unconstrained has no kinematic dependence?

[mobeets]

this doesn't seem to be an assumption of todorov's minimal intervention. in fact, in the 2009 JNP paper he seems concerned about "signal-dependent noise", i.e., row-space-dependent variability.

in any case, there could be changes in mean/variance as a function of row-space-output simply due to the non-negativity constraint at the level of spikes.

on the other hand, you don't want to have unconstrained implicitly assuming the cloud stays the same, so focusing on the intuitive-predicts-perturbation case doesn't really fix anything.

[mobeets]

Basically I really think the Habitual hypothesis is more like the Minimal intervention hypothesis, because it's saying for a given class of movement, the only corrections occur in the irrelevant space.

The Unconstrained is just a sort of toy example that motivates conditioning on task variables.

And this toy-ness of Unconstrained, along with the fact that we don't spend too much time with the baseline and minimum firing models, this leaves us with just Habitual-like models. Is this reasonable? I guess we're not necessarily ruling other hyps out, we're just saying the Habitual-like ones do very well.

[mobeets]

and as far as Habitual being "good enough" vs. "optimal control", I think that's out of scope, since the real distinction there lies with optimality concerns. here, i think calling it a "minimal intervention" is sufficient to cover both camps.