Predicting null space activity - mobeets/nullSpaceControl GitHub Wiki

Problem set-up

Inputs:

• Y1 [nt1 x 10] - latent activity during intuitive block
• theta1 [nt1 x 1] - cursor-to-target angle during intuitive block
• YR2 [nt2 x 2] - row space latent activity during perturbation block
• theta2 [nt2 x 1] - cursor-to-target angle during perturbation block

Outputs:

• YB2 [nt2 x 8] - null space latent activity during perturbation block

Perspectives

So basically the intuitive block serves as our training data, and the perturbation block as our testing data. But we get additional information: the row space activity.

Intuitively, then, there are three sources of information to predict each [8x1] null space latent activity:

1. Dependence on theta
2. Dependence on row space activity
3. Correlations amongst the [8x1] vectors.

I'll discuss each of these in turn.. (Figures from plot.highLevelCorrelations.)

1. Dependence on theta

Perturbation null space activity projected into space with maximal correlation with (sin(theta), cos(theta)), with points colored by thetaGrp:

2. Dependence on row space activity

3. Correlations amongst the [8x1] vectors

These appear to be somewhat different in different blocks, but still:

Approaches