Setting non optimized errors in the take_step function #2
Description
Hello there, Mr. John! I'm sorry this is coming so long after you've published your work, but I noticed some things I'm having problems understanding. In Mr. Platt's paper, in the take_step function, he gives a line to updating the error cache, which I assume is usually done by reapplying the decision function to the two support vectors at those indices and getting the error. Instead, you do none of that and actually update the error cache of every support vector excepting the ones that have just taken steps. Is there some reasoning for this? I would really appreciate your input, as I'm trying to write a translation of Mr. Platt's paper in Rust, and I have used your notebook for some guidance