Summary: weekly mtg 20160727 (Byron, Steve, Matt, me) - mobeets/nullSpaceControl GitHub Wiki

Summary

illustrative example connect to muscles at all? we go straight from neurons to movement

  • i.e., why is the answer to our illustrative example question (i.e., square vs. circle) important?

Why is answering this question important? And how does our answer connect to past literature?

Past literature

  • Todorov
  • Loeb
  • Diedrickson (two hands to control single cursor, how to distribute action across two hands)
  • Kaufman
  • Wolpert? Fagg and Bartot?

Fig 1B: add output-null axis

Uncontrolled:

  • which view? keep in spikes view
  • are we implementing this right? not obeying non-negative constraints
    • for a given latent prediction, check if there's any spike value that could generate that value; throw out otherwise

Baseline and Uncontrolled are two extreme ends of the spectrum:

  • Baseline/Minimum = small variance
  • Unconstrained = huge variance filling the space (this is a very strong interpretation, that errors will fill this space)

Fig 1D: call this "new", not "learned" Fig 1E: return to white, where you don't color in the full green ellipse

Need to account for non-physiological predictions

  • talk to Matt about linear program to find if latent value is generatable without negative firing or too large firing
  • Steve: bound all spikes with 0 and, e.g., 100 spikes

NEW APPROACH:

  1. Fig1: a. cur Fig1A b. old Fig1A (i.e., A is human and B is monkey)
  2. Fig2 (1x4): a. cartoon spike space B,C,D: Minimum, Baseline, Unconstrained
  3. Fig3: predictions/evaluation a. old Fig1C b. the concept, not the details
  4. Fig4: results

NEW FIGS:

  • task:
  • prediction
  • evaluation

Write as if methods comes after results

email analysis of prev-row dims

Non-ML:

  • Aaron?

  • Geoff OR Cosma

  • Rob Kass

  • send blurb of what I'm interested in

    • five sentences or less
    • or attach SfN abstract
    • cc: Byron and Steve

Planned

  1. PNC talk: who's the third person? And pick a date.
  2. Discuss figures
  3. Discuss hypotheses (especially motivation for Unconstrained)

Hypotheses

Minimal energy

  • note: Strick's paper is on long-term learning, and in their paper, they find that firing rates stay the same but metabolic use goes down...this may actually be support for a habitual result, where habitual patterns of activation require low metabolic use

Unconstrained

  • new name: Uncontrolled
  • question: what theories might predict no dependence of null space on row space? is this even reasonable?
  • note: this is not the "minimal intervention" hypothesis, because that is about deviations from average behavior relative to a fixed goal (i.e., a fixed target location).

Habitual/Cloud

  • references: theory of neuronal group selection (Sporns & Edelman), minimal intervention (Todorov), "good enough" control (de Rugy, Loeb, Carroll)
  • new name: Habitual --> Minimal Intervention
  • new name: Cloud --> Reassociation

Minimal intervention

  • note: the "uncontrolled manifold concept," as used in Todorov's minimal intervention paper, calculates variability during a hold period and shows that there is less variability (i.e., noise) in the output-potent space than in the output-null space. "The minimal intervention principle...states that task-irrelevant deviations from the average behavior should be left uncorrected to maximize performance." Isn't the "average behavior" here then movement for a given target? This is why I see this as a habitual prediction. Key to our Uncontrolled hypotheses is that null space activity has no kinematic dependence. Is this just a motivatingly-impoverished hypothesis meant to highlight the importance of the null space activity being task-dependent?
  • or, does a "task-irrelevant deviation" just mean a change in your average behavior that doesn't have a different value under your objective--which could be, say, speed, or endpoint variance?
  • in any case, "average behavior" must be for a fixed end goal, which is in this case a fixed target location.

Maybe one prediction of this, actually, is that among activity in the null space of the perturbation, the activity that used to move the cursor in intuitive and now no longer does should change, since it's now no longer being controlled, whereas the activity that was also null in intuitive should not change, since it wasn't being controlled before and now still isn't. So we could compare cloud/hab as we did in issue #234.