Deep Learning

This course covers the foundations of deep learning and the modern architectures behind today’s AI systems. It starts with multi-layer networks, back-propagation, optimization and generalization, then studies convolutional networks for vision, recurrent networks and attention, and the Transformer. The second half covers large language models, state-space models, deep generative models (VAEs, GANs and diffusion models), self-supervised learning and image–text foundation models.

Homeworks combine theoretical problems with practical notebooks in Python and PyTorch. TA workshops on PyTorch, parameter-efficient fine-tuning (PEFT) and multi-modal models complement the lectures.

1. Foundations: MLPs, universal approximation, back-propagation, optimization, generalization, training techniques
2. Vision: convolutional networks, CNN architectures and visualization, segmentation and detection
3. Sequences: RNNs and LSTMs, word embeddings and language modeling, attention and Transformers (BERT, T5, GPT, ViT)
4. Large language models: adapting LLMs to new tasks, from LLMs to assistants; state-space models (Mamba)
5. Generative models: VAEs, GANs, diffusion models
6. Representation learning: self-supervised learning, image–text models (CLIP, DALL·E)

Homeworks do not necessarily carry equal weight; each one’s share depends on its importance. With the homework bonus points, the total can exceed 20.

• Announcements and questions: all announcements are posted on Quera; check it regularly. Ask your questions about the course and the homeworks on Quera as well.
• Quera invitations are sent to the email you registered with the university. If you have not received one by the end of the add/drop period, email the head TA with your name and student number.
• Session feedback: a short feedback form (QR code) is shared in each session. Filling it in earns a small bonus.
• Programming: homeworks are done in Python, using PyTorch.
• Late days: you have 8 free late days in total for the homeworks; this cannot be extended. The maximum lateness for each homework is stated in the homework. It may be shorter for homeworks just before an exam, since their solutions must be released in time.
• Late penalty: after the free late days are used, a penalty of 0.5% per hour applies. Lateness is not counted linearly: a submission between 1 AM and 12 noon counts as submitted at noon, so staying up late does not save late time. Free late days are distributed across homeworks so as to minimize your penalty.
• Deadlines are firm. Homework workload is designed for the given time, so please do not ask for extensions; plan early and use your late days wisely.
• Submission is through Quera. Practical homeworks also have an in-person delivery session.
• Academic integrity: enrolling in the course means accepting the homework honor code and the code of conduct for using AI tools (part B). Cheating, including copying code or answers from other students or online sources, is dealt with through the strictest formal procedures.
• Special circumstances: if you face special difficulties during the term, discuss them with the instructor or the head TA when they happen.
• Workshops and TA sessions cover prerequisites for the homeworks and topics beyond the lectures. If you need an extra problem-solving session, ask the instructor or the head TA.
• Office hours: the instructor and the TAs hold weekly office hours (with prior arrangement). Bring ideas, research questions and general questions to the instructor; ask detailed questions about the material or methods from the TAs.
• Bonus practical exercises can also be submitted at the end of the term, after the final exams, at half weight.