Section A ========== Prob 6: ------- Typo: the proof gives an equation for u3. The u_2's in one term in the numerator and one term in the denominator should be u_1's. Section B ========== Prob 1: ------- Q. What is the stopping criterion for the Perceptron training algorithm? A. It should stop when it makes no errors at all on the training data. For purposes of this assignment, you may assume that the training data is always linearly separable and the Perceptron will always stop, after enough updates. Q. Should we allow for offsets (i.e. \theta_0) in the classifier? A. No. The decision boundary should always go through the origin, i.e., \theta_0 = 0 always. Prob 2: ------- Q. I've implemented the SVM in matlab, but it seems to be taking a really long time to train the data from Prob #1. Is this normal, or am I doing something wrong? A. We just did a quick svm_train() implementation and it seems to run fairly quickly (i.e. in seconds) on data from Prob #1. It also seem to be giving the right results as well, so I don't know what's going wrong for you. Try the MATLAB on Athena, if you aren't sure your version is the latest and greatest. Q. When we implement SVM classifier, do we need to include 'c' as one of the input? A. No, you needn't. Q. Should we use the generalized rule with the offset etc.? A. No, you needn't. For this problem, just implement a simple SVM without offset or regularization. Prob 3: ------- Q. In part (a) we are asked to plot the image which was mis-classified. Could you tell me how I should do that? I looked into the data file, and it seems to be some integer values. But how can I plot them? A. look at the strimage.m code in the "data" section of the pset: http://courses.csail.mit.edu/6.867/hw1/ Also look at the link to actual PNG images (also available from pset page): http://alawi.csail.mit.edu/~alawi/6867/hw1/images/