Machine Learning

I took (and withdrew from) this class as my 6th in the program.

I was taking this course as an elective. Don’t do it!

This course is awful. The assignments might be worthwhile exercises if you already know ML, but otherwise it’s more like busywork or hazing. In a single report, the timeline for which is about 3 weeks, you must implement then “analyze” and discuss several different models on two datasets. This could be useful, except the course content does not at all prepare you for what’s in assignments. Thus, in order to be able to somewhat genuinely write on these topics, you must spend hours and hours researching outside of the course content to get a handle on what the assignment is asking you to do. Then, you must implement these models, which might take hours to tune/run, and write a report discussing all of these topics you have to self-study.

This genuinely could work as a fruitful learning experience, except you’re also burdened with quizzes and an exam. These are more closely related to the course content, but again totally disjointed from the assignments. So, in the same 3 week span, in addition to the time you must spend on self-studying for and working on the report, you must watch lectures/read/study in preparation for the proctored unit quiz.

I submitted the first report, having spent >40 hours on it, and felt like I hadn’t learned anything. My writing felt like slop, just BS I put down to get through the assignment.

In every other course I’ve taken thus far, whenever I finished a very difficult assignment, I felt a sense of pride, achievement, and fulfillment. Not in ML.

The concept of this course is interesting, but the execution is very flawed. Machine Learning focuses on the experimentation portion of machine learning. How do you interpret the results of a model? Is this model appropriate for this dataset? How do you verify a model's "correctness" for larger scale use?

Unfortunately, the course itself isn't focused on teaching you how to go about doing those things. The reports that make up a bulk of the class work are entirely focused on interpreting and analyzing your experiment results. However, the class has no real introduction/lecture on the metrics that are commonly used or how to appropriately interpret models, leaving us to hopefully stumble into an effective teaching of these on our own.

As of Fall 2025, we now have 10 - 20 page assignment docs - which provide some guidance into what metrics to look into by virtue of mentioning that a particular plot should be included, but even these aren't complete due to a lack of a formal rubric and additional instructions scattered on Ed. Only after the grades for A1 were out, and sample reports from other students were posted, did I get a better idea of what the "expected" interpretations/metrics were.

This approach is certainly exploratory, but combined with the grading lottery others have mentioned, the lack of rubrics + scattered/unclear requirements, and the lack of guidance in appropriate exploration of a model... I found myself learning a lot less than I hoped, and instead spent most of my time guessing requirements and formatting LateX papers and plots to fit a strict page limit.

This was my last course of OMSCS and the one that I was most worried about from reading the reviews on here. The course has seen major changes that have genuinely made it easier to pass the class. However, I still feel that this course doesn't do an adequate job of giving you concrete knowledge of ML.

Pros:

1. Regrade Requests: The course has introduced the ability to correct issues found by graders for half points. So if you receive a 70% on a report and adequately address all issues highlighted by reviewers, then you could receive a maximum of 85%. I personally did not use this mechanic but others seemed to benefit greatly from it.
2. Quizzes: The quizzes allow for 1 sheet of notes and a calculator. You are also given multiple attempts. The multiple attempts alone made it extremely useful for me as I typically used 1 attempt to get a baseline of my current knowledge and gaps that needed to be filled.
3. Exam: The course has done away with the mid-term exam due to the workload of the reports to where we only had a final exam. The final exam is different from the quizzes and I'd argue is easier as it focuses more on concepts centered around the algorithms covered in the class.

Cons:

1. Hidden and known rubric issues: In recent years, the requirements have become more clearly defined. You are told exactly what libraries you need to use, what plots are required, and given SOME details on how to structure the papers. However, the known requirements were honestly difficult to cover. For example, In A1 they required an absurd amount of plots and specified that they must be visible without zooming in. This forced me to drastically weaken my analysis just to include all plots. The known rubrics also went through MANY iterations and changes as the assignment deadline approached which I found frustrating. In A1 for example, the TAs defined a requirement, removed it, and defined it again in the span of a week. The hidden rubric also still exists and graders are using requirements that are not clearly defined when grading your assignment.
2. Grading delays: The grading was incredibly delayed and students had no idea whether they should continue or drop the class as grades for A1 did not come out until days before A2 was due (between these two assignments, that's 25% of your grade in limbo while the drop deadline was closing). The course also tells you that every report is iterative from the previous report in terms of the feedback that you receive. So you're expected to address feedback from A1 in A2 and the other reports. This didn't have a major impact for me and I only had to re-write small portions of my A2 report but this definitely affected other students. Additionally, I found it frustrating that regrade requests for A4 (the final assignment) were not available, with the stated reason being that grading delays pushed the timeline to essentially the end of the semester.

