Summary: weekly mtg 20160721 (Byron, Steve, Matt, me) - mobeets/nullSpaceControl GitHub Wiki

three classes:

  • used to matter, now
  • used to matter, now doesn't
  • didn't use to matter, still doesn't

fit cloud in full space, assess in those two row space dimensions

  • just compare hypotheses in that space
  • assess in the 2d space (the row of intuitive)

pick cloud based on closeness in angle

  • as opposed to norm

also show ellipses of data in that nowIrrelevant space (the 2d scatter)

spikes as supplemental activity

  • in terms of spikes: PCA
  • in terms of factors: like PCA, minus observation noise too
  • see Kaufman: they just use PCA

on using or not using IME in perturbation fitting

two things to focus on (worry about):

  1. needs to be accessible! might not get to hybrid hyps. might be better to put out clear and easy-to-understand with less material
  2. are data too processed? (dim reduction, how we make predictions, learning/second mapping, IME, etc.)

Figure 1: more conceptual, in line with Figure 2

  • 2d neural activity, 1d movement
  • monkey wants to generate one value of movement (v=1)
    • show two points that would both generate this movement: which one does he choose?

output-null and output-potent, then just null and potent

For language check out:

  • Law and Schieber JNP 2014
  • Kaufman