Lukas F. Münzel

Hi there!

I’m fascinated by many things, in particular by how intelligence can emerge from simpler principles (i.e. single neurons or a few hundred lines of code). My interests are accordingly rather interdisciplinary, from semiconductor Physics over neurobiology to information theory That is not to say that I have significant background in all of these fields. I suppose a unifying theme of the sort of positions I can imagine myself in, say, five years is the need for a solid mathematical foundation (and most probably coding, which arguably is one of the most powerful tools of the 21st century so far) and exciting and potentially impactful research questions. Specifically, I want to work on making machines more intelligent and research how to best harness the thereby generated, potentially enormous capabilities.

If that sounds interesting to you, I’d love to connect :)

Sincerely,

Lukas

Do you know me personally? Please give me anonymous feedback here

Short summary of my CV:

In 2021, I had the pleasure of participating in the Research Science Institute, a six-week research program held at MIT with amazing people from all over the world
I participated in the Informatics Olympiad with reasonable success, being one of the four people to represent Switzerland internationally in 2020 and 2022 respectively. I also qualified for the final rounds of the Swiss Biology and Physics Olympiad

some of my work

Radiograph Automata - A novel model of computation

Lukas Münzel

Jul 2023

PDF
The Influence of the Orthogonality of Feature Functions on Artificial Neural Networks

Lukas Münzel

Oct 2021

Abs PDF (English) PDF (German) Presentation recording

Conventional methods for adjusting the weights of artificial neural networks only consider the output layer when calculating the loss. Our method also takes into account infor-mation from the functions generating the outputs of the neurons in the second to last layer,which are called feature functions. Specifically, we measure how orthogonal the feature functions are using a covariance measure. We incorporate a term in the loss to explicitly rewardorthogonal feature functions. We were then able to generate nearly orthogonal feature functions without compromising on accuracy. Due to the convenient mathematical properties oforthogonal feature functions, we will in a future next step be able to find the optimal choiceof weights between the second to last and the output layer with low computational cost.

To cuckold and to be cuckolded - An essay on parallels between Shakespeare’s Othello and Greek mythology

Lukas Münzel

Nov 2021

Bib PDF

@article{paralells_othello_odyssey,
  title = {To cuckold and to be cuckolded - An essay on parallels between Shakespeare’s Othello and Greek mythology},
  author = {Münzel, Lukas},
  year = {2021},
  month = nov,
}