Recommendations:

1. Reports: I personally got 92%, 99%, 100%, and 93% across all reports. You are given the option of doing extra credit for an additional 10% but with the already defined requirements, I found it difficult to manage the page limit as it is and opted out of doing the EC. My recommendation would be to clearly define the hidden rubric requirements as early as possible so that you do well on A1 and are not compromised on A2 with the grading delays. Ed Discussion and Office Hours were my main source of assessing the requirements. The TA's will not give clear answers and will give suggestions. I would treat those suggestions as requirements for your report. Also, the known rubrics will change over time so keep an eye on Ed Discussion for the latest version of these documents. I also think there is a disconnect on the purpose of these reports that is not adequately covered in the known and hidden requirements that trips up some students. The primary goal of these reports is to analyze the algorithms and their behavior on the provided datasets. The objective is not to maximize performance metrics or aggressively tune hyperparameters, but rather to explain why the algorithms behave as they do, interpret patterns and trends in the results, and connect those observations back to the underlying theory. Finally, I believe that completeness is valued more than accuracy which is unfortunate. I found myself constantly cutting or reducing my analysis to meet all of the requirements which made my analysis very light and surface-level. A4 was EXCELLENT in my opinion as it was light on plots and heavy on analysis.

Overall, This course is clearly improving and It's clear that they're taking the feedback from previous semesters into consideration. However, I still think there is room for improvement.

This review is for Fall 2025 term.

Just finished the class with an A. It has been a tough class, probably the most laborious and time-consuming among all I’ve taken so far (see below). One piece of advice to those who are going to take this class: have a LOT of time on hands to spend on this class. No matter what your background is, if you have time to spend on this class you are likely to succeed. Also, when you write your reports, it helps to have a dedicated “Hypotheses” section with a few numbered hypotheses that are drawn from the course material or papers (explicitly sighted) and then in the Results/Conclusion sections you explicitly accept or reject each of these numbered hypotheses.

Background: 7th class in the program (after AI4R, KBAI, ML4T, AI, DL and CN). The material in ML4T, AI and DL was very useful as a pre-courser for ML - most items in this course looked familiar which helped a lot. Prior to that – no/minimal CS but some STEM academic background (from about 20 years ago). I do have some experience with academic writing and LaTeX, both the previous academic background and from current work. Other than that, demanding full-time on-site job, family and other obligations. Taking the program mostly for self-development, not for career change (although – who knows nowadays). I value the courses I take as the ratio of how much new I learned vs. the amount of time and stress it took. For this class, the denominator goes to infinity whereas the numerator, while large, is finite. Hence, it was not my favorite class (so far, the best class for me was AI4R, and by far the worst – KBAI (please do not waste your time with this class)).

The good:

• Professor LaGrow and all teaching staff do their best to encourage the students. You may (and most likely will) feel miserable at times but they don’t want you to. With the number of extra points, reviewer response, etc. anyone who spends the proper time on this course can succeed.
• The material is quite interesting. Maybe not the very cutting edge but essential.
• I obtained (or started obtaining) a very useful skill: using AI code generating tools to produce the code that does the desired analysis. I would not be able to write all the code for all the analysis by myself – nor apparently this is needed any longer.

Not so good:

• I felt that the amount of material was overwhelming. With its breadth, it was not possible to dive deep enough. Every 3 weeks a super large topic was covered, and in these 3 weeks, one had to (preferably) understand what’s going on, write (or generate) bunch of code, run it on two very different datasets (one of which is huge and super noisy), design and run your experiments that should test a few dozen things and then write an 8-page report.
• The course puts an emphasis on academic writing. Although undoubtfully an important skill (plus helps to organize the subject knowledge), I believe in this case this emphasis was excessive. I’d much rather prefer to understand the subject matter deeper.
• There is a lot of debate on the quality of the lectures. Some love this format, some hate it. I would have been OK with it if the lectures were not so long. With the number of things to do for the course, I could not afford to spend several hours a week listening to the lectures. It feels that the same could have been said in a more concise manner.
  As many students complain, the report grading feels random. My grades were 89, 89, 62 (+19 recovered in reviewer response), 101 but the structure of report 3 mimicked that of 1 and 2. The main complaint of Report 3 grader was that there was no a dedicated section with the “numbered” hypotheses (although they were formulated in the Methodology and Data sections of the report). Most likely the need for the separate section was discussed in the OH (that I mostly could not attend) but there was no such requirement in the project description or the FAQs. What I’m trying to say is that I had no feeling at all about what grade I was going to get for the reports (I personally felt that my report 3 was stronger than 1 and 2).

