Probably shouldn't use a really light marker in a 200+ person lecture hall, but here are my major faults with this class:
Homeworks I heard are completely copied off of a different machine learning course from UIUC. Midterm and Final questions also heavily are influenced from Dan Roth's course.
Overall the course wasn't conceptually hard. However, a lot of the test questions gave almost no partial credit. Eg, if you had the right answer when the question was "Show X concept", almost no partial credit was given; they looked more for proofs than anything else.
Only a two hour final, and the test questions can be written quite poorly / not clearly. If you're a text-learner, this is a great class! If you're a visual learner and want diagrams on your test, expect to be disappointed.
He also doesn't curve his class at all, rather he scales the margins (eg A- at 88% or something similar, and each letter grade following). However, people did well on his midterm (86-ish median), so he made his final alot harder to artificially decrease the grade distro. IMO this would have been fine if he curved the class, but he doesn't.
Also, he doesn't really consider test statistics that much. Everything is straight scale, so your percentiles on each test dont matter.
Overall, good lecturer, however I found him hard to understand in class sometimes, but bruincast helped out a lot (1.5x speed ftw). However his testing and grading schemes could be a lot better. My suggestions are to: make the midterms and finals more consistent in same difficulty, test on more concepts (almost no SVM questions on final iirc), have more diagrams, and have consistent writing on tests. (EG telling students not to ask questions about the test-questions on the final due to unclear language in the final seems wrong IMO. Especially on a test that ends up having a typo.)