Summary: weekly mtg 20160907 (Byron, Steve, Matt, me) - mobeets/nullSpaceControl GitHub Wiki

To discuss

• stats tests: tuning [?], mean/cov lower for cloud vs. hab [sign test? wilcoxon?]
• nul/row overlap between spikes/latents

Discussed

nul/row overlap: FA is a linear projection, but not an orthogonal one

• check out Byron's FA notes, 8 pts, exercise
• orthog projs are like (P'*P)^{-1}*P'

take svd of loading matrix

• use that U along with fa mean
• to just do orthogonal projection

minimal energetics hyps: do in latents (spikes in supplement)

• minimal firing point: orthogonal point
  • svd on L from FA ->
  • U'*(0 - mu) where mu is FA mean

why do we present hyps in FA?

• want to find the effects describing large changes in data
  • pre-processing step
  • de-noising step
  • we need to compute errors in mean and errors in covariance
    • 10d vs. 88d covariance errors
  • easier to visualize
  • a space that the monkeys can move around in during learning
  • the null space in the same space that contains the mapping

tuning hyp test: anova or cosine fit

write shorter abstract for PNC talk, send out ASAP

• no results, make them curious
• maybe one sentence to hint at what we found

presenting

• maybe for learning hyps, maybe remove implementation page