COMP9444: Neural Networks and Deep Learning

Prerequisites:

A pretty boring course personally. Teaches you about different architectures of neural networks and how deep network work, along with the theory behind them, but only a little content beyond COMP9417 of interest. You'll need to have some way of using a good GPU, either by having one or by using an online service, or training the neural networks required for the assignments will be hell and take forever. If you're interested in AI, then you might as well take this course; it's not as bad as some of the other AI courses I've done. Otherwise, a little bland, and doesn't teach you much beyond what you can find online, as long as you're motivated enough.
Content-wise, the course goes through several architectures of neural networks, taking a look at specialised architectures for language processing and image processing. Later in the course, you look at generative adversarial networks, and autoencoders, ending with a more theoretical look at deep learning and reinforcement learning. Overall a pretty decent overview.

2x 2hr Lectures

1.5/5

Yes, all recorded.

Dr. Alan Blair

Lecture slides, quiz and exercise solutions, all available online.

2.5/5

2020 T2

97 HD

- 30% assignment x2
- 40% final exam

COMP2521 or COMP1927 or MTRN3500

This is one of the level 3+ courses offered by the CSE faculty for the Artificial Intelligence specialisation. Loosely speaking, neural networks learn to accomplish various machine learning tasks but in ways that somehow mimics the human brain (hence their name).

The course isn't actually 3/5 in terms of content difficulty (in fact, some might argue the content difficulty was as low as 1.5/5). The final exam appeared to be quite lenient (especially in contrast to its sister course COMP9417). The assignments weren't exactly hard to code. The difficulty lies in fine tuning the parameters in each network you had to code for said assignments. Accuracy was one of the important aspects we were examined on this term. Unless we achieved sufficiently high accuracy, it was not guaranteed that we'd receive the marks.

Note also that training a neural network takes time (at least, once they become complicated). In the grand scheme of things, there's only so little time you have available to get to the high results. COMP9444 also uses PyTorch as of last year. (Prior to then, TensorFlow was used.)

The quizzes and exercises were good preparation for the finals. It was highly recommended to do them.

The problem I had with the course was that it felt like a bore. I felt like I was learning a lot about neural networks, instead of how to actually do them. The course definitely covered the basic stuff (e.g. backpropagation), but I felt that I just got given a bunch of equations and had to accept them. I couldn't really understand anything about why the equations worked the way they did.

There was also an emphasis on the applications, which genuinely were cool, but like they didn't tell me much. I didn't see the point on being examined on all of these seemingly context-focused questions either. Also, PyTorch felt like something we had to self-learn. Fortunately, it was not hard. But the single lecture on it felt quite vague.

That wasn't with every bit of the course though. To be fair, some stuff like reinforcement learning was made clear. It just felt boring for the most part.

And lastly, despite the final exam being quite an easy one, the negative marking was also a bit stressful to deal with leading up to it. The negative marking felt significantly more punishing for this course than back in COMP3231 (Operating Systems).

During COVID:
- 2 x 2 hours live sessions (didn't usually take up the full 2 hours).
- However much time required to watch recordings from previous years. (2-4 hours per week.)

3/5

Yes (live sessions were recorded)

Dr. Alan Blair

Textbook attached below. Otherwise, there were lecture slides, quizzes, and exercises.

1/5

Deep Learning By Ian Goodfellow, Yoshua Bengio and Aaron Courville (Online link available: http://www.deeplearningbook.org/). Didn't consider using it so can't comment on it.

20T2

95 HD