• Quiz #1: Kalai and Vempala, Efficient algorithms for the online decision problem, COLT 2003.
• Quiz #3: Blum and Mitchell, Combining Labeled and Unlabeled Data with Co-Training, COLT 1998.
• Quiz #4: Dasgupta and Gupta, An elementary proof of a theorem of Johnson and Lindenstrauss, Random Structures and Algorithms, 22(1):60-65, 2003.
• Quiz #5: Taskar, Guestrin and Koller, Max-Margin Markov Networks, NIPS 2003. Most of the quiz questions were written by Ben Taskar.

Comments: Problem 2(b) was significantly more difficult than the rest of the problem set. You need to be able to distinguish an easy problem from a hard problem. Only two students (Chris Miller and Eugenia Lyashenko) gave solutions that didn't require the use of the epsilon-net theorem (and actually made use of Problem 2(a)).

Unfortunately, many students don't know how to write proofs. Even if you don't plan to write proofs for a living, you still need to be able to communicate your ideas clearly. Here is a good set of lecture notes on mathematical writing. Grading was not competitive, so solutions that received the same number of points are, regrettably, not of the same quality.

Your grade is based on the sum of scores accumulated throughout the course. All scores were sorted and clustered to determine letter grade boundaries.

Final exam: Only the final exam was "curved". The curved score was computed by adding to your absolute score the difference between the top student's score and the maximum possible score (the difference happened to be 10 points). All undergraduate students got 5 extra points. (Another "curved score" was obtained by adding 5 points to each absolute score, and then scaling the resulting scores so that the highest score is 25 points. Undergraduate students got 5 extra points added to the scaled score. Both curved scores were pretty close and resulted in the same final letter grades.) Common errors were graded uniformly. We tried to be as fair as possible in grading both the exams and the projects. Deergha had specific guidelines on how to assign partial credit.

In hindsight, I would have changed the grading policy to create more incentives to think independently (e.g., rewarding partial answers obtained independently more than correct answers that were clearly copied from somewhere; giving partial credit to those who thought about problem 2(b) (final theory set), couldn't solve it realizing that it's difficult, and didn't submit any answer, instead of searching for an answer in a book and finding a wrong one).