Summary: weekly mtg 20160413 (Byron, Steve, Matt, me) - mobeets/nullSpaceControl GitHub Wiki
First we look at my evidence for a case where the IME decoder helps drastically:
Left column is IME decoder, Right column is true decoder. Bottom panel is change in mean for each column (null: 1-8, row: 9-10) from intuitive activity to perturbation, conditioned on activity from the same thetaGrps, both in the perturbation's mapping. Top panel is the average of each row in the bottom panel: blue is average of null (columns 1-8), red is average of row (columns 9-10).
So under IME decoder (and not under the true decoder), we see that changes in the mean row space activity are correlated with changes in the mean null space activity for a sparse set of aiming angles and columns.
- question: how can we relate changes in row space to changes in behavior? more change in row space means more learning, maybe?
- question: do row space pdfs change? (they're just 2d, so I can view them in whole)
Then we went through my cartoons.
- next week: smarter bounds control (don't just bound predictions by the marginal bounds)
- next week: mean shift hypothesis (sample from points in intuitive cloud that match the mean row space activity of the current aiming angle)
- next week: pair cartoons with evidence for/against for each hypothesis
- future-direction: SfN abstract (if we convince ourselves we have enough convincing results in the next few weeks)
- future-direction: sample all points instead of just one for each hypothesis