Course syllabus
Course-PM
TIF360 / FYM360 Advanced machine learning with neural networks lp4 VT22 (7.5 hp)
Course is offered by the department of Physics
Contact details
Examiner
Giovanni Volpe giovanni.volpe@physics.gu.se
Lecturers
Mats Granath Mats.Granath@physics.gu.se
Kristian Gustafsson Kristian.Gustafsson@physics.gu.se
Bernhard Mehlig Bernhard.Mehlig@physics.gu.se
Daniel Midtvedt daniel.midtvedt@physics.gu.se
Teaching Assistants
Homework A, Graph neural networks: David Fitzek, davidfi@chalmers.se
Homework B, Playing Tetris using Reinforcement Learning: Henrik Moberg, henrik.kleinmoberg@chalmers.se
Homework C: Video frame prediction using transformers: Jesus Pineda, jesus.pineda@physics.gu.se
Homework D: Non-linear time-series analysis with reservoir computers: Ludvig Storm, ludvig.storm@physics.gu.se
Project Groups 1, 3, 5, 7, 9: Henrik Moberg, henrik.kleinmoberg@chalmers.se
Project Groups 2, 4, 6, 8, 10: Harshith Bachimanchi, harshith.bachimanchi@physics.gu.se
Project Groups 11, 13, 15, 17, 19: Yu-Wei Chang, yu-wei.chang@physics.gu.se
Project Groups 12, 14, 16, 18, 20: Jesus Pineda, jesus.pineda@physics.gu.se
Review groups: 1-2, 3-4, ...
Course purpose
This course introduces students to recent developments and state-of-the-art methods in machine learning using artificial neural networks. This advanced course builds on Machine learning with neural networks (FFR135) and provides an in-depth analysis of many of the concepts and algorithms that were briefly introduced in that course, with particular emphasis on applications in the natural and engineering sciences. The goal is to become familiar with several advanced machine-learning methods, and to code them efficiently in Python using current neural-network packages. An essential part of the course are projects in deep learning, recurrent learning, and reinforcement learning.
Schedule
See below, the TimeEdit schedule is NOT fully correct.
Course literature
There is no single book that covers all the topics in the course.
Suggested reading for background: Bernhard Mehlig, Machine Learning with Neural Networks. Cambridge University Press (2022) www.chalmersstore.se or www.adlibris.com
In addition we provide a set of compulsory (and additional) readings assigned to each lecture that covers and broadens the content of the lectures.
Course design
Important!
Most classes will be in person.
The remote classes and other activities will be with this link:
https://gu-se.zoom.us/j/61054184445?pwd=emJzSDVNMWlsZStGVUlDN2ZWOUVNdz09
Meeting ID: 610 5418 4445
Passcode: 537945
| Topic |
Readings / Materials to prepare |
Additional readings | |
|
Tue Mar 22
|
Lecture 1 Introduction Course information |
|
|
|
Wed Mar 23 |
Lecture 2 Introduction to machine learning with neural networks |
Bernhard Mehlig, Machine learning with neural networks. Cambridge University Press (2022). |
Slides |
|
Fri Mar 25 |
Lecture 3 Introduction to Graph neural networks |
Two nice YouTube videos on the topic: Theoretical Foundations of Graph Neural Networks Deep learning on graphs: successes, challenges, and next steps | Graph Neural Networks Mats lecture notes. (coming) |
Two papers that give an overview of the subject: A comprehensive survey on graph neural networks. Benchmarkinggraph neural networks Application to drug discovery: A Deep Learning Approach to Antibiotic Discovery.pdf For more in-depth: |
|
Tue Mar 29 |
Lecture 4 Introduction to Q-learning and Deep Q-networks [slides] |
Mehlig, Introduction to machine learning with neural networks (Sections 11.0 and 11.3) Sutton & Barto, Reinforcement Learning, An Introduction (Section 16.5) |
Mnih et al., Human- level control through deep reinforcement learning. Nature 518, 529–533 (2015) |
|
Tue Mar 29 10:00-11:45 |
HW help session |
|
|
|
Wed Mar 30 |
Lecture 5 |
Original paper introducing transformer: https://arxiv.org/pdf/1706.03762.pdf
|
Youtube video explaining recurrent neural networks: Youtube video explaining transformers and language processing: https://youtu.be/S27pHKBEp30 |
|
Fri Apr 1 8:00-9:45 |
Lecture 6 Recurrent nets and reservoir computers |
Bernhard Mehlig, Introduction to machine learning with neural networks (Chapter 9) |
Pathak et al. (2017) Using machine learning to replicate chaotic attractors and calculate Lyapunov exponents from data |
|
Fri Apr 1 10:00-11:45 |
Christofer Åkerström (Invited speaker) Syntronic |
|
|
|
Fri Apr 1 |
Project Group Announced |
The project groups will be formed randomly. If, for a very good reasons, this is not optimal for you, inform the instructor before Tue Mar 30. |
|
|
~Apr 1-15 |
1st meeting with tutor |
Arrange a 45-minute meeting with tutor to discuss and receive feedback on the project idea and outline |
|
|
Tue Apr 5 |
Daniel Langkilde (Invited speaker) |
|
|
|
Fri Apr 8 |
HW help session |
|
|
|
Fri Apr 22 |
HW help session |
