Deep Network Development
2025/26/2
For this semester of Deep Network Development (IPM-21FMIDNDEG), the course leadership has changed: Viktor Varga takes over as the lead instructor, and I continue as co-instructor. The course is also supported this round by four demonstrators (Arpan Ekka, Emma Tayla Elliott, Benedek Fegyó, Arron Pirku) who handle homework feedback. As such, this iteration looks quite different from the previous one: the format is now a single weekly lecture block (no separate practice session), the assessment model has been restructured around two theory tests and two homework assignments, and the Coding Test was dropped mid-semester, so its score is not counted.
Course materials, assignments, assignment feedback, and grading are administered on Canvas. Announcements and public discussion happen on the Teams group, joinable with code mzy01jp.
For questions about course administration, requirements, schedule, or assessment, please reach out to Viktor Varga directly. Teams messages are preferred. For homework feedback follow-ups, reply to the demonstrator on the same Canvas submission page. Please note that the slides may contain small mistakes or typos. If you spot any, a note to my email is appreciated.
Lecture Content
Theory Test 1, plus Unstable gradients problem, transfer learning.
Spring break, no lecture (HW1 normal deadline).
Coding Test (removed from requirements, results not counted), no lecture.
Unsupervised learning with neural networks.
Homework Assignments
There are two homework assignments this semester. Both are graded on an accept/reject basis (5 points each if submitted by the normal deadline and accepted). HW1 covers an MLP-type neural network in NumPy and PyTorch, HW2 covers a more complex architecture and is defended during the oral exam. Detailed specifications and submission pages are on Canvas.
Theory Tests and Oral Exam
The two Theory Tests are short paper-based assessments (about 30 minutes, possibly multiple-choice) covering lectures 1 to 6 (Test 1) and lectures 7 onward (Test 2). Each test has one retake. Sample versions are released as part of the lecture materials above.
The Oral Exam during the exam period has two parts (about 10 minutes each): defense of the HW2 submission, and a few theoretical questions from the lecture material. Eligibility requires meeting the minimum thresholds on both Theory Tests and both homework assignments.
Example versions of the two written theory tests, matching the format of the real assessments held in weeks 8 and 14.
Course Syllabus
Schedule
Lecture:
- Schedule: Tuesdays, from 17:00
- Location: South Building, Room 0-821, Bolyai terem
Note:
- Hungarian: Déli Tömb 0-821 Bolyai terem
- There is no separate practice session this semester.
Description
This course introduces students to the fundamental theory of neural networks and the popular software libraries used to implement them. It features weekly face-to-face lectures, with additional practice materials provided in video and notebook format.
The course covers Supervised Deep Learning end to end: from linear and logistic regression through MLPs and backpropagation, to Convolutional Neural Networks, Recurrent Neural Networks, the Attention mechanism and Transformers, and finally unsupervised learning with neural networks. Implementation work happens in NumPy and PyTorch, with the homework assignments designed to build hands-on familiarity with each layer of the stack.
The course also emphasizes ethical considerations in AI development, ensuring that students not only learn the technical aspects of Deep Learning but also understand its broader impact on society.
Grading
A total of 70 points can be collected during the semester. Minimum requirements must be met in order to pass, and the Coding Test (originally in the requirements) is no longer counted after the requirements were updated mid-semester.
| Component | Points | Minimum |
|---|---|---|
| Theory Test 1 | 15 | combined min. 10 across both tests |
| Theory Test 2 | 15 | combined min. 10 across both tests |
| HW1 (accept/reject) | 5 | must be accepted by final deadline |
| HW2 (accept/reject) | 5 | must be accepted by final deadline |
| Oral Exam, theory | 15 | min. 6 |
| Oral Exam, HW2 defense | 15 | min. 6 |
| Total | 70 | min. 28 overall |
If minimum requirements are not met, or if the total is under 28, the grade is 1 (fail). Otherwise the grade follows the table below.
| Final Score Range | Grade |
|---|---|
| 56 - 70 | 5 |
| 49 - 55 | 4 |
| 42 - 48 | 3 |
| 28 - 41 | 2 |
| < 28 | F |
There is also an optional coding extension available during the oral exam, intended to compensate for the removal of the Coding Test for students who perform significantly better in coding than in theory. It must be requested from the instructor at least one day in advance, can only be taken once, and only by students who scored at least 12 points on every Coding Test they took. See the Canvas course page for full conditions.
Prerequisites
- Linear Algebra
- Probability Theory
- Programming Skills (Python)
Tools and Frameworks
- Programming Language: Python
- Frameworks: PyTorch
- Libraries: NumPy, Matplotlib, torchvision, torchaudio
- Additional Tools: Canvas, Microsoft Teams (announcements)
Guidelines for the Use of AI
The university trains professionals capable of practicing independently without AI tools.
- The use of AI and external assistance is prohibited in closed examinations (Coding Test, Theory Tests, Oral Exams). Unauthorized circumvention of the on-site exam conditions (e.g. unauthorized internet access, external devices) results in immediate exclusion from the course.
- For homework, AI use is not prohibited (it cannot be reliably verified), but using AI for HW1 is seriously discouraged since HW1 is preparation for closed-circumstances coding work. For HW2, AI may be used only to the extent that students fully understand every line of code in their submission.
- Content generated by AI is not necessarily correct. Students use it at their own risk.
Learning Objectives
- Understand the fundamentals of Deep Learning
- Understand and implement Neural Network architectures from scratch and in PyTorch
- Build, train, and evaluate models for classification, regression, sequence modelling, and beyond
- Gain practical familiarity with the PyTorch ecosystem
- Reflect on the ethical implications of Deep Learning systems
Recommended Reading
- Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep Learning. MIT Press.
- Paszke, A., et al. (2019). PyTorch: An Imperative Style, High-Performance Deep Learning Library.
- Hastie, T., Tibshirani, R., & Friedman, J. (2009). The Elements of Statistical Learning.
- Brownlee, J. (2019). Deep Learning for Computer Vision. Machine Learning Mastery.
- Montreal Declaration for Responsible AI.
For questions related to course administration, please contact Viktor Varga directly. For homework feedback follow-ups, reply on Canvas to the demonstrator who reviewed your submission.
back to teaching