I don't have anything to say that others haven't already said, Professor Kao is truly one of the best lecturers at UCLA and I would highly recommend this class if you are interested in Neural Networks and Deep Learning. Also, the TAs for this class were amazing, especially Tonmoy Monsoor. Tonmoy is insanely knowledgeable about the topic and his discussions were super useful for the homeworks!

Grading:
Homework: 40% (5 homeworks)
Midterm: 30%
Final Project: 30%
Extra Credit: 0.5% for filling out class eval, up to 1.5% for participating on piazza (in a useful way), and some extra credit given on the midterm (final question on the exam is optional extra credit)

Prof. Kao is SUPER clear on concepts. The lectures are well designed, and is really good. However, the homework are so difficult, and always takes like 6-7+ hours to finish. Also, the final is really hard(luckily I found the tricks and did well).

In general, do take this course with Prof. Kao if you want to learn, but do not if you hate challenges. Overall, I like Prof. Kao so I will take his subsequent courses!

This class was enjoyable and interesting up until the final. Kao is a very good teacher. The midterm was relatively easy and the homeworks took a while but were doable if you went to the discussions or reviewed the discussion material. However, the final was impossible, with an average of 65. The curve was not enough to account for the horrendous final.

Jonathan Kao is my favorite professor at UCLA. Highly recommend this organized, fair, and passionate lecturer.

The class is good overall, and Kao is indeed a good professor. Kao’s lecture is very clear and well-paced, and he often has time to answer any questions people may have. At the beginning of the class, Kao warned us about the workload of the class, but I found it pretty manageable. The concepts of this class are very well explained, and Kao even manages to go over some real-life applications (including his own research) to make the class less “theoretical” than it is. I have been to Kao’s office hours only once or twice, and he mainly just went over the HW problems. The piazza of this class is pretty well organized and the TAs are very responsive to questions.
However, I would only give this class a 4 out of 5 overall due to two things: the grading scheme and the exams. Let’s first talk about grading. Kao uses an absolute grading scale, which means that he doesn’t curve unless the class performs really badly overall. This is great in theory because it means that no one is competing against each other. However, in reality, they definitely cannot run a class where everyone gets an easy one, so they inevitably have to try to control the class grade distribution by the exams. But the problem is the TAs kinda just suck at designing exams. The midterm according to Kao and the TAs was designed to be fairly easy, but the median score was lower than what they expected. Honestly, I did feel like the midterm was easier than I thought it would be, but a lot of questions were just so easy that they essentially became trivial. For example, for some questions, you can just tell the answer by inspection, so you wouldn’t try to solve it the way they want you to do it. Anyway, I got 5 points back after fighting for regrades. Because of how badly people performed on the midterm, Kao promised that he would replace the midterm grade with the final grade if people scored higher on the final. So then you would think that the final might be roughly the same difficulty as the midterm, if not easier.
The final was extremely difficult and surprising in terms of content. It was so bad that people were still talking about it on piazza after the exam. Content-wise, “modulation”, a concept that showed up on every single final from past years, was not covered. Instead, we got multiple questions about a really niche concept of Fourier Transform that showed up only once on the HW. Also, the TAs were obsessed with putting the exact same proofs that we went over in class on the exams (both midterm and final). If you went to the review sessions, the TAs would “hint” that there were some proofs from the lecture that you needed to know for the exam, which were the same ones that actually showed up on the exam. I don’t know how I feel about this, to be honest. Maybe the TAs were doing this to make certain parts of the exam easier, but I feel like this was really unfair because people who had the proofs on their cheatsheets obviously had an unfair advantage over those who did not. And you didn’t even have to actually understand the proofs, you just had to be lucky enough - or sensitive enough to their hints - to copy down the proofs. In the end, they gave out a very generous curve and basically raised everyone’s grade by 3%.
Also, your HW is worth more than everything else (40% of your grade), so whether or not you get 100% on it affects your grade significantly. For instance, even though I got a 96/100 on the midterm, because my HW average was only ~95%, I still had to get close to 90% on the final to get an A in this class, which was pretty ridiculous. Also, something that nobody tells you is that you can go to the TAs’ office hours and literally ask them to check your answers with their HW solution manual. Again, I don’t think this is fair for people who cannot make it to office hours or simply don’t know about this hidden secret. So if you are taking this class, make sure to take advantage of this.

Professor Kao was honestly a little overhyped. In terms of teaching it was pretty standard he would just go over proofs during lecture that often became math heavy. He seemed nice and considerate after our midterm with an average "lower than expected" and let us replace the midterm grade with our final. However, he only gave us false hope; the final was alot harder than the midterm with an average being around 20 percent lower. I felt like the question asked on the final were obscure and a little unfair given the last four years finals had similar format. Alot of the final also depended on thinking about a specific application that wasn't really emphasized. Also the grading rubric for the test was a bit questionable as it wanted us to solve the questions in a very specific way. Class was pretty interesting getting to learn about filters was pretty fun.

I would highly recommend this class to any interested in deep learning and machine learning. Professor Kao is a very good lecturer and he does an amazing job explaining concepts. I never truly understood how backpropagation worked until he explained it in class. Anyone interested in research/ML should definitely take this class. You will learn so much.

