I don't know if anyone is still reading this, but I discovered an interesting "trick" in MATLAB. I'm sure most of you found that computing the feature function values took quite a while. So I decided to try splitting up the training set, so that I calculated all the features for 500 examples at a time (passing only the 500 examples into the function), and just appended the calculated arrays. This was blazing fast compared to calculating the entire set at once...so that's something to try if you have this problem in the future.