Глубинное обучение 2 2025
Содержание
Общая информация
Курс предназначен для студентов 4 курса ФКН ПМИ (МОП и КНАД).
Занятия проходят по понедельникам 14:40-17:40 (переносы будут сообщаться в чате).
Полезные ссылки:
- Чат с обсуждением: https://t.me/+8dcjl4gHlyEwYzcy
- Репозиторий курса: https://github.com/thecrazymage/DL2_HSE
- Таблица с оценками: https://docs.google.com/spreadsheets/d/18rfZbf7Zsm-xmbiQn_eJVDXXbDfesbHT/edit?usp=sharing&ouid=115132401804687564737&rtpof=true&sd=true
- Anytask: https://anytask.org/course/1219
Формула итоговой оценки (округление арифметическое):
- МОП: Оитог = 0.25 * Осоревнование + 0.75 * ОДЗ,
- КНАД: Оитог = ОДЗ,
где ОДЗ - средняя оценка за практические домашние задания.
Преподаватели и ассистенты
Кому писать, если кажется, что все пропало: Мишан Алиев
| Группа | Семинарист | Ассистенты | Чаты групп | Инвайт в anytask |
|---|---|---|---|---|
| 221 (МОП) | Федя Великонивцев | Динар Саберов, Анна Василева | МОП 221 | yvZZTIS |
| 222 (МОП) | Ева Неудачина | Александр Матосян, Полина Кадейшвили | МОП 222 | W38CfZf |
| 223 (МОП) | Иван Ершов | Андрей Уткин, Георгий Фатахов | МОП 223 | ph91Jlz |
| 224 (МОП) | Степан Беляков | Анна Пономарчук, Татьяна Яковлева | МОП 224 | iPwd342 |
| КНАД | Даня Бураков | Анастасия Лапшина, Иван Галий | КНАД | zTn4sRP |
Лекции и семинары
Лекция / Семинар 1 (08.09).
Тема: Essentials of GPU, Deep Learning Bottlenecks, and Benchmarking Basics
Аннотация: In this session, we will explore the reasons behind the dominance of GPUs in Deep Learning and examine the common sources of performance bottlenecks in DL code. You will learn how to identify these bottlenecks using profiling tools and apply techniques to optimize and accelerate your code.
Лектор и семинарист: Fedor Velikonivtsev
Локация: лекция - R208, семинар - D102.
Материалы: запись лекции, запись семинара, материалы.
Лекция / Семинар 2 (15.09).
Тема: On Transformers and Bitter Lesson
Аннотация: In this talk we’ll dive into the landscape of model architectures in deep learning, with a focus on the world around transformers. We’ll briefly recall what a transformer is, trace the evolution from encoder–decoder to encoder-only and decoder-only models, and touch on the rise of “efficient mixers” such as state space models, linear attention, and beyond. We’ll conclude by reflecting on the role of data and compute. The lecture is inspired in part by The Bitter Lesson and blends a brain dump with some entertaining insights from recent years of architectural exploration.
Лектор и семинарист: Ivan Rubachev
Локация: лекция - R206, семинар - R208.
Материалы: запись лекции и семинара, материалы.
Лекция / Семинар 3 (22.09).
Тема: Modern LLMs essentials
Аннотация лекции: This week, we will discuss LLMs. We will discuss why they are so effective for text generation, how they can be applied to different NLP problems, and the risks they pose. You will learn the details of RLHF, PEFT, and RAG, which make LLMs robust in various cases.
Лектор: Alexander Shabalin
Локация: онлайн.
Аннотация семинара: In this seminar, we will explore the concept of LLM-based agents and how they extend the capabilities of modern language models. We will discuss function calling as a way to integrate external tools, chain-of-thought reasoning for structured problem solving, and reinforcement learning techniques for training agents.
Семинарист: Ivan Ershov
Локация: R208.
Материалы: запись лекции, запись семинара, материалы.
Лекция / Семинар 4 (29.09).
Тема: Basics of Efficient LLM Training Infrastructure
Аннотация: In this lecture, we will study the basic rules that underpin the infrastructure for efficient large language model training. We will also examine common problems that may arise in this process and explore practical ways to address them.
Лектор и семинарист: Michael Khrushchev
Локация: лекция - G002, семинар - R208.
Материалы: запись лекции и семинара, материалы.
Лекция / Семинар 5 (06.10).
Тема: Segmentation and Detection
Аннотация лекции: This week, we'll explore the evolution of object detection and segmentation — from R-CNN to real-time methods like YOLO-World and CLIP-based approaches. We'll examine how U-Net architectures have transcended computer vision to power neural operators and how to use diffusion models for segmentation tasks.
Аннотация семинара: We will implement architectures and train a semantic segmentation model. We will discuss regularization methods for convolutional layers. In addition, we will learn how to use pre-trained models and apply them.
Лектор и семинарист: Alexander Oganov
Локация: лекция - G002, семинар - R208.
Материалы: запись лекции, запись семинара, материалы.
Лекция / Семинар 6 (13.10).
