Summary: weekly mtg 20160330 (Byron, Steve, Matt, me) - mobeets/nullSpaceControl GitHub Wiki

External

• talk to Doug (Byron: it could work, but how?)
• email alan about BCY

Covariance Error

• Why is cov error big for single kinematics? (spiky, not smooth)
• understand coverror units
• understand coverror itself

Other error metrics

• can we get a single number?
  • timepoint by timepoint, still binned by kinematics
  • find new error floor, say samping from the same observed but randomly
    • shuffling actuals to get null distribution
  • any method of comparing two distributions
    • matt's method
      • candidate points: simplex points
    • any distance metric, like KL-divergence
      • could it be non-parametric? or parametric

Learning and behavior

• compare behavior by kinematics to errors by kinematics, in single plot

Conditional hypothesis

• isn't just the same as fitting conditional gaussian distribution of (YN1,...YN8 | YR)
  • just fit 10-d gaussian, then find conditional distributions
  • is this is the same?