News
Currently, no news are available
Neural-Symbolic Computing
The way our brain forms thoughts can be classified into two categories (according to Kahneman in his book “Thinking Fast and Slow”):
System 1: fast, automatic, frequent, stereotypic, unconscious.
- Is this a cat or a dog?
- What does this sentence mean in English?
System 2: slow, effortful, logical, conscious.
- 17*16 = ?
- If a -> b does b -> a?
The traditional view is that deep learning is limited to System 1 type of reasoning. Mostly because of the perception that deep neural networks are unable to solve complex logical reasoning tasks reliably. Historically, applications of machine learning were thus often restricted to sub-problems within larger logical frameworks, such as resolving heuristics in solvers.
In this seminar, we will explore new research that shows that deep neural networks are, in fact, able to reason on “symbolic systems”, i.e., systems that are built with symbols like programming languages or formal logics.
Example Topics:
- What are your chances against an AI in a programming competition?
- Is it possible to teach temporal logics to neural networks?
- Can neural networks learn the intuition of mathematicians to improve automated theorem proving?
Requirements
Participants should be strongly interested in logical reasoning and/or machine learning. There is, however, no formal prerequisite.
Organization
The structure of the seminar is as follows:
In the first week of the seminar, you will choose a research topic on neural-symbolic computing. The topics include graph neural networks, reinforcement learning, and transformers. You have a selection of two neuro-symbolic papers on each topic.
Topic Phase: In the first phase, you will prepare an informal lecture on your selected machine learning topic and background, presenting the basics to your fellow students. You are assigned to an advisor who will help you with your preparations. This informal lecture will not be graded so that you can see it as a rehearsal. Thus, this phase gives you and your fellow students the foundations for the following paper talks.
Paper Phase: With the help of your advisor, you will prepare a research talk on one of the two papers in your chosen topic. This talk is weighted most in your final grade.
Project Phase: You will be given a neural-symbolic computing task. Each team must solve this problem using the methods and ideas explored in this seminar, for example, by applying deep neural network architectures. This project has to be passed, and it will not be graded.
Dates
Kick-off meeting: Tuesday, April 23rd, at 4:00 pm
Weekly meetings: Tuesdays at 4:00 pm