Summary: weekly mtg 20160511 (Byron, Steve, Matt, Aaron, me) - mobeets/nullSpaceControl GitHub Wiki

Discussed

First, Byron asks me what are the things we need to do in order to publish. I list two things:

  1. fit hypotheses using IME.
  2. differentiate results of hypotheses within the habitual class (i.e., habitual vs. pruning)

We mostly discuss IME.

Next, we discuss ways of evaluating the hypotheses without involving learning, or two mappings. Steve comes up with idea of fitting activity from some targets and using it to predict activity for different targets. This could test unconstrained, which is great, because now we'll have a way of testing all non-habitual hypotheses in ways that don't involve two mappings.

Notes

updates IME will make:

  • cursor-target relationship
  • direction he thinks cursor is moving
  • null space

questions regarding ime:

  • does IME estimate depend on null space?
  • cross-validated estimates of IME model?

motivation for ime:

  • want to make sure monkey knows what null space is
  • think about 1st trial of perturbation: his null space activity is still the old mapping's null space

ime to do:

  • evaluate habitual with CORRECTION in ime model
  • evaluate all hypotheses with null space defined by ime
  • fit ime on latents vs. on spikes, compare
  • plots of whiskers and errors for ime fits

byron concern:

  • monkey clearly never learns perturbation mapping as well--are our results solely driven by this?

fit null space for orthogonal movement directions

  • e.g., use targ=0 to fit targ=90
  • but then there's no unconstrained activity

future direction for estimating unconstrained:

  • 2d moving BCI vs 3d moving BCI

unconstrained:

  • null space distribution that doesn't change with task
    • steve says now call this "task invariant"
  • vs. null space distribution takes up full range of possible activity (e.g., even when that space is not in the null of the current mapping)

to do:

  • fit task invariant (hypothesis formerly known as unconstrained) hypothesis to intuitive data using different sets of targets
  • fit intuitive using perturbation