The concept is not clearly explained at all. The instructor seems to be just reading a script. He did not try to explain the math. Instead, he uses graphs to try to fool us. The other instructor hardly teaches anything but just to show his face and say a few openning setences. I took Andrew Ng's courses and was impressed, but I am very disappointed by the quality of this course. Deeplearning.ai, please have some quality control over the courses you offer, otherwise it hurts your brand name!

The course is rather disappointing. Videos are short. They give you an intuition, why something works but don't go much into the details. When teacher said "Now you are an expert in transformers" it sounds like a mockery. The course material is split into four weeks, however you can obtain certificate after spending a few days.

Homeworks won't teach you much. For you to understand, by now the most hard exercise according to course's Slack is to write a function with model and input tokens as input, which should predict next token. It's body contains only 8 lines of code, some of them is already given, your task is well explained.

Trax, a deep learning framework, that is used in homeworks might be a great framework, but not for learners. All you need to do, is just to pick a layer, put it in right place and

voila. But instructions makes a situation even worse. It is so detailed, that you can just copy a code from instructions, paste it into your code and obtain a working solution. Sometimes you should look at documentation just to see the argument's name. You won't have to think about dimensions, you won't have to think about structure of a model. When you decide to write a transformer from scratch with Pytorch then, you will struggle hard, but the price is much deeper understanding.

Would I recommend taking this course? I think, that course team did a nice work to provide you an overview of the state of the art techniques in NLP. Some references are amazing. So if you treat this course like intorductory, you could take it. But don't expect too much. When you are said, that you will "build a chatbot using a Reformer model" take in mind that the crucial skill to do it, is just a copy-pasting.

Superficial coverage of topics, lack of mathematical depth and sophistication. Dumbing down and simplification never help.

I haven't had similar issues with previous courses by Deeplearning.ai, but with this one I was worried I'm overly stupid the moment I started, because I noted I was "missing" a lot and was not understanding easily what's going on (Note: I have all required background from the ML Course and DL Specialization). Then I saw the existing reviews and was happy to see I'm not alone to feel like that:

- Overly superficial coverage of theory in videos; too many things not explained well (if at all). For example: last week's videos are about... 18 minutes. REALLY? I thought we were talking more complex stuff here. If one can be taught this in 18 minutes, then... oh well...

- Lots of "copy-paste this here" parts in assignments, too (not much thinking/effort required).

- The quizzes are (as in most courses) a joke, they're there just for the sake of it; I just skip them.

- Looks to have been created in too much of a rush; I don't know if that's the case, but that's the feeling I get from the content quality...

Based on the success of the original Andrew Ng courses, the quality bar is high as are the expectations. I hope there is better quality control in future specializations, either in-house or by better selection of external beta-testers. I can't believe several reviewers bring this up, but no one else did before the release.

the NLP spec course has ~30min video on every week, and sum-ups are ~1/4 of it. Thus, one cannot expect a good and profound theory knowledge, only some intuition and insights.Without theory, it can be expected that program tasks should contain something practical and superficial. Like crash-course into the most popular packages in the field. But tasks are huge – x6 time more than a theory – and boring. A lot of spaghetti code with few levels of enclosed IF’s, with constructions like `A[i][j:k][l+1]`, low code reuse, global variables, and `from utils import *`.The student will spend time doing the bad implementation of 100K times implemented things, and that will not provide him with enlightenment on how they are implemented because of a lack of the theory.And nobody will teach him how to use standard tools on simple and understandable examples. It is boring, exhausting, and impractical. And in most cases, students can't do just part of tasks, because the auto checker will raise an error.

To anyone looking to learn the content for the first time, I would suggest by reading the original papers and some blog posts. The videos are short and do not go in-depth much at all. The real meat for this course is in the homework assignments. The videos tend to oversimplify to the point of not explaining the concepts correctly or being flat out wrong and fail to give critical context to fully understand what is being explained. On the other hand, the homework was interesting (especially when compared to other courses out there) and did go in more depth, making students think through the details of some of the algorithms and models. tldr; learn the content elsewhere, take the course for the homework + to learn about trax.

Very disappointed by this course. I took the specialization to better understand Attention and these few videos are very unclear... I saw in the forum that my sentiment in shared by many people. Hope that Andrew will react and give us a better learning material.

Tricky course, not well explained. I had to struggle a bit to understand the various concepts.

I'm a master/graduate student who took an NLP course in Uni.

I think that overall this is a very a good introduction to the topic. Some concepts are really well explained - in a simple manner and with a lot of jupyter-lab code to experiment with.

In general in this specialization - the first 3 courses are good. There are some quirks (e.g. why Lukas is needed at all? He doesn't really teaches, just passes it on to Younes) but nevertheless I learned from it. And I think they have good value in them.

The 4th one, however, is completely disappointing. First 2 "weeks" are confusing, not really well explained, but somewhat "bearable". The last 2 weeks are a complete sham. They claim to teach "BERT" and "T5" but don't really give any value. You're better off going elsewhere to learn these concepts.

