Summary: weekly mtg 20160914 (Byron, Steve, Matt, me) - mobeets/nullSpaceControl GitHub Wiki

To discuss

Talk related

• Added spiking noise to minimal/baseline--cool?

• Should I just show the top 2 dims of Output-null? (Or else, how to explain why they shrink to zero? And are uninteresting?)

• Simple hyp results: hists, then tuning? Or skip tuning and say "This is where we're going..." Could cut tuning totally from results.

• Should I present Uncontrolled before Habitual?

• Could call Uncontrolled simple just "Uniform"? Then the later Uncontrolled could still just be Uncontrolled

• Is the normalized error plot (with some hyps ignored) reasonable?

• Conclusions slide okay?

• Why FA: "88d is a high-d space, so we focus on only the top 10 dims as found through FA. This is convenient because the mapping is..."

Not talk related

• minimal point in latent space: using orthogonal projection doesn't look right...
• FA shrinking towards zero not explaining the problem

Basically, FA shrinks towards zero in latent space, but if we remove all null space activity before hand, then the values in those dimensions are already zero. so shrinking overall data towards zero won't suddenly make those values non-zero

Here's the problem:

Let B be the FA plane, orthogonal or not.

Then B*RB and B*NB are not orthogonal, though RB and NB are orthogonal. Okay, so let's make B orthogonal. But then RBs and B*RB are no longer the same space.

Discussed

compute distances in latents and distances in spikes with the minimum firing rate make sure these distances are highly correlated

new minimal firing:

• min | Lz |
• among all the pts in latent space that satisfy the behavioral constraint (cursor kinematics), find the one closest to the origin in spike space

slides:

• preserve bin wheels
• first show one dim
  • then show all
  • then show first hyp and immediately direct attention to first col
    • physically stand in front of other columns
• uncontrolled-uniform and uncontrolled-empirical