Prep: weekly mtg 20160223 (Byron, Steve, Matt, me) - mobeets/nullSpaceControl GitHub Wiki

What I learned this week

  1. Cloud hypothesis is the best so far, both in mean and variance. (Note: ignore "kinematics mean" below. It had a bug.)

  1. Row and Nul activity are always very correlated. (No sign of consistent directionality of correlation change as trials go on.)

Sources of null activity

  1. No learning: Even if monkey knows model, he may not care to change what he's doing (e.g., if perturbation is close enough to intuitive for a given angle that he decides he's good enough). LOOKS LIKE: Cloud.
  2. Model error: Any mis-estimate of the model on the monkey's part will create "unintentional" null activity. LOOKS LIKE: IME?
  3. Yoking: Even if monkey knows model, he may not be able to separate Nul and Row. LOOKS LIKE: Vol, Cloud.
  4. Freedom: Given no other constraints, he may choose some residual null activity. LOOKS LIKE: Baseline, Minimum.

Note that ALL good hypotheses so far require draws from intuitive activity: Hab, Vol, Cloud.

And no real way to tell the difference between no learning and yoking.

Re-aiming hypothesis

Do a cloud fit, only sample from a rotated theta. Calculating the best rotation per kinematics angle ends up explaining why 20120601 does so well at intermediate angles, but bad on the 0, 315 cases: in those cases, it looks like the monkey is re-aiming!