However, the class is not a cake walk. It's actually fairly easy to get a good grade in this class as long as you put in the effort. There is only one exam around week 8, which won't be bad if you pay attention to lecture (our average for the exam was a 95%). The homeworks are the real killer and can take a very long time. You essentially have to build neural networks from scratch using Python and Numpy.

Overall, this is an amazing class where you can truly learn so much, but at the price of many hours of homework. Professor Kao is probably one of my favorite professors I have ever had at UCLA.

Be prepared to spend 20+ hours a week on the homework assignments. I learned a ton from this course. It makes it to where AI/ML is not a black box anymore. You can understand how things are working and how it all comes back to the math.
The lectures are very good. The professor and TAs are very helpful. It is a great course which I would recommend if you are single and have the time.

I will start this review with the facts, then my opinion.

The midterm had a mean and median of 75.33, 79.0 (out of 100), whereas the final had a mean/median of 65.6, 66.3 (out of 100). Here is Professor Kao’s final announcement regarding the grades in the course: “I relaxed the overall final grade scales as follows. 89-100 (A), 86-89 (A-), 82-86 (B+), 79-82 (B), 76-79 (B-), 71-76 (C+), 68-71 (C), 65-68 (C-), 60-65 (D+), 57-60 (D), 54-57 (D-), lower than 54 (F). Under these relaxed grade scales, the median grade in the course was a B+.” Kao has the following grading allocation: 40% HW, 25% MT, 35% Final, & 2% extra credit for a survey and piazza contribution. This is the Fall 2024 course.

My overall review of the course is not the greatest. I’d like to first start off by saying that Kao truly is an AMAZING professor. The downfall of this course comes from the exam designs and grading scheme, which are all things the TA’s have control of. I don’t like the TA’s and that mostly has to do with the fact that they SUCK at making exams. The midterm was pretty fair in my opinion despite me not doing the greatest (a tad above median). However, it’s the way the TA’s grade that make it awful. They are EXTREMELY strict on grading and remove points for the dumbest little things. Because of how badly people performed on the midterm, Kao promised that he would replace the midterm grade with the final grade if people scored higher on the final and also said that the final would be easier than the midterm. With that being said, you would expect that the final might be roughly the same difficulty as the midterm, if not easier. WRONG. The final was so insanely brutal as you can tell from the stats of the final that he had to “relax the grades” as he likes to say, to the scale shown above. Although I performed significantly better than the mean/median of the final, I honestly felt like I was having a heart attack during that test. It was trully a mentally extruiating pain that did not sit well for me. The content on the final was just stupid. Simple as that. For one, they made multiple problems on an extremely niche topic on Fourier that we literally only saw once in one part of one HW, and left out actually important stuff like mod/demodulation of singals. Also, the TAs were weirdly obsessed with putting the exact same proofs that we went over in class on both exams. If you went to the review sessions, the TAs would “hint” that there were some proofs from the lecture that you “should know how to derive” for the exam, which were the same ones that actually showed up on the exams. IMO this is not okay as it makes it really unfair to students who can’t make it to review sessions and OH, as those who had the proofs on their cheatsheet had a huge advantage over those who didn’t.

Discussions were worthless and office hours became infested with students holding up the line to check every single one of their HW answers which was very annoying. It’s okay to check one or two answers but some students would hold up the line and check every single problem which is just a little side thing that the TA’s tolerated. Again, this makes it very unfair to those who can’t make it to those OH’s. If you want to get an answer to a question its more efficient to ask on Piazza because the TA’s do have a very fast response time on there (I’ll give them that). This course is very math heavy, and I am not the greatest in my math ability, but I did manage to get an A. It is feasible. My advice to you: Ask tons of questions on Piazza and do all practice midterms/finals. Doing those practice exams will prepare you for ~80% of the exam content because they always make brand new tests. The other 20% is fair game. Good luck.

I think the winter 2024 offering of C147 was very similar to the reviews I've read in the past:

- There were some very hard/long assignments. However if you used Piazza you should be fine
- The class was a bit more mathy in the beginning than compared to the end
- The professor was excellent!

There are also several comments complaining about the "slow pace" of the class, "over-enrollment" and new 50% Midterm Weight. I'd like to address those and what my views are as a fellow student:

Personally, as someone who skipped M146 and went straight into C147, I appreciated that the professor reviewed some of the more fundamental aspects of DL in the beginning of the course. Furthermore, I'd like to point out that this is after all, a graduate class. Grad Students often don't get to take the particular sequence of courses at UCLA prior to coming here, so it make sense that the professor do some review in the very beginning to make sure everyone is on the same page.

For those that were interested in "practical" machine learning (ie learning one of the popular ML libs), the final project was an opportunity to do that. Personally, I think this class is not a torch-bootcamp and rightfully so: C147 focus on more fundamental knowledge and the theory side of DL. There are also other more coding-heavy classes offered (CS188 in W24 for example) that one can take to gain that kind of experience.

While it is true this was a very big class, personally between taking the class vs not taking the class I'd always choose the former. I definitely appreciated the effort the teaching staff put into making the logistics work for such a large class!

Finally, the median for the class was an A- (partly due to grad class, also because Kao doesn't grade harsh). I really don't understand why people complain about grades at this point considering this is an upper-div and the midterm median was a B.