About
Hi, I’m Freya!
I am a postdoc at Boston University with Aaron Mueller where I work on interpretability and machine learning theory. I am funded by a SNF Postdoc.Mobility fellowship.
I did my PhD at EPFL with Lenka Zdeborová in the Statistical Physics of Computation Lab, where I studied phase transitions in complex systems with methods from statistical mechanics.
Research
My interests are at the intersection of theoretical machine learning and interpretability. I work on building principled frameworks that characterize learning as a process that produces representations, structure, and mechanisms, and exhibits sharp transitions between different regimes of behavior. To do so, I often take phenomena from large networks, and explain them in minimal (theoretical) toy models. This makes them more understandable, and improves our ability to characterize and predict the behavior of learning systems from first principles.
More broadly, I see neural networks of one example of a complex system in which structured global computation emerges from many small local interactions. Other cool systems I studied are cellular automata and opinion dynamics.
Publications
- . Knowledge Distillation for Random Data: Soft Labels and Similarity Scores May Contain Memorized Information. ICLR, 2026. [link]
- . Counting in Small Transformers: The Delicate Interplay between Attention and Feed-Forward Layers. ICML, 2025. [link]
- . A phase transition between positional and semantic learning in a solvable model of dot-product attention. NeurIPS, 2024. [link]
- . Dynamical phase transitions in graph cellular automata. Physical Review E, 2024. [link]
- . Quenches in the Sherrington–Kirkpatrick model. Journal of Statistical Mechanics, 2024. [link]
- . (Dis)assortative partitions on random regular graphs. Journal of Physics A, 2022. [link]