March 31, 2003

Logistic Regression

Pat suggested looking at logistic regression as a natural step from correlation coefficient. It seems quite complicated though.

Posted by torque at 10:09 PM | Comments (8)

March 28, 2003

Brain tuning to do's

Now that there is something to write about, there are some additional things to try.

Finer grid
My grid has been quite coarse, a finer grid may or may not lead to better results but certainly take longer to compute.
Ideally I should run several permutations of the data and compare least squares to correlation coefficient, and both against the p=1/7 statistics.
Subtracting the "common" signal
So far I have not made any effort to locate and subtract a common signal across the prototypes. This may be worthwhile. The idea would be to average all the prototypes together, and then subtract that signal from each prototype. I tried it for the math problems but the statistics here should be able to show whether it makes a difference or not.

Posted by torque at 11:06 PM | Comments (0)

Least squares vs. correlation coefficient

For brain wave classification, which is a better measure of closeness, least-square or correlation coefficient? Today's computation gives some evidence that correlation coefficient might be a more effective measure. I took data from S6 and S7 from an earlier 60-channel EEG experiment and ran a 3 bin classification on monopolar and derived bipolar. Stimuli consisted of 7 words, first, second, third, yes, no, right and left, and were either auditory (S6A, S7A) or visual (S6V, S7V). In every case of the derived bipolar, use of the correlation coefficient increased the recognition rate by at least 0.7σ, in one case up to 1.8σ. Dfferences in rates for the monopolar montage were less conclusive, though, for every data set, classification on the derived bipolar using correlation coefficient performed the best. More details here.

Posted by torque at 9:42 PM | Comments (22)

Installed Moveable Type!

With some help from Brad Lauster, I was finally able to get Moveable Type working on my Leland account. I've been meaning to write software like this for years - so it is nice that it is available for free. There were some issues though...

I had problems after moving the docs and images directory to my WWW directory. The problem was that I unzipped and installed the files into my cgi-bin directory then moved the docs and images directory. Somehow the permissions were all messed up and I was unable to look at files in those directories from the web. I ended up having to make new directories. Then I copied the files into those directories. That worked!

Posted by torque at 2:53 PM | Comments (0)