Тема: Segmentation and Detection 2
Аннотация лекции и семинара: In this lecture, we will continue in more detail about segmentation and detection.
Лектор: Sergey Zagoruyko
Семинарист: Eva Neudachina
Локация: лекция - G002, семинар - R208.
Материалы: запись лекции, запись семинара, материалы.
Лекция / Семинар 7 (20.10).
Тема: Diffusion models 1
Аннотация лекции и семинара: Tomorrow's lecture and seminar will be devoted to an introduction to diffusion models. Diffusion models are currently the most popular approach to generative modeling due to their high-quality generation and diversity (mode coverage) of the learned distribution. The idea behind diffusion models is to consider the process of gradually transforming data into pure noise and construct its inverse in time, which will transform noise into data. In the lecture and seminar, we will work with noise processes and derive the classic DDPM model, which proposes to minimize the KL-divergence between the “true” reverse process that converts noise into data and the denoising process specified by the neural network. In the process, we will see that this procedure is equivalent to training a denoiser neural network that predicts a clean object from a noisy one. In addition, we will interpret the resulting denoising process: in it, each step corresponds to replacing part of the current noisy image with an (increasingly high-quality) prediction of the denoiser.
Лектор и семинарист: Denis Rakitin
Локация: лекция - G002, семинар - R208.
Материалы: запись лекции, запись семинара, материалы.
Лекция / Семинар 8 (03.11).
Тема: Diffusion models 2
Аннотация лекции и семинара: In this lecture and seminar, we will continue our exploration of diffusion models. We will introduce the score function and score identity, present classifier and classifier-free guidance, and derive DDIM model.
Лектор и семинарист: Denis Rakitin
Локация: лекция - R205, семинар - R208.
Материалы: запись лекции, запись семинара, материалы.
Лекция / Семинар 9 (10.11).
Тема: Diffusion models 3
Аннотация лекции: This lecture presents applied aspects of diffusion models, a class of generative methods that have demonstrated strong performance across images, text, and audio. We will review modern architectures, with a focus on FLUX, and unpack the key principles of their design and training.
Лектор: Nikita Starodubcev
Аннотация семинара: The seminar will provide a hands-on, in-depth inspection of PixArt model. Then ee will attempt further fine-tuning via DreamBooth, review common evaluation metrics, and compare different models and configurations.
Семинарист: Eva Neudachina
Локация: лекция - R205, семинар - R208.
Материалы: запись лекции, запись семинара, материалы.
Лекция / Семинар 10 (17.11).
Тема: 3D CV
Аннотация лекции и семинара: This lecture and seminar will discuss how diffusion models can be used for 3D computer vision tasks such as generation and reconstruction, and how they relate to methods like NeRF and Gaussian splatting. We will also look at how these models can be applied to problems like novel view synthesis, relighting, and camera pose or depth estimation.
Лектор и семинарист: Mishan Aliev
Локация: лекция - R205, семинар - R208.
Материалы: запись лекции, запись семинара, материалы.
Лекция / Семинар 11 (24.11).
Тема: Neural Recommender Systems
Аннотация лекции и семинара: Modern recommender systems also make heavy use of deep learning — both discriminative and generative models. A system has to understand content (tracks, products, videos, etc.), model user behavior (short- and long-term preferences), and predict which content each user is likely to enjoy. At the same time, recommender systems come with their own challenges: cold start for new items and users, huge billion-scale item catalogs with heavy-tailed popularity distributions (popularity bias), and constant distribution drift, where all the underlying distributions keep changing along with the rest of the world. We will discuss: 1) Two-tower neural networks for candidate generation; 2) Sequential recommendation and more modern approaches to generative modeling of users; 3) Semantic IDs and tuning LLMs for recommendation 4) What makes the recommendation domain different from other deep learning domains.
Лектор: Kirill Khrylchenko
Семинарист: Artem Matveev
Локация: лекция - R205, семинар - R208.
Материалы: запись лекции, запись семинара, материалы.
Домашние задания
| № | Домашнее задание | Ссылка | Дедлайн (жёсткий) |
|---|---|---|---|
| 1 | Tensor and DL Libraries | https://github.com/thecrazymage/DL2_HSE/tree/main/homeworks/homework_01 | 30 сентября, 23:59 |
| 2 | Transformers for Named Entity Recognition | https://github.com/thecrazymage/DL2_HSE/tree/main/homeworks/homework_02 | 14 октября, 23:59 |
| 3 | Image Segmentation | https://github.com/thecrazymage/DL2_HSE/tree/main/homeworks/homework_03 | 2 ноября, 23:59 |
| 4 | Diffusion Models | https://github.com/thecrazymage/DL2_HSE/tree/main/homeworks/homework_04 | 18 ноября, 23:59 |
| 5 | Retrieval | https://github.com/thecrazymage/DL2_HSE/tree/main/homeworks/homework_05 | 10 декабря, 23:59 |
Соревнование
Задание: https://github.com/thecrazymage/DL2_HSE/tree/main/homeworks/competition
Возможность перезачета: https://t.me/c/2969026465/2/1601.
Дедлайн (жёсткий): 19 декабря 2025, 23:59