|
|
|
Wed Apr 27 |
Carolina Wählby (Invited speaker) |
|
|
|
Fri Apr 29 |
HW correction - session 1 |
|
|
|
Fri Apr 29 |
Project Proposal |
Submit your finalized project proposal |
|
|
~May 1-20 |
2nd and 3rd meetings with tutor |
Arrange 45-minute meetings with tutor to discuss and get feedback on the progress of your project and its presentation |
|
|
Wed May 4 |
Antonio Celani (Invited speaker) |
|
|
|
Fri May 6 |
HW correction - session 2 |
|
|
|
Tue May 10 |
Jonas Einarsson (Invited Speaker) |
|
|
|
Fri May 13 |
HW correction - session 3 |
|
|
|
Tue May 17 |
Vilhelm Verendel |
|
|
|
Fri May 20 |
Draft of Project Report |
Submit a draft of your project report |
|
|
Mon May 23 |
Project Presentation |
Submit your project presentation (8-minute video, format MOV) |
|
|
Wed May 25 |
Project Presentation |
|
|
|
~May 26-31 |
4th meeting with tutor |
Arrange a 45-minute meeting with tutor to get feedback on your project draft and presentation |
|
|
Fri Jun 12 |
Project Report |
Submit your final project report |
|
Learning objectives and syllabus
Learning objectives:
Knowledge and understanding
- Describe the different available neural network models with their advantages and disadvantages
- Find relevant literature to keep up with this quickly advancing field
Skills and ability
- Implement a broad range of state-of-the-art neural network models
- Train and validate these models
- Optimize these models for a specific task
- Plan, manage and execute a small scale project in the field
- Write a report of their results of the project
Judgement and approach
- Critically analyse the advantages and disadvantages of the available neural network models
- Benchmark the results of a neural network models against other models
- Critically evaluate and discuss advances in the field of neural networks
Link to the syllabus on the Chalmers Studieportalen:
Link to the syllabus on the GU website:
https://utbildning.gu.se/kurser/kurs_information?courseid=FYM750
Examination form
There will be no written final examination in this course.
The examination consists of homework assignments and a group project.
Specifically, the examination is based on
- 30% homeworks (10% for each)
- 20% final project presentation
- 50% final project report
A necessary (but not sufficient) requirement for passing grade is that 5/10 points are achieved in each homework.
Homework information
The purpose of the homework problem sets is to provide you with hands-on experience of advanced deep learning techniques. You are strongly encouraged to team up and collaborate on the homework, to ask your classmates if you are stuck at some point, and to assist classmates in need of advice. But you must write and run your own code, and have your own work assessed. No late assignments will be accepted.
Grading: Out of the four available homeworks, you need to solve and get assessed on three homeworks.
Each homework accounts for 10 points. A necessary (but not sufficient) requirement for passing grade is that 5/10 points are achieved in each homework.
Overall these three homeworks account for 30% of your total course grade (10% for each).
Computer labs: The computer lab sessions serve two purposes:
1) To get help if you are stuck in a problem set (and have not managed to get unstuck by consulting some of your classmates). The time we have available for computer lab is limited, so it is essential that you come prepared to the lab, with specific questions you need to have answered.
2) To have your homework graded.
Assessment: We will assess your work during lab hours.
Be prepared for the assessment: Generate all your figures and/or videos up front. There will be no time for running code during the assessment. Check that you have answered all questions carefully. You should be able to give reasoned answers regarding your programming choices as well as being able to discuss the implications of your results.
Project information
Group advanced deep learning project
The purpose of the group project work is to provide you with training in
- Developing your own small research-style project.
- Executing a collaborative project.
- Presenting your independent work in writing and as a summary talk.
The deliverables are a 8 minute presentation + 10 minute questions (20% of the course grade) and a written report (50% of the course grade).
The deadline for submitting a project proposal to your tutor (see below) is by 23:59, April 29, 2022. However, it is probably good for you to have it done earlier. Groups with late submission may not receive tutoring.
The draft report is due by 23:59, May 20, 2022.
The presentation video is due by 23:59, May 23, 2022.
The project will be presented to the class in week 21 (23-27 May 2022).
The final written report is due by 23:59, June 12, 2022. No late reports will be accepted.