All in all, I’m not regretting taking this course, just wish it was not taking ALL my time (and more).

I took this course in Fall2025 for the third time (got A) and as the last one for my ML track (yes i lost interest in prev versions and withdraw immediately for very obvious reasons) and i agree it's been improved a lot. Now contents are more up-to-date (with some new videos and practical homeworks!), useful and also with enough depth (thanks to reports, though tough, they are really worthy of time). It's really nice to have centralized threads inside Ed and also there're nice people summarizing OH (which could be done by TAs using LLM nowadays to save time for everyone.) Yes the scaled up course is still alive and evolving!

I hope there're more up to date (optional) videos on practical aspects of the whole ML pipeline and also an industry talk on modern infrastructures, monitoring, and risk management, as it's becoming even more frequent and critical nowadays as everyone touches ML from various aspects. i wish the homework instructions could be even more concise, elegant (modularized, not of spaghetti structure so we dnt GOTO previous papges to callback), and with good academic sense (to write a report of 8 pages i hope instructions/OHs all together are clean/neat and LESS than 8 pages). If you can save 5 hr for one student, then it's 5000 for 1000 students in one year, which is highly impressive. Quizs are necessary and well designed, i can't complain. Just the scientific calculator makes me a bit awkward cuz i really dnt have them nowadays...

For incoming students, you should be aware that ML is wider than ANN and deeper than DL (actually Stanford CSMS could skip ML 229 and go for DL 230 directly). So people attending this course should expect a harder mode compared to DL/RL. HWs are necessary as a CS course, and this course is more of experiment report/conf paper writing style, which is a very good training for both research work and phd application (if interested and willing to take the challenge.)

I took this class in fall 2025. My background is in CS but I've never really done ML at all before this.

Thoughts: I enjoyed this class. I feel pretty prepared to speak confidently at ML because it. I think people may feel a little frustrated at the "self learning" and open-ended project but thats what made it truly feel like a graduate class. Also, the TAs and office hours are a great resource.

The lecture format is a little hard to follow along with, I suggest getting the topics from there and then reading the book with those topics in mind.

This review is for Fall 2025

This class was extremely frustrating to me. Unless you've already had an introductory machine learning course before this introductory machine learning course, you're going to have a bad time. They say in the syllabus that you need to be proficient in python, but you need to be proficient in python and two specific libraries. You need to have used Scikit-Learn and PyTorch at a minimum. The other libraries you'll use aren't difficult to learn how to use for your reports, but if you are completely unfamiliar with those 2 libraries, then you're going to put in more work than what I've listed on this review. The reports are fine. They are a time sink to complete well, but they are not the main problem. If you come into this class with minimal knowledge on how to operate the python libraries, you will have to learn the material, learn how to write it up with the library that you will have to learn, and then write up the results in LaTex. The reports don't always cover the lecture content so you'll spend the majority of your time learning about the things to write about in your report followed by quizzes over all of the video lectures and readings. The reports breadth becomes anti-pedagogical.

The TAs and professor are great, passionate, and a treasure trove of knowledge, but after the first report is due, it felt like their time was consumed by grading the reports. Most of my questions went unanswered or answered by other students. Before this stage, I felt like my questions were met with more factual responses. For example, I asked about how a specific cross validation method worked and to see a mini example of it, but I received a "yes this cross validation technique is used for this algorithm" answer. This semester they added the reviewer response so you could edit your report, but I didn't have time to correct the reports due to needing to work on the next report and learn the next material.

In general, the material is great. You will learn a lot. The reports are a fine pedagogical tool, but due to their breadth they are anti-pedagogical to some concepts in the course. You will spend most of your time writing code, waiting for results of your code, writing what you can of your report while you wait, and write about the results of your code. You will rinse and repeat this cycle for many algorithms. My overall recommendation would be to watch the material for the public version unless you are aiming for this specialization for your degree. If you do end up taking this course, you will probably want to watch the public versions first anyways. The class is doable, and you can succeed. My experience boiled down to two factors: the amount of time I had to dedicate to the class and how much frustration I was willing to tolerate writing reports.

This review is for Fall 2025, which is not available yet.

My background: Dabbled in undergraduate data science projects without much theoretical grounding. Bachelor's in liberal arts (majored in computer science).

