Pinson, H., Lenaerts, J. & Ginis, V. (2023, July). Linear CNNs Discover the Statistical Structure of the Dataset Using Only the Most Dominant Frequencies. In International Conference on Machine Learning (pp.27876-27906). PMLR.
Hannah Pinson
Assistant Professor of AI
at Eindhoven University of Technology
News
15/11/2025: Shawon Ashraf joined the group as a PhD student, co-supervised by Joaquin Vanschoren. Welcome Shawon!
01/11/2025: Yannis Kaltampanidis joined the group as a PhD student on the MedGPT project. Welcome Yannis!
30/07/2025: end of our ‘deep learning’ course, which I taught together with Vlado Menkovski for 200 students at the master level. We received excellent student evaluations (8.3). On to the next year!
20/07/2025: new paper at the ICML workshop on High-dimensional Learning Dynamics by Aurélien: The Interplay Between Implicit Bias and Adversarial Robustness in Linear Convolutional Neural Networks
25/03/2025: start of the MedGPT project (13.3 million euros): “To create safe and ethical AI applications, we need to better understand how AI models work. Our research on the basics of AI models will help us improve their privacy, robustness, and compliance, especially for healthcare applications”
01/09/2024 – 31/1/2025: maternity leave to welcome a little girl
20/7/24: new paper at the ICML workshop on High-dimensional Learning Dynamics: Exploring the development of complexity over depth and time in deep neural networks
01/07/2024: my 30min talk on how simple CNNs learn and which information about the dataset they store in their parameters, is now on youtube.
01/05/2024: Mudassar Ahmad joined the group as a summer research intern. He’ll try to extend our insights in learning with attention mechanisms. Welcome Mudassar!
15/01/2024: Aurélien Boland joined the group as a PhD student. Previously he obtained a M.Sc. degree in mathematical engineering, and an advanced M.Sc. degree in AI. He also worked two years as a data scientist in the industry (AGC). Now he’s interested to tackle questions about what and how neural networks learn. Welcome Aurélien!
Research
The goal of my research is to obtain fundamental insights in deep learning:
- How do neural networks really work?
- What do they learn, and when do they learn it?
- Why does it sometimes go wrong?
In more technical terms, I try to solve questions about inductive bias, learning dynamics, and generalization. Currently I have a specific focus on understanding the interplay between dataset structure, neural network architecture and gradient descent.
My main subject
I am interested in obtaining fundamental insights in deep learning. My current focus is on questions about inductive bias, implicit regularization and generalization, and I mainly study the interplay between dataset structure, neural network architecture and gradient descent.
I also teach about the use and understanding of neural networks, and I supervise several research projects in AI.
Previous experience
I obtained a master’s degree in Physics and one in Computer Science/AI, both at the Vrije Universiteit Brussel (VUB), Belgium. My master’s studies were completed with a year-long research stay with the Max Tegmark group at MIT in 2017. I later obtained the Agoria award for the best thesis in technology and innovation for my thesis project.
In 2018, I started my PhD as an FWO fellow under the supervision of Prof. Vincent Ginis at VUB. During my PhD I remained a visiting researcher at MIT, and I also learned to experiment with real neurons and analyze neural network recordings in the labs of Prof. Michael Levin at Tufts University and of Prof. Mustafa Sahin at Harvard Medical School. In June 2023, I defended my PhD thesis with the title Form and Function in Neural Networks: On the Interplay Between Dataset Structure, Network Architecture, and the Evolution of Learning.
Collaborations
During my studies and the beginning of my PhD, I enjoyed collaborations on a wide range of different topics, from photonics to medicine to neuroscience. The two main papers of this period are:
Sundberg, M., Pinson, H., Smith, R. S., Winden, K. D., Venugopal, P., Tai, D. J., … & Sahin, M. (2021). 16p11. 2 deletion is associated with hyperactivation of human iPSC-derived dopaminergic neuron networks and is rescued by RHOA inhibition in vitro. Nature communications, 12(1), 2897.
Pinson, H., & Ginis, V. (2019). General framework for the frequency shifting of electromagnetic pulses using time-dependent surfaces.
Physical Review B, 99(20), 201407.