Haitz Sáez de Ocáriz Borde

🚨 "Mathematical Foundations of Geometric Deep Learning", co-authored with @mmbronstein 📚 Read the paper here: arxiv.org/abs/2508.02723 🧠 We review the mathematical background necessary for studying Geometric Deep Learning. #GDL #mathematics #deeplearning #AI

A handcrafted synergistic example:

✨ Key insight: Multi-head attention can enhance information propagation beyond parallel computation by creating synergistic pathways across diverse feedforward DAGs, reducing mixing time and amplifying minimax fidelity.

🚀 New paper “Beyond Parallelism: Synergistic Computational Graph Effects in Multi-Head Attention” has been accepted to the Proceedings Track of the #NeurIPS 2025 Workshop on Symmetry and Geometry in Neural Representations!

🎉 Thanks to all collaborators Carlo Saccardi, Maximilian Pierzyna, Simone Monaco, Cristian Meo, Pietro Lio, Rudolf Saathof, Geethu Joseph & Justin Dauwels, and especially huge congrats to Carlo!

In this work, we benchmark state-of-the-art generative models for climate downscaling and introduce a Power Spectral Density (PSD) loss that helps models recover fine-scale spatial detail and improve physical realism.

New paper “Assessing the Geographic Generalization & Physical Consistency of Generative Models for Climate Downscaling” has been accepted at #NeurIPS 2025 AI4Science! 📄 Paper: arxiv.org/abs/2510.13722 💻 Code: github.com/CarloSaccardi/…

Group theory originated with Galois, who introduced permutation groups to prove that general quintic polynomials cannot be solved by radicals. It later found deep applications in physics, and more recently, in Geometric Deep Learning. arxiv.org/pdf/2508.02723 w/ @mmbronstein

We combine sequence modeling + graph representation learning for ICU length of stay prediction: modeling each patient’s time-series while using inter-patient correlation graphs as a geometric prior.

'Bridging Graph and State-Space Modeling for Intensive Care Unit Length of Stay Prediction' w/ Shuqi Zi, @09Emmar and Pietro Liò, accepted at #NeurIPS2025 (GenAI for Health Workshop)! Paper: arxiv.org/abs/2508.17554 Code: github.com/ShuqiZi1/S2G-N… #AI4Healthcare #GNN #Mamba #ICU

Dataset (PyG) 👉 pytorch-geometric.readthedocs.io/en/latest/gene… Measurement of long-range influence (PyG) 👉 #torch_geometric.utils.total_influence" target="_blank" rel="nofollow noopener">pytorch-geometric.readthedocs.io/en/latest/modu…

🔍 Alongside the dataset, we propose a model-agnostic measurement of long-range influence (now available in PyTorch Geometric v2.7.0) that lets you quantify how many hops truly affect a GNN’s prediction.

“Towards Quantifying Long-Range Interactions in Graph Machine Learning: a Large Graph Dataset and a Measurement” w/ Huidong, Baskaran, @mmbronstein , and @epomqo has been accepted at the New Perspectives in Graph Machine Learning Workshop @NeurIPSConf 🎉 arxiv.org/pdf/2503.09008

The manuscript contains numerous historical annotations and margin notes that contextualize the mathematical concepts under discussion, while also highlighting direct connections to modern GDL architectures.

Our paper "Mathematical Foundations of Geometric Deep Learning" (co-authored with @mmbronstein ) has now been officially added to the Geometric Deep Learning Book website as the recommended preparation material! (special thanks to @PetarV_93) geometricdeeplearning.com/book/

With CLIP rewards, the model can learn to generate images beyond the pre-training distribution.

Original distribution versus using an Aesthetic Reward

🚀 “Fine-Tuning Next-Scale Visual Autoregressive Models with GRPO” w/ @MatteoGallici We adapt DeepSeek-R1’s RL techniques to image models using aesthetic + CLIP rewards. 📄 openreview.net/pdf?id=lZK2svo… #GRPO #RL #VisualModels #AutoregressiveModels #DeepSeek #CLIP #ICML