Research
Explanations in AI as Claims of Tacit Knowledge
An article resulting from my master’s thesis, in which I argue that the philosophical concept of “tacit knowledge” can be applied to AI systems under certain conditions, and that this can be used to provide reliable explanations of their behavior.
Available via Minds and Machines or read directly here.
Black-box systems, multi-level explanation, and cognitive architectures
My (philosophical) master’s thesis in Artificial Intelligence from December 2020, supervised by Brandt van der Gaast. It is about explainable AI, the field that tries to make black-box machine learning systems understandable. I argue for an mechanistic approach to explanation, where an explanation is justified by a demonstrable causal connection between a mechanism in the system and the result to be explained. For inspiration on how a complex system could be modularized at a high level to demonstrate such a causal connection, I look at cognitive architectures from psychology, such as Soar, CLARION and ACT-R. The article above came out of this thesis.
Dimensionality reduction and trajectory grouping
A Complex Systems research project at Utrecht University from August 2020. The trajectory grouping problem is about finding groups among moving entities from their trajectory data, like animals moving in packs. I tried to tackle it with dimensionality reduction, comparing PCA, ICA and a custom clustering procedure. I evaluated them on synthetic data from the Kuramoto model, with the method by Buchin et al. as a baseline. PCA turned out to predict the number of groups accurately. The group structure itself was not clearly accurate enough to replace the baseline method, but the dimension grouping algorithm does provide an interesting general approach to the problem.
Read the report here. The code is on GitHub as DimensionGrouping.jl and Kuramoto.jl.
Determining suitable models for C. elegans locomotion by evaluating Lyapunov exponents
My bachelor’s thesis in Physics and Astronomy at the UvA and VU from July 2018, supervised by Greg Stephens. Earlier work showed that the movement of the worm C. elegans follows an attractor in a 6-dimensional phase space, with a symmetric spectrum of Lyapunov exponents. I simulated two standard dynamical systems to see if either could produce a similar spectrum. The first was a pair of coupled Nosé-Hoover oscillators, and the second a pendulum driven by a periodic force. The driven pendulum can’t produce a symmetric spectrum without extending it in some way. Coupling the oscillators also breaks the symmetry, but they show a wide range of behavior, so a different coupling might work. Chapters 2 and 3 give an accessible introduction to dynamical systems, Lyapunov exponents and how to compute them.
De Gebarende Wijs: Nederlandse zinnen met persoonlijk voornaamwoorden omzetten naar Nederlandse Gebarentaal
My bachelor’s thesis in Computer Science (Informatica) at the UvA from June 2018, supervised by Floris Roelofsen. It is written in Dutch. I looked at whether the Grammatical Framework (GF) can serve as the base for a translation program between a fragment of Dutch and Dutch Sign Language (NGT), focusing on personal pronouns. I wrote formal grammars for both languages as multiple context-free grammars and implemented translators with GF. Correct input sentences were translated correctly, but the grammars overgenerate and cover only a small part of each language. The thesis also has what I think is a nice introduction to multiple context-free grammars and GF.