General information
You will be randomly divided into groups (exceptions are possible only for very good reasons).
The reason for this randomness is twofold:
(1) it will assure a broad distribution of skills and interest;
(2) it is a realistic scenario for later work in academia or industry.
- Each group will be paired with another group, and they will act as review groups for each other.
- Each pair of groups gets a tutor assigned and has 4 meetings, 45 minutes each, with this tutor during the course.
- The tutor’s job is to give you general advice on anything you need as best as he or she can. The tutor’s job is not to formulate your project for you, nor to debug your code.
- Should you experience problems with the group assembly, you need to take this up with your tutor as soon as possible.
To start your project you need a challenge you will address. You can get a challenge in one of the popular challenge websites, such a Kaggle, CodaLab, etc.
The first meeting with the tutor is to ensure that you’re embarking on a feasible project. Before this meeting you should:
- Identify your challenge.
- Write a two-paragraph proposal: one explaining the background of the project, possibly with references to relevant literature, and another outlining what you aim to do and why.
- Discuss your project with your review group and revise your proposal accordingly.
- Send the proposal to your tutor.
- Prepare for the meeting; prepare questions, comments, simulations, results, or whatever is relevant for a most efficient outcome of the meeting.
The following meetings with the tutor are to ensure that you have progress and direction so that you are on course to completion. Before these meetings you should:
- Have ready an implementation of your model.
- Have a set of preliminary results from simulations.
- Have ready the outline of a report.
- Have met with your review group to discuss model, results and report outline.
- Send the report outline to your tutor.
- Prepare for the meeting; prepare questions, comments, simulations, results, or whatever is relevant for a most efficient outcome of the meeting.
Presentation
The presentation of the project should take maximum 8 minutes and be prerecorded (MOV format), with an additional 10 minutes for questions and discussion.
Put the emphasis on what your problem formulation is, a general discussion about how you tackled it, what problems you had, and what your conclusions are.
Report
The report should be written in RevTex 4.2 (see https://journals.aps.org/revtex), using the format of Physical Review X with a limit of 8 pages in double-column reprint format (including pictures and references). If needed, you can have an extra document containing an unlimited amount of Supplementary Materials.
Structure your report like a scientific article, with an abstract summarising the rationale and results of the project; an introduction shortly explaining its background and motivating why the question is interesting; methods and/or results section(s) describing your model and what you do with it; a discussion section; and ending with a conclusion. Put some work into the discussion and conclusion sections. This is where you demonstrate that you truly understand the implications of your work, including shortcomings and uncertainties. It is important that the discussion does not fall out as a simple summary of what the figures show.
Write in plain language and write enough to say what you need to say. Don't think "the report feels short, better throw in some extra figures." If you can say the same thing in fewer words, do so.
Be sure to reference any source you use. The report should be readable and understandable on its own, but there is no need to reproduce for example derivations of equations from your sources; a citation is enough.
Figures are an important part of the report, but only those that substantially contribute to your analysis should be included. Make sure that each figure is well designed with informative captions (not just "Fig. X shows how quantity A depends on quantity B"). If you find it hard to do this, you probably are not clear on why you include an figure, so cut it out.
Most importantly, use your own judgement and try to write a report you would like to read.
Evaluation criteria for report
Criteria that will be considered are:
- Structure: Is the project report well structured (see above)?
- Motivation: Does the project have a clear research question?
- Originality: How original is the research question?
- Background: Is the background of the research question well explained, including references to the literature?
- Methods: Are the methods and analysis appropriate for convincingly answering the research question?
- Execution: Does the project answer the initial or iterated research question?
- Figures: Are the figures included informative with descriptive captions?
- Format: Is the format of the report as required?
- Outlook: Are the implications of the results clearly stated?
- Contributions: Is there a contributions statement in the report?
Note that the research question does not have to be answered in the affirmative. A negative result is equally valid as long as an analysis is carried out that properly explains the negative result and what eventually would work better.
Evaluation criteria for presentation
Criteria that will be considered are:
- Structure: Is the presentation well structured for a 8-minute presentation?
- Clarity: Is it clear from the presentation what is being investigated?
- Format: Does the presentation fit the requested format?
- Motivation: Is it explained well why the project is important?
- Conclusion: Are conclusions and implications clear?
- Methods: Are the methods explained sufficiently but not excessively?
- Presentation quality: Is the presentation engaging and visually appealing?
Calculation of final grade
You can get at most 30 points from the homework and 70 points from the project. The maximum combined score is 100.
The grade limits will be 50 points for grade 3, 70 points for grade 4, and 90 points for grade 5.
For GU the limits will be 50 points for G and 80 for VG.
Course summary:
| Date | Details | Due |
|---|---|---|