666

ths

非常好

666

top

ok

<3

kk

gj

:

)

A

-

-

A

i

赞

The Sequence Models course was the one I sought out in the deep learning specialization. Very interesting assignments, e.g. neural machine translation, music composition, etc. - much more interesting than the convolutional network models, in my opinion. However, it is also much more difficult to follow; probably the most difficult one of the five courses.

Prof. Ng did a wonderful job in the delivering the materials, as always. However, I expected a lot more details about the sequence models, and recurrent networks as much as the ones given in the previous courses. I was looking forward to learn more in-depth about this model, but I didn't feel I get all that I wanted. For example, I wish there an example, step-by-step walkthrough of the backpropagation through time (BPTT) algorithm, especially for the LSTM and GRU models.

The assignments were a little more difficult to follow, I think. To me, the instructions were not as clear as the previous courses (in my opinion), especially when using Keras objects/layers - "use this *object/layer*" but it wasn't clear whether or not to fiddle with the arguments. Usually when it does require a specific value for the argument (e.g. axis=x), it will be mentioned either in the text or code comments. I guess it's a good challenge, but I find myself doing more trial-and-error with the coding to get it to work instead of having some guidance on how to use those Keras objects/layers. The discussion forums do help, however. Lastly, some of the assignments involved building a recurrent model using Keras layers, I felt like there was not enough explanation why such architecture, layers, or hyperparameter values were chosen.

Overall, I liked the course, I did learn a lot from the course, and enjoyed the models we get to play with in the assignments. I think I will still run into problems trying to devise my own sequence models, and fumble with Keras. I wish there is a more in-depth course on the sequence model. Prof. Ng's delivery was excellent; I enjoyed listening to every one of his lectures (even at 2x speed) :)

Thank you to Prof. Ng, and all the people who worked hard to develop the course.

(09/2019)

Overall the courses in the specialization are great and provide great introduction to these topics, as well as practical experience. Many topics are explained clearly, with valuable field practitioners insight, and you are given quizzes and code-exercises that help deepen the understanding of how to implement the concepts in the videos. I would recommend to take them after the initial Andrew Ng ML course by Stanford, unless you have prior background in this topic.

There are a few shortbacks:

1 - the video editing is poor and sloppy. Its not too bad, but it’s sometimes can be a bit annoying.

2 - most of the exercises are too easy, and are almost copy-paste. I need to go over them and create variations of them in-order to strengthen my practical skills. Some exercises are quite challenging though (especially in course 4 and 5), and I need to go over them just to really nail them down, as things scale up quickly. Course 3 has no exercises as its more theoretical. Some exercises have bugs - so make sure to look at the discussion board for tips (the final exercise has a huge bug that was super annoying).

3 - there are no summary readings - you have to (re)watch the videos in order to check something, which is annoying. This is partially solved because the exercises themselves usually hold a lot of (textual) summary, with equations.

4 - the 3rd course was a bit less interesting in my opinion, but I did learn some stuff from it. So in the end it’s worth it.

5 - Slide graphics and Andrew handwriting could be improved.

6 - the online Coursera Jupyter notebook environment was a bit slow, and sometimes get stuck.

Again overall - highly recommended

Note: this review also applies to the specialization as a whole. I thoroughly enjoyed the courses and learnt so much. The content delivery was excellent. I am not sure how unfeasible it was to re-produce the videos, though it would have been nice to see fewer corrections. In regards to the programming assignments, I think that they are great overall. The way the code was mostly pre-filled to ensure logical workflow helped me to stay on track rather than trying to skip straight to the final model (I assume beginner programmers do this a lot). It provided a great template for thought and I will definitely be referring back regularly. However, I also think that descriptions and exercises could be integrated better, especially in the later assignments where students are compiling more complex models and prior skills in programming (from earlier courses) are expected. Specifically, there are certain Exercise descriptions with multiple parts 1a,b,c 2a,b,c etc and hints before you even get to the start of the actual implementation. I think this is inefficient and was a bit frustrating at times. It felt a bit disjointed and overwhelming to read the instructions without no context (i.e. reading about step 2c before implementing step 1a). On the whole, I know that the knowledge, understanding and skills I obtained through this specialization will serve me extremely well throughout my PhD. Thank you very much.

I enjoyed Prof Ng's excellent lectures, but felt the material moved too quickly. This 3-week course could easily have been extended to (say) 5 weeks to allow for more depth in covering the various RNNs, applications and model details.

The homework/programming assignments were more difficult and time-consuming than prior courses, particularly for implementing models with Keras. The structure of these programs was hard to understand (a bit of spaghetti-code, in my opinion). Some experience with Keras and tensorflow is essential. I spent a lot of time just trying to construct the programs with correct syntax, etc...while this is useful to know, I'd rather focus more on fundamentals of the learning algorithms.

However, its clear that a lot of effort went into constructing the programming exercises for this course, and they covered a lot of ground, with a bit more sophistication than the exercises from most of the earlier courses.

While I loved listening to Andrew Ng's lectures and I find him very lucid in his presentation and pedagogy, I feel that the practical aspect has suffered -by giving enough hints on how to solve the programming exercises, the challenge is reduced. There were quite a few issues I also faced when connecting to the server which resulted in rewriting the code a couple of times (in hindsight, I should have always made a local copy and tested it before submitting it). I would rather that each of these courses becomes a 2 month course (much like Stanfords convolutional networks course) so that the practical aspect is also given equal weightage. While presenting the lectures, Prof Andrew Ng could also lay it out how you would implement in a particular framework like tensorflow and there should be enough exercises that walk a person through them before attempting the programming exercises.

This course has great material on sequence models, presented with the usual energy and enthusiasm that Andrew Ng brings to every course. The model diagrams are great for visualizing what's actually going on in the complex assignments, and the assignments are generally designed with 1) additional code and commentary to make the examples informative and 2) the "guard rails" (e.g., insert code here, with related hints) to clarify the specific lines to edit. Having said that, I feel that the assignments were a little less polished/refined relative to earlier courses in the concentration, and I spent significantly more time in the discussion forums than I had in prior courses. I also recommend investing in additional training in Keras and Tensor Flow as a prerequisite or in parallel to this course to help get the most out of the practical applications of the material.

As i progressed through the specialization i liked the content less for the following 3 reasons. 1) With the exception of the first course there are no in-video questions which would help learn/exercise on the spot. There could easily have been one question per 3-4 minutes of video content. 2) The programming exercises although useful, became more disconnected from the lectures as the specialization progressed- ie they focused on keras documentation which could have been accompanied by a lecture video dedicated to that goal (as in the second course with tensor flow). The result was that I just tried to get through the programming exercises in courses 4 and especially 5 without understanding much of what I was doing. 3) There were too many errors in the lecture. Overall a good specialization but at least points 1 and 2 should be fixed

this was a really interesting course. Too bad the details of the math are not really there.

That being said, its greatest redeeming factor is that Andrew cites the research papers, and with his overviews of the various models, I can read those papers, and build up my own library of relevant material.

I don't really care for Jupyter Notebooks after all...I much prefer the Spyder IDE, as it has intellisense, true debugging, and is not prone to crashing the kernel.

I came across some of Andrew's course lecture notes for his CS courses at Stanford - now those have much more mathematical detail - perhaps Coursera can provide the links to the online material? ( unless, of course, that is problematic due to copyrights?)

All in all, I did enjoy the entire deeplearning.ai material as it is....the rest I can dig into myself.