If it wasn't for this, I would give the overall experience a 5 stars, but because of this, I think the overall is more like 3 or 4.

Great specialization, however the 4th course was not cooked enough. It is the most complicated material, sure, so this is the place to put extra effort in preparing the lectures and labs. Instead, I got the feeling you push much too much into 1 course. You should consider splitting it. I hope someone read this feedback...

Started out nicely, but for Week3 and Week4 a lot of the concepts and details are skipped over or copy pasted.

Can the instructors make maybe a video explaining the ungraded lab? That will be useful. Other students find it difficult to understand both LSH attention layer ungraded lab. Thanks

This course glossed over everything and as a result I learned pretty much nothing. The constant congratulations for having done things, when I haven't done anything is aggravating.

The videos need more explanation. Even the assignments were quite challenging because of 'trax'

Token one star.

I was very disappointed in the overall low quality of this course. The labs were confusing (poor formatting, misleading comments), and even though I completed the assignments, I do not feel I obtained any solid grounding of the underlying concepts.

This course is easily the worst course I have taken on Coursera. Why the drop in quality?

TRAX absolutely made it super hard to learn and follow.

If it was explained using Tensorflow or Pytorch it would have been very beneficial.

The course overall introduces a ton of interesting and current concepts.

However, it has the depth of a puddle. While they provide links to existing papers, it would have been nice to hear them discussed in detail by the instructors.

Additionally, the choice to use trax really hinders the experience. It's a new framework with basically no usage outside of this course. I'm just going to learn how to do all this in tensorflow afterwards anyways, so it would have been nice if they used tensorflow (or pytorch) instead.

A helpful suggestion would be to offer separate assignment workbooks for each library (trax, tensorflow, pytorch)! Then, each user can pick the library most relevant to them.

Two last weeks were a bit disappointing. Videos 3min long are honestly not enough to explain how these models work. I barely catched anything.

The course introduces state-of-the-art techniques in NLP.

What is good about the course: (1) the Lab notebooks and assignments are well documented. Much of the material covered in the lectures is covered in much more detail in the notebooks; and the instructions facilitate very much model implementation. So much that one could finalize the assignments in a single afternoon. (2) The reading list and web resources listed are very helpful to understand the models' intuition and how they improve on earlier NLP models.

What is not so good about the course are the video lectures. The lecturer attempts to explain the content in the notebooks but regrettably, his efforts fall short. The script in the video lectures is too repetitive and does not explain the material at the required depth.

After the second week I decided to skip the lectures altogether and proceed to learn the material from the labs, the assignments, and the references. To understand the material fully I watched the corresponding YouTube videos of Stanford's CS 224N. In this regard, the Labs and the assignment served as a good complement of the Stanford course's videos.

Explanation in video lectures is very shallow. Have to read research papers or blogs for better understanding. Lecture videos can be made much better.

Seemed like this course was rushed together. The lectures were very high level and labs did not provide much depth beyond what was already presented in the previous NLP series.

This should really be two separate courses, instead of one. In general, there should also be a separate course for TRAX syntax.

sometimes the videos were a bit short. Instead of making 3 3 min videos with part I II and III with each and intro and outro of a minute, just make one solid video explaining the concepts. Some weeks only had a duration of 30 minutes of theory with 15 minutes of useless intros and outros. I felt like these minutes could used to go a bit deeper into the material. Furthermore, whats op with the names of the weeks? A week is named Q&A but actually explains al the Transformer models and Q&A is like a small application of it?? Last but not least, you could fill in the assignments without even reading or understanding the concept because it is just an auto complete coding exercise. Normally I always better understand the theory because of the assignments, but now I felt the assignments were sometimes useless and didn't really teach me to implement a solid ml model. It doesn't really match up with what you will have to implement if you are working at a company, I think. The assignments could be a bit more of indepented coding with a non auto-complete explanation. On the other side, I think this is ofcourse difficult to correct and thus to scale up the course.

Dr. Ng's Deep Learning specialization is so good: 5 stars. For me, this course was not nearly as good as the courses in that specialization. I felt like I could have just read the papers on BERT, GPT-2, T5, and the Reformer, and would have learned the same amount. The one exception was the lecture video on the history of Transformers (the evolution from ELMO->BERT->T5, etc.). Also the ungraded Reformer labs; those were good. But I personally didn't get very much value out of all the other lectures and labs.

I'm very disappointed with the whole deeplearning.ai NLP specialization in general, but this course was icing on the cake.

The course treats advanced and state-of-art techniques in NLP with neural neutworks, but the theoretical lectures are confusing and imprecise. The framework programming assignments are totally useless, since the user is asked to implement the network architectures discussed in the lectures using a "fill the dots" approach with a very restrictive starter structure. In my personal experience, this yielded a close-to-zero learning outcome, but a lot of frustration in trying to get around some bugs in the auto-grading system, by desperately browsing in the posts from the learners community.

I came here after the very nice Deep Learning Specialization held by Andrew Ng and wasn't expecting this.