Source
Juarrero (2023), LIT-1995, ch. 15, pp. 223–232; read in NOTE-1532. The chapter reports Churchland, Cunningham et al. 2012, Shenoy, Kaufman et al. 2011, Kaufman, Churchland et al. 2014 and Mante, Sussillo et al. 2013; none of these is held or read here.
What was actually shown
As the chapter reports the studies. Single neurons in the early preparatory period are heterogeneous and uncorrelated; dimensionality reduction finds a low-dimensional subregion whose end state predicts reaction time and movement, with rotational population dynamics in the same direction across conditions (pp. 225–227). In the color/motion task, PFC activity lines up along an “axis of choice”, a line attractor, before the dots appear; the sensory axes are transient while the axis of choice persists until the response; and a “selection vector” makes the same input move the state along the attractor in one context and not the other, which Mante et al. call “context-dependent relaxation” (pp. 227–232). The selection mechanism is from a simulation (p. 230).
The book’s reading. The subspace is “top-down governing constraints for the correct response” (p. 230); its persistence “can be interpreted as the standing cause of the response” (p. 230); the axis of choice “might be considered the neural correlate of a prior intention” (pp. 229–230) and the selection vector that of “a proximate intention” (p. 231); behavior “is constrained top down, according to the relevant, task-defined contextual cue” (p. 228).
What could have come out otherwise: the studies could have found no low-dimensional structure, or structure that did not predict behavior. They found both. Nothing reported could have shown the subspace not to constrain, since nothing reported intervenes on it.
What this does not say
- That the pattern causes the behavior over and above its neurons. The book holds that governing constraints are “not spatiotemporally other than” their components (p. 81), so the reading is compatible with a bottom-up mechanism summarized in low dimension.
- That this is control in CLAIM-533’s sense. It is confinement of the state to a task subregion plus context-gated integration of evidence; nothing about sensitivity, amplification or noise.
- That the boundary of the system matters in THEORY-1314’s sense; the studies are of task context, not of a boundary between brain and surroundings.
- What the primary papers say. Read only through the chapter.