Coursera: Sequence Models Review

Sequence Models is the fifth and final course in the new deep learning specialization offered by Coursera taught by Coursera founder Andrew Ng. The three-week course focuses on recurrent neural networks and related architectures for applications related to sequential data like speech recognition and machine translation. It is a good idea to complete courses 1, 2 and 4 in the specialization before taking this course, as it builds on the notation and concepts covered earlier, although you could also take the course if you've had prior exposure to neural networks outside the specialization. This course requires a subscription for access to graded quizzes, assignments and a certificate, but you can still watch all the lecturers and view the programming assignments if you audit the course.


Sequence Models wastes no time getting started, covering sequence models and recurrent nets in general, as well as gated recurrent units (GRUs) and long short-term memory networks (LSTMs) all in the first week. Weeks 2 covers text-related sequence modeling topics such word embeddings, sentiment analysis and debiasing word embeddings, while week 3 discusses sequence to sequence architectures, speech recognition and trigger word detection. Each week of content consists of about 90 minutes of lecture followed by 2 or 3 programming assignments. The lecture style is identical to the previous courses in the specialization: each video consists of prepared slides augmented by text written by Andrew Ng in real time along with voice-overs explaining the slides and text. Each video also has a short intro and outro where Ng appears on screen and highlights key takeaways. The video production quality is mediocre, but the instruction is excellent, which is a good thing since the course tries to cover a lot of topics in a short amount of time.

The programming assignments in Sequence Models guide you through the process of implementing and applying lecture topics in code. The assignments are administered through Python notebooks you can open in your browser, so you don't need to worry about setting up a local programming environment. Each programming assignments has an expected completion time listed, which can be used to gauge how long they are although such estimates sometimes understate time requirements. The assignments provide you with code skeletons containing all the functions, arguments and control flow necessary to complete them, so all you have to do is read long and implement key lines of code, usually having to do with calculating values in the forward pass of recurrent nets. There is a lot of text and specific guidance for each line of code you write, which helps keep the assignments moving along despite covering complex topics. Several assignments use the Keras neural net package to build more complex networks which makes those assignments a bit more difficult as Keras isn't given enough attention in the specialization for students to really be comfortable using it. For some reason the introduction to Keras offered earlier in the specialization was optional despite it being necessary to complete assignments.

Sequence Models provides a nice introduction to sequence-related neural networks and gives learners a hands on look at a variety of tasks you can tackle with them through the programming assignments. The course probably tries to pack a bit too much into three weeks, but covering all the basics of recurrent net architectures in the first week allows more time to cover interesting applications later in the course.

I give Sequence Models 4.25 out of 5 stars: Great.