Глубинное обучение (курс лекций)/2020
Материал из MachineLearning.
(Новая: __NOTOC__ This is an introductory course on deep learning models and their application for solving different applied problems of image and text analysis. '''Instructors''': [[Участ...) |
|||
| (19 промежуточных версий не показаны.) | |||
| Строка 5: | Строка 5: | ||
The timetable in Autumn 2020: Fridays, lectures begin at 10-30, seminars begin at 12-15, [https://us02web.zoom.us/j/83283454898?pwd=SENBQUNqMVRURXhPdmZZRTdsdUF1dz09 zoom-link] | The timetable in Autumn 2020: Fridays, lectures begin at 10-30, seminars begin at 12-15, [https://us02web.zoom.us/j/83283454898?pwd=SENBQUNqMVRURXhPdmZZRTdsdUF1dz09 zoom-link] | ||
| - | |||
| - | |||
| - | |||
| - | |||
| - | |||
| - | |||
== Rules and grades == | == Rules and grades == | ||
| - | + | We have 6 home assignments during the course (among assignments on VAE and GAN you should choose GAN in case you have already made an assignment on VAE in previous semesters and you may choose any of these two otherwise). For each assignment, a student may get up to 10 points + possibly bonus points. For all assignments a student is allowed to upload his fulfilled assignment during one week after deadline with grade reduction of 0.3 points per day. All assignments are prepared in English. | |
| - | + | ||
Also each student may give a small 10-minutes talk in English on some recent DL paper. For this talk a student may get up to 5 points. | Also each student may give a small 10-minutes talk in English on some recent DL paper. For this talk a student may get up to 5 points. | ||
| - | The total grade for the course is calculated as follows: Round-up (0.3*<Exam_grade> + 0.7*<Semester_grade>), where <Semester_grade> = min(10, (<Assignments_total_grade> + <Talk_grade>) / | + | The total grade for the course is calculated as follows: Round-up (0.3*<Exam_grade> + 0.7*<Semester_grade>), where <Semester_grade> = min(10, (<Assignments_total_grade> + <Talk_grade>) / 6), <Exam_grade> is a grade for the final exam (up to 10 points). |
{| class="standard" | {| class="standard" | ||
!Final grade !! Total grade !! Necessary conditions | !Final grade !! Total grade !! Necessary conditions | ||
|- | |- | ||
| - | | 5 || >=8 || | + | | 5 || >=8 || 5 practical assignments are done, exam grade >= 6 and oral talk is given |
|- | |- | ||
| - | | 4 || >=6 || | + | | 4 || >=6 || 4 practical assignments are done, exam grade >= 4 |
|- | |- | ||
| 3 || >=4 || 3 practical assignments are done, exam grade >= 4 | | 3 || >=4 || 3 practical assignments are done, exam grade >= 4 | ||
|- | |- | ||
| - | |}-- | + | |} |
| + | |||
| + | == Exam == | ||
| + | Exam for the group 517 is scheduled on 11th of January. Exam for the group 622 is scheduled on 15th of January. All students are [https://docs.google.com/spreadsheets/d/1jpSehhwRNXXRpFzNh8YuLHDpyqp9ksAR59CBpQE1oc4/edit?usp=sharing distributed beforehand] for particular time slots. In one hour before the scheduled time a student get an e-mail with exam question and zoom-link. During this hour a student writes an answer to the exam question (it is allowed to use any materials during this time). Then at the scheduled time a student enters the corresponding zoom-conference. During examination a student from his part should be able to write formulas as answers to examinator questions in online mode. Here a student may use a graphic tablet or use his mobile phone as web-camera, put it over a table and write using pen and paper. | ||
| + | |||
| + | Please pay attention to questions from theoretical minimum section. Examination starts with discussing this part. A student should be ready to answer any of these questions without preparation. Unsatisfactory answer to any of these questions leads to failed score for the exam. | ||
| + | |||
| + | [[Media:DL2020_exam_questions.pdf|Exam questions + theoretical minimum]] | ||
| + | |||
| + | == Student presentations == | ||
