy), or variance filling the allowable space (uncontrolled).

A. Equation-based

  • Minimal/baseline energy
  • Uncontrolled

But maybe there's some allowable range of null space activity, and so we want to sample observed activity for it. So we split up the trials into train and test, and then for each timepoint we reproduce the observed row space activity. The uncontrolled hypothesis then says null and row have nothing to do with one another as long as the total activity is in within physiological ranges. The minimal energy hypotheses says that, given the produced row space value, null space activity is chosen so as to produce the final value closest to minimal energy use.

B. Sampling-based

  • Minimal/baseline energy
  • Uncontrolled

But then both of these fail. So maybe there's some empirical structure to activity. The question is then how is it adjusted after learning a new mapping. One approach says nothing at all changes (cloud). The other says corrections are made depending on context (cursor-target angle), and then only in the row-space, which are the minimally-required directions (habitual).

C. Learning-focused

  • Cloud (yoking)
  • Uncontrolled (no yoking)
  • Habitual (no yoking)