Alonso Urbano García

I am a PhD candidate in Mathematics at the IOL Learn lab at the Zuse Institute Berlin and TU Berlin, where I work in efficient deep learning.

Before starting my PhD, I worked as a Consultant at Deloitte (Madrid) and as a Machine Learning Engineer at GrwNxt (Amsterdam). I hold an MSc in Artificial Intelligence from Vrije Universiteit Amsterdam and a BSc in Mathematics from Universidad de Málaga.

Broadly, I am interested in how baking structure (e.g. hierarchy or symmetry) into deep learning architectures can improve quality and efficiency. Lately, that question has taken me to linear attention sequence models: structuring their recurrent memory can buy better recall at lower compute. The same spirit runs through my earlier work; for example, LH-NeF represents arbitrary continuous signals as geometry-adaptive token groups, where increased efficiency and quality come from introducing hierarchy and locality biases into the tokenization and field decoding.

Publications

Neural Field Tokenizations with Hierarchy and Spatial Locality Priors
A. Urbano, D. W. Romero, M. Zimmer, S. Pokutta
Preprint, 2026. [arXiv]

RECON: Robust symmetry discovery via Explicit Canonical Orientation Normalization
A. Urbano, D. W. Romero, M. Zimmer, S. Pokutta
The Fourteenth International Conference on Learning Representations (ICLR) 2026. [arXiv]

Self-supervised detection of perfect and partial input-dependent symmetries
A. Urbano and D. W. Romero
Geometry-grounded Representation Learning and Generative Modeling Workshop (GRaM) at ICML 2024, PMLR 251:113-131. [Paper]