Summary: weekly mtg 20170503 (Byron, Steve, Aaron, me) - mobeets/nullSpaceControl GitHub Wiki

To do:

  • cite Matt's Cosyne abstract
    • mention that this (my version) is a stronger finding than that one in this particular way...
  • send to Aaron post-discussion
  • are we close to the word limit?
  • scaling factor activity by number of neurons
    • take length where FA loading vector would intersect hypercube as normalizer. eg [0 1] has length 1 but diagonal has length sqrt2

Discussion:

  • Summary:
    • no need to revisit motivation. just an executive summary
      • just revisit what we found, and why it's important
    • Main finding: Neural redundancy is solved differently than muscle redundancy.
  • why do we study redundancy in the first place? what's the contribution?
    • A: these findings gave rise to cost functions, which gave rise to better models of control theory
      • these cost functions have led to a deeper understanding of how muscles are recruited when making movements (i.e., Byron says don't focus on just making better models, say it leads to us gaining a better understanding)
    • maybe an understanding of the neural redundancy problem will lead to similar models of how neural activity is modulated to lead to better control
  • why it seems like a book report
    • Byron says it's when the flow seems organized around the references
    • instead, you want to have each paragraph make a point
      • a more nuanced point
      • or a specific comparison to other work
  • overall, use the same meat i have now, just reorganize it to be around the big ideas, one main point per paragraph, and trim out any tiny points

High-level major points

  1. Summary
    • no need to revisit motivation. just an executive summary
      • just revisit what we found, and why it's important
    • Main finding: Neural redundancy is solved differently than muscle redundancy.
  2. Comparing solutions to neural and muscle redundancy problem
    • go deep
    • e.g., why would it be different for neurons than for muscles? like minimum energy.
    • see: that point where Steve says "this is an appropriate level of detail"
  3. Estimating output-null dimensions?
    • Byron says maybe put this in Methods. This might over-emphasize this as a potential problem.
    • But, it is important to point out how critical it is to know what the null space dimensions are
  4. How generalizable are these results?
    • e.g., longer timescale
    • effect of practice (say, with the same WMP, day after day)
      • this is one place we can talk about learning, but we should still emphasize it's that learning emphasizes the null space
    • e.g., other brain areas
    • mention we don't consider dynamics. would considering dynamics help?
      • okay to leave this out if it doesn't fit somewhere
    • is this just a local area effect vs. inter-area?
  5. Are there advantages of neural redundancy? (CONCLUSION)
    • entropic constraint?
      • maybe activity on manifold is just more probable (i.e., more neural states representing that activity), and activity tends to "relax" to the more probable state
    • discuss other ideas of why it might be useful...
      • e.g., seems weird that you would use all of the redundancy; let's talk about that
    • maybe you have to use the whole repertoire in order to preserve that network state, so you have to use the whole null space to visit all the states
      • maybe cite Litwin-Kumar & Doiron (Nature communications maybe? look for "maintenance")

WASHOUT STUFF

  • with the WMP strategy, what would happen through the intuitive mapping?
    • Steve wants some simulations for this
  • say, conditioned on the same cursor-target situation
    • let's assume he's got two static internal models: the intuitive mapping, the perturbation mapping
    • you've got the intuitive neural activity conditioned on this
    • and you've got the perturbation neural activity conditioned on this
    • during washout, which of those is he selecting from? can we tell these apart? does he have an a-ha moment and just switch immediately?
  • in 20160722, the left target has 3 in one path, 2 in another
    • were they adjacent? which came first?

suggested analysis:

  • for each timepoint in the first few washout trials, reach back in time to your most recent visual feedback, and propagate two internal models forward:
    • the IME from the intuitive session
    • the IME from the last bit of the perturbation session
  • now, compare the angular errors under these two models
    • see photo from whiteboard