Edwige Cyffers

Phd Student in Privacy Preserving Machine Learning

Dimension Reduction

Responsible of 24-hour lecture in Master of Machine Learning

During my PhD, I have the opportunity to teach at University of Lille. I was lucky to get a Master course, with the freedom to design lectures and tutorials.

Part of a computer science Degree specialized in Machine Learning, I aim to give students a better intuition of high dimension, and of the trade-off between dimension reduction and preservation of data characteristics. The goal is also to strengthen their mathematical background and their ability to code in Python. The lecture covers the curse of dimensionality, PCA, SVD, kernel-PCA, manifold-learning, LLE, t-SNE, spectral clustering…

I continue this year to get involved in the Rendez-vous des jeunes mathématiciennes et informaticiennes. We now include more computer-science-oriented tutorials, and ot only mathematic ones. I was looking for a topic easy to understand and to explain to parents and relatives. Maze generation was looking fun and reasonable for 3-hour work.
We generate mazes with two different algorithms. For small mazes, the construction can be seen step by step with animation. We finish with mazes following the form of a picture.
You can find the notebook that we use during the tutorial (in French).
Once again, I was surprised by the quality of the questions and solutions found by the girls. Let’s share also share a fun fact. I got one expected question that I wasn’t prepared to. How can we draw the figure in pink? Well, usually, I’d prefer to use other colors, but we can look at the possibilities. Is purple enough? (You can see the result)

Ethics of algorithms

12 hours tutorial

I design three practicals from scratch for people coming from social sciences, with Python programming. I taught this tutorial at the University Paris Eiffel in 2021.

  • Compas example: How to compute the fairness metrics in this simple but interesting example.
  • Introduction on Privacy: what is re-identification, and can we do it at scale ? Which solutions ?
  • Bias in NLP: why and how we keep historical gender bias.

The goal is to open some black boxes and give them some tools that they could reuse in other projects.

Recent Posts