Michael Collins

Dynamic programming assignments can be unintuitive and will take some time if you don't have prior experience. Everything else is just too reasonable (lighter than most CS classes)

Remarkably, at times disturbingly, little. Two problem sets, both assigned near the end of the course and requiring more time to typeset in LaTeX than they did to solve. The main event is the final project, but even that is pretty low-key: the writeup is only six pages long if you're working by yourself, and Vinod seemed surprised when I gave him something in the form of a journal article rather than a lab report. I don't know if MIT students are all crybabies, or if Collins feels compelled to dumb himself down for us (I hope it's the former); either way, don't let the 6 in front of the course number fool you into thinking that the material will make you sweat.

Simply the GOAT. Brilliant: just check out his Google scholar list and Wikipedia page. Down-to-earth: flipped class room format is very intimate and you have plenty opportunities to interact with him Caring: late policy is so so generous and give you home during difficult times of the semester Fair: final (40%) was very well structured, a nice balance of questions of all difficulty levels Handsome: while he's no Channing Tatum, Prof. Collins is a stud Give the GOAT a GOLD.

An enjoyable, informative, and (really, really) laid-back class. Like the title suggests, it's about ML applications to NLP, but since just about every machine learning model anyone's come up with has been used for NLP at some point, it tends to focus on problems Collins himself has worked on or is working on now. This is obviously a Good Thing: he was obviously excited about most of the material in the class (I don't think I've ever heard the phrase "fun algorithm" used more or with such earnestness) and made it exciting to learn as well. He's a phenomenally clear lecturer---I took this class concurrently with ML, and the contrast between Collins' presentations and Jebara's insane ramblings couldn't be clearer. I'd encountered a couple of the models we talked about in this class before and struggled with them; MC made everything effortlessly comprehensible. The slides are also good enough that you can get away with skipping the assigned papers (though you really shouldn't; I stopped after about two weeks but am hoping to catch up during the summer). At times it felt like half the department was taking the class; other days there were only five people there. Obviously he's a big fuckin' deal, and just about everyone in the NLP group, faculty included, showed up at some point to watch the man himself teach. I'm still not sure how many people there were auditors rather than enro…

4 Assignments; we were supposed to have analytical part and programming part for all of the assignments, but ssignment 3 became analytical part only assignment and assignment 4 became programming part only assignment. Midterm and Final

5 homeworks. Each was split into an analytical and a coding section. The analytical weren't too bad, but the coding was quite time consuming. You had to implement certain algorithms to produce specific output. There were generally hundreds of tiny bugs you could make that would drive your performance on the task from like 91% to 3%, and it often felt like a bug hunting exercise. Starting early enough that you have an intelligent design for your system is highly recommended, since you'll generally iterate on part1 in part2. Each one was probably 10-20 hours of coding. midterm + final, both of which were quite doable. However, the types of questions he asks are mostly of the "do something with a toy grammar" type that you're asked in the flipped classrooms. Which means that often understanding the bigger picture is less important that seeing the tricks in the toy examples. Still, with the flipped classroom exercises + cheat sheet, you should do ok on the exams.

This is for the flipped classroom in fall 2014. We were assigned small (10-15 person) sections, and each week had about 60-90 minutes of coursera videos to watch before hand. At the session he would spend 20 minutes recapping the key points from the videos, then maybe 30 minutes we'd be doing questions (very similar to the exam questions), and discussing the answers with him. Lots of opportunities to ask questions and talk with him, especially if you get a less crowded session. (Note that his original plan was that he'd only do 2 of every 3 weeks for each section, and a TA would do the other. However it seemed like he abandoned this after a while, and combined a few sessions so he could do all of them himself.) I really enjoyed the format. Keeping up with the lectures before your session was a bit difficult when you have a million other things to do, but pretty important. I fell behind about 3/4ths through when we switched away from machine translation, and it took me a while to catch up again. It was also a little disappointing that you almost never saw who else was in the class, since a bit of discussion about the homework with other students was quite valuable. There is no textbook, but he has a set of classnotes online that's basically an alpha version of a textbook, along with the slides he used in the coursera video. The biggest issue I had with the class was general non…

weekly online lectures (1-1.5 hours), weekly discussion section (1 hour) 4 homeworks (35%) - each had a theory part and programming part, due separately 1 midterm (25%), 1 cumulative final (40%)

Prof. Collins did a flipped classroom format this semester, and I learned more in this class than I've ever learned in any other upper level CS class. I enjoyed the class and I'd recommend it. I also wish that every CS professor would switch to flipped classroom. Every week, he assigned about an hour and a half of video lectures from Coursera. These lectures are incredibly clear and the great thing about them being online is that you can watch them 80 times over until you understand the material. I usually space out during CS lectures, so this was invaluable for me. Then you attend a weekly (~10 people) discussion session where you do problems taken from previous exams. This was more like him handing out the questions, everyone staring blankly at them for a few minutes, and him explaining the answers. Repeat for an hour. These discussions were led by him and the TAs, and they usually felt awkward, but they were useful. They were supposed to be mandatory but he stopped taking UNIs after a few weeks, and I don't think they really ended up being mandatory. Toward the end of the semester, a lot of people stopped going. He posted textbook style notes about the lectures as well as classroom questions and solutions online after each week. As someone whose programming skills were rusty and mediocre at best, I found the homework to be challenging but not at all impossible. I recommend…

3 HWs (5 were expected in the beginning) 2 Midterms 1 Final