On the geometry of learned representations

I am a final-year PhD student at the University of Basel, supervised by Alexander Navarini and Marc Pouly. I also spent time as a visiting researcher at EPFL with Maria Brbić.

My research is about the geometry of learned representations, the structure inside a neural network’s embedding space that every downstream decision (similarity, outlier detection, cross-modal alignment) ultimately reads from.

The Platonic Representation Hypothesis claims that as neural networks scale, their representations converge to a single shared geometry. We revisit this claim [1] and find that the reported convergence largely tracks model size rather than reflecting real alignment, and that once we correct for the confound, what remains is agreement on local neighborhood structure rather than global shape. We call this narrower picture the Aristotelian Representation Hypothesis, and it pins down which signals from a representation downstream tools can rely on across encoder choices.

A learned representation also carries traces of the data it was trained on, beyond what the labels reveal. Within a single dataset, isolated samples are usually off-topic, tight pairs near-duplicates, and label outliers mislabeled, which turns the representation space itself into a label-free auditor for image [3, 4] and audio [5] collections. Across two modalities, preserving each side’s local structure regularizes alignment when paired data is scarce, taking the requirement from millions to tens of thousands [2]. Across collections of an entire scientific field, harmonizing many representations into a single latent space lets us audit not one dataset but a whole discipline, as we do for digital dermatology, where 1.1 million images across 29 public archives yield the first quantitative atlas of the field and expose what those archives do and do not cover [6].

Outside research, I spend my time with my family and on the bike.