| + | Each student may prepare a presentation on some recent DL topic. This activity is compulsory for the final course grade 5 and optional for all the other cases. Presentation must be in English, 10-minutes long and cover some papers from the last 3 years (2018, 2019 and 2020). Please register for particular talk on either 11th or 18th of December [https://docs.google.com/spreadsheets/d/1gBDNmhzJ5NTT1LCTPr089aetOBjHlCIri8OADPlsnmo/edit?usp=sharing here]. The maximum capacity for each of two days - 12 presentations. | ||
== Lectures and seminars == | == Lectures and seminars == | ||
| Строка 35: | Строка 38: | ||
!Date !! No. !! Topic !! Materials | !Date !! No. !! Topic !! Materials | ||
|- | |- | ||
| - | | | + | | rowspan="2"|11 Sep. 2020 || rowspan="2" align="center"| 1 || Introduction. Fully-connected networks. || |
| + | |- | ||
| + | | Matrix calculus, automatic differentiation. || [https://drive.google.com/file/d/1Yu790uIPyxp9JIyysxfJDor_LJQu83gQ/view?usp=sharing Synopsis] | ||
| + | |- | ||
| + | | rowspan="2"|18 Sep. 2020 || rowspan="2" align="center"| 2 || Stochastic optimization for neural networks, drop out, batch normalization. || | ||
| + | |- | ||
| + | | Convolutional neural networks, basic architectures. || [https://drive.google.com/file/d/1uSVdPsn5wznk510gS9N1K9DXITpxNFXt/view?usp=sharing Presentation] | ||
| + | |- | ||
| + | | 25 Sep. 2020 || align="center"| 3 || Pytorch and implementation of convolutional neural networks. || [https://github.com/nadiinchi/dl_labs/blob/master/lab_cnn_english.ipynb ipynb 1]<br> [https://github.com/nadiinchi/dl_labs/blob/master/loss_surfaces_lab/lab_loss_surfaces.ipynb ipynb 2]<br> | ||
| + | [https://github.com/nadiinchi/dl_labs/blob/master/lab_pytorch.ipynb ipynb 3] | ||
| + | |- | ||
| + | | 02 Oct. 2020 || align="center"| 4 || Semantic image segmentation. || [https://yadi.sk/d/jel16JzCmHLgBQ Presentation (pdf)]<br>[https://portrait.nizhib.ai/ Portrait Demo] ([https://github.com/nizhib/portrait-demo source]) | ||
| + | |- | ||
| + | | 09 Oct. 2020 || align="center"| 5 || Object detection. || [https://yadi.sk/i/vmJJgDAAvtY6Pw Presentation (pdf)]<br>[https://yadi.sk/i/5gLFLx1R7Qfjjg DS Bowl 2018 (pdf)] | ||
| + | |- | ||
| + | | 16 Oct. 2020 || align="center"| 6 || Neural style transfer. || [https://yadi.sk/i/Hp9wbpaIEHz_pw Presentation] | ||
| + | |- | ||
| + | | 23 Oct. 2020 || align="center"| 7 || Recurrent neural networks. || [https://drive.google.com/file/d/1KvSzzctOjRhYwJH_9LJJeZhMp4USTcDV/view?usp=sharing Presentation] | ||
| + | |- | ||
| + | | 30 Oct. 2020 || align="center"| 8 || Recurrent neural networks memory and attention mechanisms. || | ||
| + | |- | ||
| + | | 06 Nov. 2020 || align="center"| 9 || Reinforcement learning. Q-learning. DQN model. || | ||
| + | |- | ||
| + | | rowspan="2"|13 Nov. 2020 || rowspan="2" align="center"| 10 || Policy gradient in reinforcement learning. REINFORCE and A2C algorithms. || | ||
| + | |- | ||
| + | | Reinforcement learning implementation and multi-armed bandits. || [https://github.com/nadiinchi/dl_labs/blob/master/lab_reinforcement_en.ipynb RL notebook]<br>[https://www.youtube.com/watch?v=kopoLzvh5jY Multi-Agent Hide and Seek video]<br>[https://github.com/nadiinchi/dl_labs/blob/master/lab_bandits.ipynb Bandits notebook]<br>[https://learnforeverlearn.com/bandits/ Bayesian Bandit Explorer] | ||
| + | |- | ||
| + | | 20 Nov. 2020 || align="center"| 11 || Generative adversarial networks || [https://yadi.sk/i/wNmNOSipwhRbWQ Part1] [https://yadi.sk/i/s5goIhh_0WxLwg Part2] | ||
| + | |- | ||
| + | | 27 Nov. 2020 || align="center"| 12 || Variational Autoencoders || | ||
| + | |- | ||
| + | | 04 Dec. 2020 || align="center"| 13 || Reparameterization methods || | ||
| + | |- | ||
| + | | 11 Dec. 2020 || align="center"| 14 || Student presentations || | ||
| + | |- | ||
| + | | 18 Dec. 2020 || align="center"| 15 || Student presentations || | ||
|- | |- | ||
| - | |||
|} | |} | ||
| + | |||
| + | ==Recommended Literature== | ||
| + | * [https://deepmachinelearning.ru Deep Machine Learning] - Victor Kitov, Lomonosov Moscow State University. | ||
| + | * [https://education.yandex.ru/handbook/ml Machine Learning ] - Yandex School of Data Analysis. | ||
Текущая версия
This is an introductory course on deep learning models and their application for solving different applied problems of image and text analysis.
