Course syllabus
Course-PM
DAT615 DAT615 Neuro-symbolic AI lp2 HT24 (7.5 hp)
Course is offered by the department of Computer Science and Engineering
Contact details
- Lecturer/Examiner: Moa Johansson (moa.johansson@chalmers.se)
- Lecturer: Devdatt Dubhashi (dubhashi@chalmers.se)
- TA: Andrea Silvi (andrea.silvi@chalmers.se)
- TA: Sólrún Halla Eínarsdottir (slrn@chalmers.se)
For questions about their lectures, contact the respective lecturer. For questions about assignments, contact the TA.
Student Representatives
- MPCAS marcushansen56@gmail.com Marcus Hansen
- MPENM mengfanbo0920@outlook.com Fanbo Meng
Course purpose
The purpose of the course is to introduce the students to the field of neuro-symbolic AI. Recent advances in neural networks have quickly moved the state of the art forward for many applications, not least natural language processing and computer vision. But neural models still have weaknesses, for instance when it comes to reliably reasoning logically or planning. This is were methods from symbolic AI (e.g. rule based and heuristic methods) still dominate. The aim of neuro-symbolic AI is to combine the two approaches to get the best from both worlds. It will also allow us to tackle problems which neither method on its own is particularly good at. In this course, we will focus on two application areas: cognitive modeling and languages for emergent communication (with Dubhashi), and program synthesis and mathematics (with Johansson). Towards the end of the course, students should have sufficient knowledge to be able to follow recent research papers in these areas. The final two weeks are dedicated to presentation and discussion of such state of the art research articles.
Schedule
Please refer to TimeEdit for details of the schedule. Note that the rooms are different each week (for reasons beyond our control). Below is a rough overview of the lectures:
| Week | Tuesday 10.00 (Johansson) | Friday 10.00 (Dubhashi) |
| 1 | Introduction to neuro-symbolic AI | |
| 2 | Introduction to Program Synthesis | Crash course in RL + MCTS |
| 3 | Neural and Probabilistic Methods for Synthesis | Sparse rewards, exploration, exploration in deepRL, credit assignment |
| 4 | Library Learning: Invention of New Concepts | Intrinsic Motivation and RL, intrinsic motivation and information theory |
| 5 | AI for mathematics | Emergent commuication and Efficient Communication |
| 6 | Neuro-symbolic AI in the International Math Olympiad | Learning Math with RL, Intrinsic motivation for math. |
| 7 | Guest lectures on neuro-symbolic methods in current research. |
Research papers: |
Office Hours with TA: Getting help with Assignments
If you want to get help with the assignments you may come to the Office Hours and speak to the TA. Starting on 20 November these are in EDIT building, room EDIT 5128 on:
- Week 3, Wednesday and Friday 20/11/24 and 22/11/24, 13:00-14:00
- Week 4, Friday 29/11/24, 14:00-15:00
- Week 5, Friday 6/12/24, 13:00-14:00
- Week 6, Wednesday 11/12/24, 13:00-14:00
- Week 7, Wednesday 18/12/24, 16:00-17:00
Should you have additional questions about the assignments outside these times, please contact Andrea Silva via email (see contact details above). Canvas direct messages is not preferred. While we of course aim to reply quickly to emails, please note that the TA may need up to 48 hours, so please don't leave it too close to the deadline. You can only expect to get replies during office hours, and not during evenings and weekends (this applies to all teachers).
Course literature
Links to reading material related to each lecture will be linked to from each lecture page. After the lecture, you will also find copies of the slides there.
- Motivation
- B. Lake et al, "Building Machines that Learn and Think Like People", Behavioural and Brain Sciences, Vol. 40, 2017.
- K. Collins et al, "Building Machines that Learn and Think With People", 2024.
- F. Chollet, "On the Measure of Intelligence", 2019.
- H. Kautz, "The Third AI summer Download The Third AI summer", AI Magazine (2022).
- RL
- E Pignatelli et al, "A Survey of Temporal Credit Assignment in Deep Reinforcement Learning", TMLR 2024
- Program Synthesis
- S. Chaudhuri et al "Neuro-symbolic Programming" (2021).
- Gulwani et al "Program Synthesis" (2017).
- AI in Mathematics
- Kevin Buzzard "Mathematical Reasoning and the Computer", Bull. Amer. Math. Soc. 61 (2024).
- Bulletin of the American Mathematical Society 61 (2024): This special issue contains many interesting articles on the topic of AI in mathematics from a mathematicians perspective, in addition to the one by Buzzard above.
- Mirzadeh et al, "GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in Large Language Models" (2024).
[MORE TO COME]
Course design
The course has two lectures per week. The lectures will introduce topics from both symbolic AI, machine learning and neuron-symbolic AI. Students are expected to attend lectures and are encouraged to read the material linked from each lecture page. To contact the teachers, please use email. There might be a delay in getting a reply to messages sent in Canvas.
There will be three hand in programming assignments. These should be done in pairs and submitted via Canvas by the deadline stated. They are not strictly obligatory, but if completed well, will contribute to the final grade (see below under Examination for details). Assignments handed in before the deadline will be graded. There will be no second submissions. Deadlines are firm, unless granted an extension by the examiner for valid reasons such as illness. Vacation trips, failure to read the deadline dates etc does not count as valid reasons. It is the student's responsibility to keep themselves informed about the course by reading updates and announcements on the course Canvas page frequently.
Dates for the three assignments are:
- Assignment 1: Release Friday week 2 (15/11). Due Monday 25/11.
- Assignment 2: Release Tuesday week 4 (26/11). Due Monday 9/12.
- Assignment 3: Release Friday week 5 (6/12), Due Friday 20/12.
Changes made since the last occasion
This course is given for the first time in 2024.
Learning objectives and syllabus
Learning objectives:
After completion of the course the student should be able to:
- Separate what characterise symbolic and neural AI.
- Explain what neuro-symbolic AI encompasses.
- Apply and implement methods and algorithms for neural AI
- Apply and implement methods algorithms for symbolic AI.
- Apply and implement neuro-symbolic AI methods and algorithms.
Link to the syllabus on Studieportalen.
https://www.chalmers.se/en/education/your-studies/find-course-and-programme-syllabi/course-syllabus/DAT615/?acYear=2024/2025
If the course is a joint course (Chalmers and Göteborgs Universitet) you should link to both syllabus (Chalmers and Göteborgs Universitet).
Examination form
The course will have a hall exam in January and a re-exam in August. No aids are allowed. The exam is compulsory.
The final grade is calculated based on a combination of the final exam score and your score on homework assignments:
- Your total score on homeworks will be halved from a maximum of 30 points to a maximum of 15, call this x.
- Your score on the exam z, out of max points 60 will be normalized to a max of 45, y = 3/4 z.
- Your final score s = max(z, x+y). Grades will be based on s:
- Chalmers: 3 (28 points), 4 (36 points) and 5 (48 points)
- GU: G (28 points), VG (48 points).
Note that:
- Assignments are not mandatory but strongly recommended for understanding!
- Assignment bonus points (max 15) are only valid for exams in this course round: i.e the January 2025 and August 2025 exams, not for any subsequent reexams in later years.
Examination dates and locations: Please refer to https://www.chalmers.se/en/education/your-studies/examinations-and-other-summative-assessments/find-examination-dates/?search=DAT615
Course summary:
| Date | Details | Due |
|---|---|---|