Awesome course

I hope it's ok to review the specialization as a whole here as I'd really like to give some overal context in one review :) ..... I'm sorry to have to write a negative sounding review of the back of covering this specialization given that I also appreciate the efforts involved, never mind the background expertise but if it were not for the high quality of the jupyter notebooks, I would have requested a refund. I have followed Machine learning material on Coursera before and the overall standard of instruction was significantly higher that I what I encountered here. I'm always after a balanced learning experience to help me ramp up on any given topic and while the notebooks were very comprehensively put together and were certainly helpful in completing the programing assignments associated with the last 2 courses even more so which btw. were clearly trickier (both conceptually as well as practically) than the first 2 courses, the lecture delivery wasn't great overall - Please take a leaf out of Andrew Ng's book here - When you are showing basic slides, ***annotate as you are talking***** or at least make use of animations rather than narrating through static slides alone - and for the non-math majors out there who also well positioned to cover this material once they have the stated pre-reqs, use more numbers to move from the more concrete to the more abstract - I've seen Andrew do it very well and any learner such as myself who is happy to invest time + money+ a heap of interest frankly deserves a more comprehensive as well as a more ***consumable delivery*** - I'm also, sad to say, seen far better introductions to LSTMs for example and to help me really appreciate their power as well as their shortcomings, I researched and follow better introductions elsewhere from several sources, never mind the transformer related topics. I was also disappointed at the overly brief treatment of the truly excellent hugging face resource at the time I covered this content - again, after taking more time out to explore how to work with huggingface and consult much better learning material on it to help me ramp up efficently, I felt that I had a proper introduction to it outside what was presented in the last course and that I'm afraid also disappointed me - even more so, given the expertise behind the material. It was great to have a good fine-tuning lab but even better to have a more generic way of working with preparing a data set tbh. that the one presented which is not exactly for everyone (or even most people I might suspect). The content at this stage felt like I was at the end of a day at some seminar where the presenter is running out of time and you end up with a sketchy whistle-stop tour of the platform + concepts - I'm sorry if this sounds harsh to some but unfortunately, given what this course could have been with the clear level of expertise behind it, the material as a whole was quite unbalanced - There were clearly too many individual learning points left to the programing assignments making some of the explanatory text overally long winded - felt really crammed in - I along with many other students should have been prep'd much better with more code samples particularly as when I took the course, I was more familiar with pytorch + keras as opposed to working directly with tensorflow - More instructional support would have been welcomed here given it wasn't on the pre-reqs list when I consulted it. I'm happy to have put plenty of time in effort but the instructional content should match up to the student's efforts - sorry. To avoid making this review much longer, I won't say anything about the quality of the quiz material - What's been written in the forums bears out the shortcomings more than sufficently in my view. Small point I know but there were numerous pop-ups immediately correcting the narrative - I'd always be ok with one or two but there were several that really interrupted the flow - low hanging fruit to fix so please fix and please bear the above points in mind. That said, I did learn and was very happy to so again, appreciate the efforts overall. The reason I took this course was of the back of Machine Learning material I followed previously on coursea which was of a very high standard and really helped me to use time better to research further - for this specialization though, I had to immediately research to get the basics down better to compensate for a less effective delivery. I'm really sorry to say all of the above as I appreciate the experienced minds and the efforts made - Whilst I certainly benefited from completing this specialization, the overall standard of delivery wasn't effective enough for me and I suspect, a significant number of folks relatively new to NLP topics. I would have loved nothing more than to give positive review but all of the above is more than fair comment I'm afraid - This course has the potential to be a fantastic resource with some considered instructional rework.

Overall I found this course underwhelming, especially in comparison to the other three courses in this specialization. The lecture videos don't seem to flow coherently, lots of terminology is introduced without being defined, there seems to be hand-waving over details, and it feels a bit infantilizing when videos ends with statements like "you now understand the T5 architecture" or "you now know how to fine-tune transformer models" when those concepts are not really explained in any meaningful detail in the videos. There also seems to be a big gap between how complex/advanced these topics are and how trivially easy the programming assignments are, with most of the logic implemented for us; completing most exercises doesn't require much more than reading comments to replace "None" values or copying code from preceding cells. In previous courses in this specialization the assignments felt like truer assessments of what we've learned. I hope this course gets a refresh for future students!

