Onderwerpen voor het profielwerkstuk
Lesgeven met behulp van AI
Begeleider
Martijn Boussé
Introductie
Hoe kunnen we AI-taalmodellen, zoals ChatGPT, gebruiken om de leerervaring van studenten te verbeteren? Het klaslokaal verandert en studenten leren niet langer alleen van een leraar voor de klas, uit boeken of na het invullen van oefenboekjes. Leren kan (veel) actiever zijn en kan worden verbeterd door recente AI-tools. Het doel van deze opdracht is om nieuwe manieren te bedenken om taalmodellen buiten de klas te gebruiken om de leerervaring te verbeteren. Studenten kunnen nieuwe interactieve en door AI ondersteunde formats, opdrachten of studietechnieken voorstellen. Op deze manier leren studenten over recente taalmodellen, hun mogelijkheden en beperkingen, en hoe ze op de juiste manier kunnen worden gebruikt voor leren buiten het klaslokaal.
Op te leveren
In de beginfase analyseren studenten bestaande leer- en onderwijsopzetten en identificeren mogelijke verbeterpunten met behulp van taalmodellen. Om dit te doen, zullen ze de literatuur verkennen, experts aan de universiteit interviewen, studenten enquêteren, docenten observeren, enz. Vervolgens zullen ze een nieuw AI-ondersteund leer- en/of onderwijsopzet voorstellen met een concreet implementatieplan. Ze passen hun opzet toe in hun eigen onderwijs en vatten de resultaten samen in een rapport, waarin ze onderzoeksvragen stellen en deze op wetenschappelijke wijze beantwoorden. Het rapport moet hun keuzes duidelijk motiveren.
Bronnen
https://onderwijscommunity.nl/artikelen/chatgpt-ai-in-het-onderwijs/
“Ik heb gelijk”: hoe overtuig ik iemand?
Supervisor
Tjitze Rienstra, Nico Roos
Introduction
Argumentatie is een belangrijk onderzoeksgebied in Artificial Intelligence dat brede toepassingen heeft buiten de AI. Op veel plaatsen wordt argumentatie gebruikt om geschillen op te lossen. Denk bijvoorbeeld aan de rechtspraak, het parlement (in het ideale geval), projectgroepen, enz. Door partijen worden argumenten en tegenargumenten aangedragen. Binnen de AI beschrijven we dit d.m.v. een aanvalsrelatie tussen argumenten. Op grond van deze aanvalsrelatie kunnen we beslissen wat (mogelijk) winnende argumenten zijn. De inhoud van de argumenten blijkt daarvoor niet langer van belang te zijn.
Het doel van deze opdracht is het bouwen van een systeem waarmee we argument kunnen invoeren en vervolgens de aanvalsrelatie tussen de argumenten kunnen aangeven. Vervolgens moet het systeem (mogelijk) winnende argument presenteren als ook aangeven waarom dit winnende argumenten zijn. Verder kan men onderzoeken of AI language models, zoals ChatGPT, de aanvalsrelatie tussen argumenten kan identificeren. Hierbij is de inhoud van de argumenten wel van belang.
Deliverables
In de initiële fase bekijken de deelnemers de theorie van “abstract argumentation frameworks” en specifiek de ‘labeling approach’ voor het evalueren van argumenten. Vervolgens realiseert men een systeem waarmee argumenten ingevoerd kunnen worden en de aanvalsrelatie tussen argumenten gespecificeerd kan worden. Tot slot wordt het algoritme voor het evalueren van argument gerealiseerd. Parallel hieraan wordt bekeken of AI language models de aanvalsrelatie tussen argumenten kan identificeren.
Sources
- Dung, P.M.: On the acceptability of arguments and its fundamental role in nonmonotonic reasoning, logic programming and n-person games. Artificial Intelligence 77, 321–357 (1995)
- Diehl, D.J.S.: Labellings and their algorithms for Abstract Argumentation Framework Semantics, Bachelor Thesis Artificial Intelligence, Universiteit Utrecht.
