理工学研究科Graduate School of Science and Engineering
HUI500X3（人間情報学 / Human informatics 500）知的情報処理特論２Intelligent information processing (Ⅱ)
彌冨 仁Hitoshi IYATOMI
授業コードなどClass code etc
| 学部・研究科Faculty/Graduate school | 理工学研究科Graduate School of Science and Engineering |
| 添付ファイル名Attached documents | |
| 年度Year | 2024 |
| 授業コードClass code | YB017 |
| 旧授業コードPrevious Class code | |
| 旧科目名Previous Class title | |
| 開講時期Term | 秋学期授業/Fall |
| 曜日・時限Day/Period | 水2/Wed.2 |
| 科目種別Class Type | |
| キャンパスCampus | 小金井 / Koganei |
| 教室名称Classroom name | 小東館-Eゼミ室3 |
| 配当年次Grade | |
| 単位数Credit(s) | 2 |
| 備考（履修条件等）Notes | |
| 実務経験のある教員による授業科目Class taught by instructors with practical experience | |
| カテゴリーCategory | 応用情報工学専攻 |
すべて開くShow all
すべて閉じるHide All
Outline (in English)
In this course, we will focus on generative models, which have been developed in recent years among machine learning techniques. We will first look at practical generative models (LLMs and diffusion models) through exercises, and then learn about each of the major techniques.
授業で使用する言語Default language used in class
英語 / English
授業の概要と目的（何を学ぶか）Outline and objectives
In this course, we will focus on generative models, which have been developed in recent years among machine learning techniques. We will first look at practical generative models (LLMs and diffusion models) through exercises, and then learn about each of the major techniques.
到達目標Goal
Develop a general understanding and ability to apply state-of-the-art research, especially in generative models.
Develop the ability to identify, formulate and solve practical problems in the field of machine learning.
この授業を履修することで学部等のディプロマポリシーに示されたどの能力を習得することができるか（該当授業科目と学位授与方針に明示された学習成果との関連）Which item of the diploma policy will be obtained by taking this class?
ディプロマポリシーのうち、「DP1」「DP2」「DP3」に関連
授業で使用する言語Default language used in class
英語 / English
授業の進め方と方法Method(s)（学期の途中で変更になる場合には、別途提示します。 /If the Method(s) is changed, we will announce the details of any changes. ）
Each class usually consists of lecture, discussion and exercise. Students are requested to do exercise in each class and some homework assignments.
# Contents may vary according to proficiency of students.
アクティブラーニング（グループディスカッション、ディベート等）の実施Active learning in class (Group discussion, Debate.etc.)
あり / Yes
フィールドワーク（学外での実習等）の実施Fieldwork in class
なし / No
授業計画Schedule
授業形態/methods of teaching：対面/face to face
※各回の授業形態は予定です。教員の指示に従ってください。
1[オンライン/online]：Introduction of generative models
Introduction to generative models and their state-of-the-art (LLMs, Diffusion models)
2[対面/face to face]：[Exercise] Large language models
[with exercise] Large language models (e.g. ChatGPT and related libraries)
3[対面/face to face]：[Exercise] Latent diffusion models
[with exercise] Latent diffusion models (e.g. Stable Diffusion and fine-tuning techniques)
4[対面/face to face]：Fundamental of generative models (1)
Convex set and convex function, Jensen’s inequality
5[対面/face to face]：Fundamental of generative models (2)
[with exercise] Expectation maximization (EM) algorithm, Kullback-Leibler (KL) divergence,
6[対面/face to face]：Fundamental of generative models (3)
Manifold hypothesis, Variational autoencoders (VAE)
7[対面/face to face]：Generative adversarial networks (GAN) (1)
Introduction of Generative adversarial networks (GANs)
8[対面/face to face]：Generative adversarial networks (GAN) (2)
DCGAN, WGAN, Lipschitz continuous, WGAN-GP, SN-GAN, evaluation criteria for generative models
9[対面/face to face]：Generative adversarial networks (GAN) (3)
Application of GANs (CycleGAN, Real-ESRGAN, StyleGAN etc.), Soft-IntroVAE
10[対面/face to face]：Semi-supervised and metric training (1)
semi-supervised learning (self training, co-training, graph-based training), siamese-, triplet- networks. (metric, contrastive learning)
11[対面/face to face]：Semi-supervised and metric training (2)
Adversarial example, adversarial training and virtual adversarial training (VAT)
12[対面/face to face]：Details of Diffusion models (1)
Diffusion models and introduction of the latent diffusion model (1/2)
13[対面/face to face]：Details of Diffusion models (2)
Latent diffusion model (2/2) and their fine tuning (textual inversion, DreamBooth etc.)
14[対面/face to face]：More recent Large language models and summary
Recent advancement in large language models and summary
授業時間外の学習（準備学習・復習・宿題等）Work to be done outside of class (preparation, etc.)
【本授業の準備・復習時間は、各４時間を標準とします。】Students must have basic background of machine learning.
テキスト（教科書）Textbooks
No specific textbook assigned. Materials (paper, reference, book chapter, slides) will be provided from time to time.
Students will find good references by their own.
参考書References
"Pattern recognition and machine learning" C.Bishop, Springer 2006.
Keras documentation (deep learning framework)
https://keras.io/
成績評価の方法と基準Grading criteria
60% in exercises in class and homework
40% in final report
(both on-line, off-line)
Use Hoppii system (Learning support system) to submit assignments and feedback.
Explanations of the assignments (except for the final assignment) will be given in the next and subsequent classes.
学生の意見等からの気づきChanges following student comments
Support with native language is sometimes necessary.
学生が準備すべき機器他Equipment student needs to prepare
Own computer (use google colaboratory)
その他の重要事項Others
This class is a continuation of [YB016: Intelligent Information Processing I].
Solid understanding of neural networks, including CNNs, is required.
If not, take [YB016: Intelligent Information Processing I] first.
Use google Colaboratory.
Minimum programming skills in Python is required.