The labs in the last two courses were Excellent. However the lecture videos were not very useful to learn the material. I think the course material deserves a v2 set of videos with more in depth intuitions and explanations, and details on attention and the many variants, etc. There is no need to oversimplify the video lectures, it should feel as similar level as the labs (assignments tend to be "too easy" but I understand why that is needed). Thanks for the courses. Azriel Goldschmidt

This course perhaps gives a good overview of the BERT and several other extensions such as T5 and Reformer. I could learn the conceptual framework of the algorithms and understood what we can do with them. However, I think the instructors chose an undesirable mix of rigour and intuition. The lectures are mostly about intuition. In contrast, the assignments are very detailed and go through each logical step one by one.

Nice course. Full of very interesting infomation. What a pity not having used Tensorflow. All that knowledge is unfortunately not work-ready as Trax is not widespreadly used in the industry world and it is hardlyit will ever be. In my opinion.

Explanation of Attention models with Attention mechanism itself and other building blocks of the Transformers was very confusing. It was really hard sometime to udnerstand what the lecturer really meant.

The presented concepts are quite complex - I would prefer less details as most will not understand them anyway and more conceptual information why these models are build as they are

not clear enough. the exersices warent good enough' i didn't learned from them much. it could be a great idea to give the slides at the end of every week for reapet.

Course videos are too short to convey the ideas behind the methodology. Illustration is too rough.

Another less than impressive effort in a specialization from which I expected more.

Explanations of attention/self-attention & other complex topics are too shallow

Course content more detailed explanation to follow.

elaborate and make it easy to learn

It is good

I would not understand much if I haven't known most of the material beforehand. Lots of repetition (not bad, just boring), but worse, bugs as well. Many times the lecturer doesn't know what he's talking about, and messes things up. Characteristic moment is when all of a sudden he talks about things without definition (like "shared frame", "adapter", or shows a diagram contradicting the code besides, etc.), or changes subject abruptly.

The grader is terrible crap happily returning errors but no explanation. You teach AI, you talk about LMs beating humans, yet the tool used for evaluating your students is so primitive as if written two decades ago. It's very likely that it infuriates everybody except its proud author. Either the code to fill is trivial (we learn nothing), or it requires mental work which potentially leaves some traces. The effect is that code works fine, but the grader fails miserably.

Like many of your courses, this one too teaches us more about the author's horizon and expectations, than new knowledge we pay for. This is particularly evident during quizzes where poorly formulated questions, answerable only in narrow context, abound. Also bugs like "translating french to english" require to mark "keys and values are the french words"...

The content is good because it covers many aspects of NLP. There are a lot of illustrations provided to help students understand the materials. However, the assignments are too easy because of the detailed comments provided. This makes it too easy because students could simply copy and paste the answers from the comments.

One suggestion is to improve the explanation of the materials because there re lots of details being skipped by the instructors. Personally, I would have to read other blogs in order to understand some of the details. Furthermore, separating the solutions from the codes is definitely something that must be done for instance presenting the solution in a separate notebook.

1) Information 3 out of 5:

no in depth explanations.

2) quiz are too easy, and I was missing good quizzes that were proposed at DL specialization with use cases, they cause me to think what to pick.

3) home tasks are 1 out of 5:

3.1 First of all all home tasks are done in different manner.

3.2 Some of them require additional check even all tests were passed.

3.3 Part with google collab is also a little bit strange... I want to have 1 click away home task and not setting up 3-rd party env.

What is good: for high - level overview this course is ok. Maybe have 2 versions of the course one with in depth explanations. and one more like this one.

Compared to Andrew Ng's deep learning specialization, this course requires a lot of improvement. Very often disparate facts are put together with not much connection between the ideas. This is probably because of the enormous amount of content covered. It might make sense to split the course into two. Thank you!

The course lectures were very confusing, and the course assignments were too easy, so they didn't reinforce the lecture concepts in the same way that assignments from other courses had.

Like: State-of-the-art NLP problems to be used in the industry.

Dislike: Topics were not well-explained. Difficult to grasp

Trax is no longer in active development.

Not a good course, the first 3 courses were really good but this one was too hard and the vedios didnt complement it well. The vedios just read a script and such a complex science topic is hard to digest this way. The reading was noo different in the other courses as well but since they were easy to understand , it didnt matter much. But this was too hard. Now have to search in you tube ig

Unclear explanation everywhere. Homework assignment cannot be graded