Can a machine tell a funny joke?
Supervisor
Dr. Aki Härmä (Supervision in English only.)
Introduction
How can we use large language models, such as ChatGPT, come up with a joke that is actually funny and novel? Can you prompt GPT to generate actually surprising and even inappropriate jokes? Can GPT predict how many people out of ten finds a joke funny and can we test if it is correct?
Deliverables
In the initial phase, it is good to look into the literature about jokes and AI-generated jokes in particular, see [1]. What humor actually is and what makes a good joke [2]? Next, perform a series of experiments using OpenAI GPT-3.5/4, or any other model, using a selection of prompts to generate a collection of AI-generated jokes. Score the jokes with a panel of students by the funniness on scale 0-5. Ask GPT to score the same jokes on the same scale and analyze the differences between the humans and machine. Next, can you improve the prompts so that the model produces better jokes? Can you also make it generate jokes that are inappropriate in one way or another? Finally, summarize the results in a report. The work can be done in English or Dutch.
Sources
- [2306.04563] ChatGPT is fun, but it is not funny! Humor is still challenging Large Language Models (arxiv.org)
- Humour - Wikipedia
Een computermodel van ons zonnestelsel
De aarde, en de andere planeten draaien om de zon in een ellipsvormige baan – de aarde doet er ongeveer een jaar over om een rondje vol te maken. Die beweging wordt beschreven door de wetten van Kepler, en wordt verklaard door de wetten van Newton: De universele zwaartekrachtswet beschrijft de aantrekkingskracht tussen planeten en de zon, en de tweede wet linkt kracht en beweging.
In dit project maak je een computersimulatie van ons zonnestelsel: je toont hoe deze beide grondwetten van de fysica tot de elliptische banen van de planeten in ons zonnestelsel leiden. In je model kan je dan de baan van een satelliet of komeet berekenen, of een botsing tussen planeten simuleren, enz.
Download de hele profielwerkstukbeschrijving
Game Theory in the Classroom
Supervisor
Gijs Schoenmakers (Supervision in Dutch or English)
Introduction
Game Theory is the research field that describes and analyzes situations where two or more people, who may or may not have conflicting interests, make strategic decisions and where everyone tries to maximize their own outcome. Such a situation is called a game.
The goal of this assignment is to help the classmates of the students increase their ability to think and act strategically, by having them play some game theoretic games. The students involved in this project should host these games and be able to explain to the classmates what they did right and wrong during the game play.
Deliverables
The project consists of two phases. In the initial phase, the students perform a study to get a grasp of basic game theoretic concepts like rationality, common knowledge, best replies, and equilibrium. The students should be able to solve basic game theory problems by the end of this phase.
During phase 2, the students go on a search for interesting games (from a game theoretic perspective) for their fellow classmates to play, or create some themselves. At the end of the project, the students let their classmates play these games, and explain what it is that they did right or wrong during the play of these games.
The results will be summarized in a report where they pose research questions and answer them in a scientific way. The report should clearly motivate their choices.
Sources
Mapping Misinformation on Social Media
Supervisor
Marijn ten Thij (Supervision in Dutch or English)
Introduction
In the current digital age, misinformation is a threat for society, as it erodes trust in democracy, increases polarization, and taps into extremist movements’ ideology. As a result, research on stopping misinformation is a very active field. However, this work focuses on a general audience, rather than members of fringe communities. Using an existing data set of public Facebook groups, the students will explore how these groups interact with fringe media. The goal of the project is to visualize the connections between different fringe communities in Western Europe based on their interactions with fringe media.
Deliverables
In the initial phase, students process an extended set of data extracted from facebook groups. The data set contains a large set of posts that are placed in groups that represent fringe communities in Western Europe. By extracting references to website from these groups, the necessary data can be extracted for the next phase. Based on the extracted website references, a network visualization of the connections between the fringe communities is constructed. Finally, the structural properties of this network can be studied.