Instructors: Dmitry Kropotov, Victor Kitov, Nadezhda Chirkova, Oleg Ivanov and Evgeny Nizhibitsky.
The timetable in Autumn 2020: Fridays, lectures begin at 10-30, seminars begin at 12-15, zoom-link
Rules and grades
We have 6 home assignments during the course (among assignments on VAE and GAN you should choose GAN in case you have already made an assignment on VAE in previous semesters and you may choose any of these two otherwise). For each assignment, a student may get up to 10 points + possibly bonus points. For all assignments a student is allowed to upload his fulfilled assignment during one week after deadline with grade reduction of 0.3 points per day. All assignments are prepared in English.
Also each student may give a small 10-minutes talk in English on some recent DL paper. For this talk a student may get up to 5 points.
The total grade for the course is calculated as follows: Round-up (0.3*<Exam_grade> + 0.7*<Semester_grade>), where <Semester_grade> = min(10, (<Assignments_total_grade> + <Talk_grade>) / 6), <Exam_grade> is a grade for the final exam (up to 10 points).
| Final grade | Total grade | Necessary conditions |
|---|---|---|
| 5 | >=8 | 5 practical assignments are done, exam grade >= 6 and oral talk is given |
| 4 | >=6 | 4 practical assignments are done, exam grade >= 4 |
| 3 | >=4 | 3 practical assignments are done, exam grade >= 4 |
Exam
Exam for the group 517 is scheduled on 11th of January. Exam for the group 622 is scheduled on 15th of January. All students are distributed beforehand for particular time slots. In one hour before the scheduled time a student get an e-mail with exam question and zoom-link. During this hour a student writes an answer to the exam question (it is allowed to use any materials during this time). Then at the scheduled time a student enters the corresponding zoom-conference. During examination a student from his part should be able to write formulas as answers to examinator questions in online mode. Here a student may use a graphic tablet or use his mobile phone as web-camera, put it over a table and write using pen and paper.
Please pay attention to questions from theoretical minimum section. Examination starts with discussing this part. A student should be ready to answer any of these questions without preparation. Unsatisfactory answer to any of these questions leads to failed score for the exam.
Exam questions + theoretical minimum
Student presentations
Each student may prepare a presentation on some recent DL topic. This activity is compulsory for the final course grade 5 and optional for all the other cases. Presentation must be in English, 10-minutes long and cover some papers from the last 3 years (2018, 2019 and 2020). Please register for particular talk on either 11th or 18th of December here. The maximum capacity for each of two days - 12 presentations.
Lectures and seminars
| Date | No. | Topic | Materials |
|---|---|---|---|
| 11 Sep. 2020 | 1 | Introduction. Fully-connected networks. | |
| Matrix calculus, automatic differentiation. | Synopsis | ||
| 18 Sep. 2020 | 2 | Stochastic optimization for neural networks, drop out, batch normalization. | |
| Convolutional neural networks, basic architectures. | Presentation | ||
| 25 Sep. 2020 | 3 | Pytorch and implementation of convolutional neural networks. | ipynb 1 ipynb 2 |
| 02 Oct. 2020 | 4 | Semantic image segmentation. | Presentation (pdf) Portrait Demo (source) |
| 09 Oct. 2020 | 5 | Object detection. | Presentation (pdf) DS Bowl 2018 (pdf) |
| 16 Oct. 2020 | 6 | Neural style transfer. | Presentation |
| 23 Oct. 2020 | 7 | Recurrent neural networks. | Presentation |
| 30 Oct. 2020 | 8 | Recurrent neural networks memory and attention mechanisms. | |
| 06 Nov. 2020 | 9 | Reinforcement learning. Q-learning. DQN model. | |
| 13 Nov. 2020 | 10 | Policy gradient in reinforcement learning. REINFORCE and A2C algorithms. | |
| Reinforcement learning implementation and multi-armed bandits. | RL notebook Multi-Agent Hide and Seek video Bandits notebook Bayesian Bandit Explorer | ||
| 20 Nov. 2020 | 11 | Generative adversarial networks | Part1 Part2 |
| 27 Nov. 2020 | 12 | Variational Autoencoders | |
| 04 Dec. 2020 | 13 | Reparameterization methods | |
| 11 Dec. 2020 | 14 | Student presentations | |
| 18 Dec. 2020 | 15 | Student presentations |
Recommended Literature
- Deep Machine Learning - Victor Kitov, Lomonosov Moscow State University.
- Machine Learning - Yandex School of Data Analysis.
