Summary: weekly mtg 20160224 (Byron, Steve, Matt, me) - mobeets/nullSpaceControl GitHub Wiki

I mention Byron's point last week that "We want to make sure that the monkey is generating the null space that he intends to," and then explain my conceptualizing of sources of activity in the null space:

1. No learning.
2. Model error.
3. Yoking.
4. Freedom.

(See prep).

I then explain the cloud hypothesis that works well by drawing it on the board: basically, it's a combination of habitual and the "cloud stays the same" idea we brainstormed up last week.

In this context, I bring up the "re-aiming" hypothesis. (This is a model-error hypothesis.)

Then we discuss correlations between null and row spaces. Byron doesn't like the way I've done my analysis--he suggests CCA, or a comparison of all norms of pairwise changes (over time, for various time jumps). I don't quite get it, but that's okay.

Matt doesn't get why we're talking about this. Basically, we're looking for either correlations between the activity that change over time, or else an asymptote in their activity signaling that they're stable.

Of course, I still need a behavioral metric, which seems obvious now. Cursor progress per time step.

Steve outlines three things we need to do:

1. find the behavioral asymptote (cursor progress, which is cursor movement projected on the target angle)
2. find the neural asymptote (at least small changes near the end? this will hopefully align with the behavioral asymptote)
3. maybe implement IME

Byron mentions how we need to make sure that with these new approaches we're comparing apples to apples (conceptually). So once you've corrected for where learning is happening, e.g., does one model do better than another?

Steve at one point asked for filtering out low angular error trials and seeing if models do worse or better.