SfN comments - mobeets/nullSpaceControl GitHub Wiki

Poster comments:

  • If you keep a high-variance factor from being used in one mapping, and then make it relevant in the second mapping, the monkey is going to use that factor differently [Saurabh, Shenoy lab]
    • SSS might get at this to some extent...
  • How well does cloud perform during the learning block?
  • Is a task more difficult when the allowed output-null variance in that part of the cloud is small?
  • Does this mean we just need to look in a different area?
    • Does this mean the monkey is just using a cognitive strategy to learn?

Related work [Preeya Khanna, Carmena lab]: two BCI tasks, both with same decoder: center-out, and then obstacle avoidance (which is just center-out except you have to move in a curve shape to avoid a little block) - comparing subspaces found using FA in both cases shows that monkeys use different subspaces in the two tasks; obstacle-avoidance task has higher dimensionality - cursor kinematics are similar in both cases (by one metric, at least) - using only the activity projected into the center-out subspace to control the obstacle-avoidance activity makes performance worse - not sure what woud happen if monkey learned the obstacle-avoidance task first, since they could use this same strategy in the same center-out task, but not vice versa

Related work [Sergey, Shenoy lab]: center-out BCI task with random perturbation of mouse cursor to be suddenly in a new spot - in the array, you see a response to this perturbation: some kind of perception of error - but this error response is in the output-null space of the mapping about 50msec before it is in the output-potent space - however, it might be that the odds of the response showing up in the 2 dims you choose to be the output-potent space are small anyway, so it's not like the monkey learns to put this error response in the output-null space...

Related work [Rouse & Schieber]: recording from M1 while monkey does reaching task - they find the 2d plane most associated with the reach location, and another for the grasp shape - it looks like early in the trial, neural activity projects mostly onto the reach-location plane and not really on the grasp-shape plane; later on, this swaps

Email poster to: [email protected]