Prep: weekly mtg 20160302 (Byron, Steve, Matt, me) - mobeets/nullSpaceControl GitHub Wiki

Hypothesis performance with high vs. low angular error

Compared all hypotheses using trials where abs(angError) < 40 vs. abs(angError) > 40 In this case, EVERYTHING gets worse across the board except Vol's mean error goes down.

If instead you use 20 as your cut-off:

Fitting to high angular errors gives all hypothesis lower mean error. (more trials?) Kinematics-mean always gets better for both covErrors. Vol always gets worse for both covErrors. Habitual's covOrient sometimes gets slightly better (2/4), but not always. Cloud-hab sometimes gets better for both covErrors (2/4), but not always.

Behavioral asymptote

Fit a three-parameter saturating exponential to the cursor progress. The trial where learning is said to be saturated is the trial number where the fitted curve achieves 85% of its max. Anywhere where the start height was higher than the end height was ignored.

Note that for 20131205, cursor progress decreases over time because though the monkey successfully reorients the activity to point on average towards the target, the overall magnitude drops, so cursor progress goes down. so instead of cursor progress we could also look at cursor non-progress? (e.g., you want to minimize the projection onto the vector orthogonal to the target?)

Where the shuffle starts were set before:

each block starts from trial 1 first number is where pete had it starting second number is end of block

20120525: 287 1004 20120601: 193 940 20131125: 95 405 20131205: 125 460

Correlation in changes in activity

Two plans from last week:

1. using CCA, project both Nul and Row into same space with maximum correlation, and plot that correlation over time.

this is correlation after doing CCA between cursor progress and YR or YN

shuffled block only:

this was without filtering of angular error. now, for block 2, on the left is the correlation of correlations with filtering, and on the right without:

2. take chunks of time points and calculate norms of all pairwise differences between activity, for nul and row, separately. do these norms move together?

And now instead of looking for correlations in the norm of the changes, look for maximal correlations (using CCA) in the changes. Red lines indicate where blocks change, so area in between two red lines is the shuffle block. all trials use the shuffle block's null space.

3. estimate convergence of neural activity and behavior (cursor progress)

Each panel shows data for a given kinematics condition. The black lines mark the start and end of shuffle trials. The red line indicates the region of the last 25% of timepoints for a given kinematics condition. During this time, CCA was used to to project the full latent activity and cursor progress into the highest correlated space.

The idea then is that if the activity not in these last 25% of trials, when projected into the space found by CCA, is still well correlated with cursor progress, then the monkey's model has already converged.

Here's a day where it doesn't align as well:

Reaiming

Reaiming now implemented, using intuitive block activity only to fit the rotations, and basically it does better than other hypotheses for the second monkey, but not the first. this kinda makes sense given the perturbations:

(black dots are activity in intuitive block with perturbation mapping--look how they're basically just rotated to the wrong target!)

here's the unnormalized version:

now here's 20120601, by contrast:

Here's all the reaimings: