Feedback: BCY on BCI (20160503) - mobeets/nullSpaceControl GitHub Wiki
Poster/presentation only:
- Ask the viewer: What would you expect the distribution of null space activity to be? (Force an answer, so the results don’t in the end seem “obvious.”)
- Intuitive mapping activity → “empirical distribution observed during a different experiment.”
- Start with motivations of hypotheses--not the implementation cartoons.
- Unconstrained cartoon is not clear. There are easier pictures to motivate this. (Generally, it's the whole "projecting" thing I think that could be made clearer by making the row/null axes the standard cartesian ones, so it's clear you're just looking at the marginal distribution.)
- New name for “cloud”
- Null space tuning curves should include error bars, or box-and-whiskers.
- Errors in mean and cov should include error bars.
- Show three different colored clouds when you show the true activity distribution (see slide 22 of cartoons slides).
- It’s confusing that there are two “8”s: number of cursor-target bins, and number of null space dimensions.
- Units for null space activity are in spikes/timestep (45 msec).
- In intro, foreshadow that there is one hypothesis that does really well, otherwise it’s too much to take in without knowing that it will pay off.
- Indirect neurons are all in null space, but this isn’t about that.
Abstract only:
- Doesn’t read like there’s an actual concrete result--can’t visualize the main figure from reading what we have. [Aaron, Stephen]
General comments:
- Motivations for hypotheses:
- Unconstrained: optimal control theory
- Minimal energy use: muscle movement studies (Fagg)
- Habitual: minimum intervention
- “Output potent” instead of “row space activity” [Aaron]
- Results in terms of a “null space activity tuning curve” will be a more efficient way of explaining how we’ve assessed these hypotheses. [Emily]
- “Habitual + minimum intervention” instead of “task-dependent” [Will]
Questions:
- What about dynamics? Timepoints are not independent.
- Are null and row space actually orthogonal? PPCA doesn’t necessarily do orthogonal axes. [Byron]
- What if you look at washout period? Does he return to what he was doing before, or still restrict himself to what was working in both previous blocks? [Alan]
- Fit baseline and minimum hypotheses to predict intuitive block, since they don’t use the actual data. This will provide evidence that these are just bad hypotheses--that it has nothing to do with the two mappings framework that we’re evaluating the others in.
- Would things change over days? E.g., a long-term WMP.
- Would these findings hold for an OMP?