The good: The spread of the content was useful to touch upon almost all areas of traditional ML: supervised, unsupervised, optimisation and reinforcement learning. However (perhaps because of my lack of background), I felt that cramming optimisation and reinforcement learning along with all the report writing and code needed to set up the experiments meant that I ended up with an only superficial understanding of the content.

The structure of the assignments was also useful to learn how to set up ML experiments, and deal with the messiness of dealing with real life datasets. Definitely learnt the importance of code abstraction and writing my own custom wrappers by the end of the course as there are many re-runs of the algorithms needed to experiment and get the correct plots.

The reading quizzes + final exam forced me to actually sit down and understand the concepts covered, albeit from an ML intuition point of view rather than anything mathematically rigorous.

The not so great: The lectures were great (I appreciated the banter and conversational style of lecture that really tried to build conceptual intuition rather than just going through equations), but unfortunately too verbose especially when there are only 3 weeks between assignments (which includes completing one reading quiz per module, watching all the lectures, coding the experiments, and writing the report). Resorted to ChatGPT to explain and understand the concepts.

While I appreciated the approach to the assignments (see above), I felt that the assignments had too many requirements which sometimes led to assignments being checkbox ticking exercises. For instance, for the SL, UL and OL assignments, we had to run the same analysis twice on 2 different datasets, one being much larger than the other. While perhaps the purpose of this was to teach us how to do things in a scalable manner, it was probably unnecessary in a course which is also trying to teach more fundamental intuition of understanding the algorithms. Perhaps AI would be a more suitable course since it's focused on the implementation of the algorithms.

Note: The course seemed to me to be front-loaded - the SL and OL assignments had the most requirements while UL and RL was more open-ended, while the final was MCQ only with no numerical calculations required.

TLDR: I put a lot of time into this class and got pretty consistently "good" scores (90 to 100+) on reports, but don't feel like that time translated into learning much about ML.

The Good:

• I got a lot of practice with using LaTeX and writing reports. I've never written as much for a class as I did here, and feel a lot more confident in my ability to produce writing
• I now feel very comfortable with sklearn and hyperparameter tuning
• This course forced me to learn to not procrastinate
• The instructor and a lot of the TAs clearly put an enormous amount of time into this course every semester, and are very involved, accessible, and willing to answer questions.
• I learned a lot about the role of hyperparameters in ML
• I really enjoyed a lot of the lectures

The Bad:

• I don't feel that I came away from the class with an ability to have a rigorous conversation about ML or with a full understanding of the fundamentals. More time dedicated to the course translated to better reports, but not necessarily to better understanding of the material that I was there to learn
• Assessments reward a combination of effort and understanding instructions much more than learning or understanding of ML concepts. Succeeding on the reports requires looking through office hours, the FAQ, and Ed to understand the list of requirements being checked off; once you understand those, it becomes an exercise in coding and writing something that meets those requirements (regardless of whether it makes sense or is grounded in reality)
• The above is especially true given the introduction of quizzes and reader response. Getting 100% is more about using attempts wisely and researching answers rather than learning information.
• The TA team seemed poorly recruited or organized. TAs would often answer students' questions incorrectly. Grading was all over the place- on some assignments I got very detailed feedback, while on others it was clear that the TA had not read the assignment at all
• There were lots of big egos on the instructional team. Myself and multiple other students pointed out a glaring issue with one of the two datasets used throughout all the assignments, which had clearly fake data with only linear relationships; the instructors refused to consider the idea that a mistake had been made and just kept promising that the dataset had hidden relationships that would become apparent to us later (it clearly didn't, and was created by some sort of Kaggle AI slop account).
• There seemed to be a general culture of hand-waviness masked as the scientific method
• There were a lot of promises made around materials being released that never were, dates that grades would be received, etc.
• The time required by reports prevented me from focusing on lectures, readings, etc as much as I would have liked to
• Starting early on reports is very important, but the way that information is shared makes it difficult to do so because it's hard to get a clear picture of the assignment until you've attended multiple office hours, lots of questions have been asked on Ed, etc

Tips for future students:

• Take this course when you have a lot of time
• Try to be organized in writing your code. Create your own wrapper functions to cache results that take time to produce - AI can help with this
• Be careful about using the recommended wrapper libraries -- they often caused bugs that took a long time to track down, made it difficult to create good figures, and made it difficult to regenerate/modify figures without re-running time-consuming processes. It was often much easier to adapt code from AI or online
• Take another ML-related course (I recommend ISYE 6501, which I'm taking now) first -- I did fine without previous ML experience, but feel like I might have gotten more out of this course if I'd taken a previous one