Prague’s Dive into Deep Learning

Do you want to not only use modern AI but also understand it from the inside? At Matfyz, you have the chance starting next week, as part of regular academic teaching. Here’s my experience from last year.

If you ask ChatGPT about the ideal content of a university course on ‘deep learning,‘ it seems to suggest exactly what is actually taught at Prague’s Faculty of Mathematics and Physics (MFF). But the course itself has taken shape gradually since the summer of 2016, evolving alongside rapid advancements in the field through the instructor’s continuous efforts.

Prague students have probably heard about the course through the student grapevine. I am writing more for those who studied traditional IT earlier and now wish to refresh their knowledge significantly, from the ground up, at an academic level. You may have nostalgic memories of flip-flops and logic gates in hardware or of your first encounter with branching and loops in software.

In this course, you will acquire exactly such fundamental theoretical building blocks, but for deep learning—though the complexity of the initial modules and concepts is significantly higher. The long-standing paradigm of neural networks has been revitalized with deeper models (dozens to low hundreds of layers) and modern hardware. It is precisely these methods that have given rise to the current wave of chatbots and large language models.

The course is maximalist—students are expected to complete up to 45 programming assignments per semester, about 2–4 per week, based on the preceding lecture explanations and introductory seminar guidance.

Assignments always have a two-week deadline and form the fundamental rhythm you must adapt to. Yet another batch comes just a week later. Another major rhythm is the lectures, which you will more or less need to understand in order to complete the assignments. Attending classes in person is not necessary; everything gradually appears on the course website. Lectures and seminars are available in both English and Czech, but the assignments are exclusively in English.

The main instructor, Milan Straka, is a co-author of what is currently the ‘largest‘ transformer for Czech (RobeCzech), trained on major linguistic tasks.

The course content (NFPL138 as of last year , in IS ) condenses an ever-growing body of deep learning knowledge. After an initial review of essential math, the core topics include: Basics of neural network training, including backpropagation and SGD; Convolutional networks (CNN): classification, detection, and segmentation; Recurrent networks (RNN), including LSTM and GRU; Using neural networks for traditional linguistic tasks like tagging and lemmatization; Structured prediction, CTC, and speech recognition; Word2Vec, seq2seq, NMT; Transformers, including BERT and RoBERTa (with RobeCzech or EleCzech); ViT (Vision Transformers); Deep reinforcement learning (RL); GANs, diffusion models, speech synthesis, and meta-learning.

The frameworks used for assignments vary—simpler tasks use Keras, while more complex ones use PyTorch. TensorFlow or JAX may also be used if preferred by you, and the emphasis has shifted over the years based on broader industry trends.

The remaining information, only briefly:

Key Values

• 100% Prepared Assignments: Skeleton code, training data, and correctness verification are provided.
• 90% Cutting-Edge Theory: The course covers deep learning in a solid and up-to-date manner.
• 50%–80% Clarity of Explanation: There is room for improvement—partly a trade-off for being taught at the very frontiers, and some concepts are objectively difficult to explain in a simple way.

Course Prerequisites (What You Should Know or Have Beforehand)

• Python programming is essential!
• Intellectual level of a Matfyz student* —this course is designed for them.
• Refresh your understanding of function derivatives and basic probability (watch the first lecture’s content).
• No fear of tensors—they can be thought of as n-dimensional ‘matrices‘ but are primarily used mostly in concrete implementations only. If you still dislike such structures, it need not be for you.
• Plenty of time—ideally, 3.5 months of relatively free time.

When Not to Take the Course (Avoid Frustrating Yourself)

• If you expect to be guided by the hand—it won’t work.
• If you lack any prerequisites—you won’t understand or keep up.
• If you plan to ‘wing it‘ after work but aren’t a genius—you won’t have enough time.

Can I Take the Course for Free Without Official Enrollment?

• Yes. All lectures and practicals are publicly available online online and can be ‘attended‘ remotely.
• However, you won’t be able to verify your assignment solutions, and you may lose some motivation without structured evaluation and strict rhythm.

Formal Enrollment Includes:

• Submitting an application and paying Lifelong Learning tuition (free of charge for current CTU students and few other external applicant types).
• Access to all login-protected platforms, including Recodex for solution validation.
• The ability to attend even physically lectures, seminars and individual consultations.
• The opportunity to take the exam and receive official certification (but you must complete enough assignments and pass the exam).

What Forms of Participation Make the Most Sense?

• Completing the full course in a common manner, including the final exam.
• Doing only a few assignments but verifying their correctness. Taking the exam is not mandatory even after official enrollment (uni students are advised to check whether there are any negative consequences, but if you are an external student, it simply does not matter).
• Splitting it over two years, attending everything twice.

If You’re Considering the Course from Outside any University

• Consider persuading next 1–2 colleagues or friends to enroll and attend with you, as about one-third of the hardest (competitive) assignments is allowed to be tackled in such teams, significantly distributing the workload.
• Be aware that the course will consume a lot of your time.

How Many Students Participate?

• The first competition had 164 submissions, dropping to under 100 students by mid-semester and only 40 in the final week.
• My impression was that the course was formally completed by roughly 60–80 students, with about half of them possibly being international participants.

If anything else is important to you, leave a comment—I’ll try to answer.

Other Similar Courses

This blog post isn’t intended as an advertisement, though some particularly sensitive readers might interpret it that way (ethics code 8.5–8.9). I actively discourage anyone from attempting the course, unless they have sufficient time and prerequisites. The fundamental challenge of evaluating academic programs is that you can’t truly review a school’s experience without going through its specific program. And once you have, you’re no longer ideally unbiased… To put it differently: I only feel compelled to write about one course out of dozens, and for reasons beyond myself.

For completeness, let’s at least consider some alternative options for academic deep learning education here around me.

In Prague, AI is also studied at CTU (FIT , FEL , and CIIRC ), but their study information systems are closed enough that it is difficult to determine what can actually be learned there or whether non-students can participate.

Some general links to these institutions can be found via prg.ai, for example, for ML .

A year ago, edX.org had interesting courses from Databricks, but they are now outdated and discontinued. The remaining ones don’t seem quite right, but you can try your luck. Look for topics like ‘Machine Learning,‘ ‘Deep Learning,‘ ‘Generative AI,‘ or ‘Large Language Models.‘ My broader experience with edX is that university courses tend to be didactically better then the corporate ones. In a pinch, however, you take what you can get—the field is too new and fluid even for course providers.

Definitely consider Stanford’s CS224N remote course, esp. if you can find a sponsor. Let me know if you do.

If you know of any good AI courses that are factually, technically, and otherwise accessible to us, please share them in the comments—with links. Thanks!

If you don’t have a login for the comments here, feel free to email me, for example, via ILAW

---

This post in Czech: Pražský ponor do Hlubokého učení

–-

* Being a Matfyz student (a student at the Faculty of Mathematics and Physics, Charles University) means diving deep into the world of mathematics, physics, or computer science at one of the most rigorous academic institutions in Central Europe. The faculty is known for its demanding curriculum, strong theoretical foundations, and emphasis on problem-solving. On the other hand, for this course, only the previously mentioned areas of mathematics are necessary; fortunately, you don’t have to (re)learn everything that full-time students have gone through.