Berfin Simsek

Dr. Berfin Simsek is a doctoral researcher at the École Polytechnique Fédérale de Lausanne (EPFL), affiliated with the School of Computer and Communication Sciences. Supervised by Clément Hongler and Wulfram Gerstner, her work focuses on the theoretical properties of deep learning models, including neural network loss landscapes, symmetry-induced saddles, global minima manifold geometry, and optimization dynamics in overparameterized regimes.

Education

• Double major in Electrical-Electronics Engineering and Mathematics at Koç University, Istanbul.
• International Mathematical Olympiad competitor prior to her Ph.D. studies.

Research Themes

Her research investigates the combinatorial structure of neural network loss surfaces, leveraging permutation symmetry to analyze critical manifolds. She examines how network width and overparameterization affect landscape smoothness, with applications to training failure modes in mildly overparameterized systems. Her work also encompasses generalization theory, random features, kernel methods, and out-of-distribution learning, the latter during an internship at Meta AI.