Sources
Treat signals with care
Supervisor
Pietro Bonizzi (Supervision in English only)
Introduction
Signals are everywhere around us. The chirp of a bird, the music we listen to, the signals that our smart phones receive and send to allow us communicate with one another. A signal can be viewed as a “carrier of information”. In order to listen to music through our smart phones, or to be able to analyze a signal (like a gravitational wave) with a computer, we need first to transform a signal into a finite sequence of values. This is called signal digitalization, or signal sampling. This transformation must be performed properly, so that the characteristics of the original signal are preserved. Otherwise, an improper sampling can generate “aliasing”, which means that the signal’s properties have been altered. For instance, a music piece that eventually does not sounds “right”.
Deliverables
In the initial phase, students become familiar with what signals are, and how signals can be converted to sequences of values that can be stored into an electronic device, like a computer. Next, student will try to understand under what conditions a signal can be digitalized properly, and what is the impact of an incorrect digitalization of a signal. Next, the students will design a simple app or graphic user interface to explain the concept of signal sampling to laymen, and allow people to play with simple toy examples.
Students run their format in their own classroom and summarize the results in a report where they pose research questions and answer them in a scientific way. The report should clearly motivate their choices.
Sources
What can genes tell us?
Supervisor
Rachel Cavill (Supervision in English only)
Introduction
Biology is transforming into a science of BIG data. One of the easiest datatypes we measure is the expression of genes. Naively, each gene codes for one protein and the more a gene is expressed, the more of that protein will be made (in reality it’s a lot more complicated that this!). There are thousands of public gene expression datasets. In this project, we will pick a disease that interests you, find a relevant dataset and explore it using data science and machine learning.
Deliverables
You will be introduced to different data science tools for exploring data. Depending on their interests, these can either be web-based/ point and click applications, or we can explore writing programming code to do the analysis.
We will find a relevant dataset for the disease of interest.
We will look at what the biggest differences (changes in gene expression levels) between the samples are using different data science methods to visualize our data.
We will look at what the differences are between samples from sick people and healthy people. Can we predict who is sick from their gene expression?
We will look at which biological processes the genes that change are involved in and explore how this relates to what is already known about the disease.
Sources
Databases with datasets (not easy to use without an expert!):
Wisdom of Crowds
Supervisor
Barbara Franci, Martijn Boussé (Supervision in English only)
Introduction
Imagine you are a judge at a fair, where you ask the participants to estimate the weight of an ox. Those who guess most successfully receive a prize. You obtain the most diverse numbers, very few got close to the actual weight of the ox, but… the average between all the guesses is the closest than any individual guess! This phenomenon is known as the wisdom of crowds: a large group of people is collectively smarter than single individuals. It is then natural to wonder what would happen if the individuals can talk with each other and change their opinion.
The goal of this assignment is to investigate how the communication between the individuals affects the wisdom of crowd effect.
Deliverables
In the initial phase, the students analyse existing opinion dynamic mechanisms by exploring the literature and talking with experts. This will allow the students to formulate some hypothesis on what might preserve the wisdom.
Next, they will select a model and propose some experiments to run, either virtually or by selecting participants among their peers or external individuals. The students will summarize the results of the experiments in a report where they pose research questions and answer them in a scientific way.
Sources
Databases with datasets (not easy to use without an expert!):
- https://wisdomofthecrowd.nl/startpagina/mijn-onderwerpen/de-kracht-van-samen/wisdomofcrowds/
- https://www.nature.com/articles/075450a0
- Leren met behulp van AI: Buiten het klaslokaal
- Lesgeven met behulp van AI
- “Ik heb gelijk”: hoe overtuig ik iemand?
- Can a machine tell a funny joke?
- Een computermodel van ons zonnestelsel
- Game Theory in the Classroom
- Mapping Misinformation on Social Media
- Treat signals with care
- What can genes tell us?
- Wisdom of Crowds