Summary: weekly mtg 20160323 (Byron, Steve, Matt, me) - mobeets/nullSpaceControl GitHub Wiki

Discussed

• hypothesis scores under IME model
  • they help me find a bug where I was multiplying by 0.045 instead of dividing
• simulations: we realize that our hypothesis predictions are already simulations, so we could just score the match of one hypothesis's predictions using another hypothesis's predictions

To do

relationship between learning and hypothesis errors (mean/cov) over time, both IME and not

for given theta

• error over time
• behavior over time

instead of simulating, fit one hypothesis with the others

• also fit it with itself, produced again

datahigh of 8d null space, for all hypotheses

new hypotheses

• somehow use the actual row space activity, like cloud
• steve's idea: use coupling of intuitive space to predict in perturbation, assuming the coupling stays the same
  • some kind of regression of 8, given 2
  • maybe after subtracting out effects of current decoder
  • run regression on full, then run again on residual
  • look at structure between row and null that isn't related to the task

look at CCA of Nul and Row over time under IME

• and all other views of IME things over time