noise floor estimates

[mobeets]

Estimating the noise floor for histogram error, mean error, and cov error

For each session, errors are computed between the predicted and observed distributions, conditioned on 8 different cursor movement bins. The error for the session is the average of these 8 errors.

For each session, I split the data into half, using 50% as observed and 50% as predicted. Then averaged the errors across sessions. These are the values in the plot under “0.5”. I also split it into thirds, and used the first third as observed and the second third as predicted. This is the “0.33” value. Etc...

[mobeets]

For reference, in blue I've plotted the mean error for fixed repertoire as a horizontal blue line.

[mobeets]

and here's splitting the actual data in two, using half to predict the other (x-axis), and comparing that to using the same number of points from the intuitive session to predict (y-axis).

avg errors across sessions:

hyp   histErr   meanErr   covErr
true  0.0784    0.3761    1.2882
cloud 0.1455    1.0689    1.8017

and actually, there's usually just about the same number of points in the intuitive session, so the cloud's errors here are really close on average to what they are normally:

0.1368    1.0243    1.4649