Geometric Intelligence Workshop
Instituto de Matemáticas UNAM
Mexico City, March 31 - April 4 2025
Instituto de Matemáticas UNAM
Mexico City, March 31 - April 4 2025
¡Los invitamos a explorar el papel de la Geometría en la Inteligencia Artificial!
Descubran cómo la Geometría y la Topología están transformando la Inteligencia Artificial.
Participen en un taller diseñado para profundizar en estos conceptos y sus aplicaciones en problemas actuales.
Charlas y tutoriales impartidos por especialistas en el área.
Hackatón para aplicar conocimientos en un reto práctico.
Análisis de datos reales con aplicaciones en neurociencia.
We invite you to explore the role of Geometry in Artificial Intelligence!
Discover how Geometry and Topology are transforming Artificial Intelligence.
Participate in a workshop designed to dive deep into these concepts and their applications in current problems.
Talks and tutorials given by specialists in these areas.
Hackathon to apply your knowledge in a practical challenge.
Analysis of real data with applications in neuroscience.
Ponentes / Speakers
Omar Antolín Camarena, IM UNAM
Yalbi I. Balderas-Martínez, INER
Guillermo Bernárdez Gil, UC Santa Barbara
Louisa Cornelis, UC Santa Barbara
Miguel A. Evangelista, U. Rosario Castellanos
Rodrigo Fritz, IM UNAM
Sergei Gukov, Caltech
Sarah Kushner, UC Santa Barbara
Anayanzi Martínez, FC UNAM
Miguel A. Maurin, IM UNAM
Víctor Mijangos, FC UNAM
Nina Miolane, UC Santa Barbara
Adele Myers, UC Santa Barbara
Mathilde Papillon, UC Santa Barbara
Luís Pereira, UC Santa Barbara
Haydeé Peruyero, CCM UNAM
Sergio Rajsbaum, IM, UNAM
Nelly Sélem Mojica, CCM UNAM
Eduardo Velazquez Richards, IIMAS UNAM
Programa - Schedule
31.3
9:30 Guillermo Bernárdez - Topological Deep Learning Tutorial 1: Train Your Own Topological Deep Learning Model
(Important : please bring your laptop!) Graph Neural Networks excel in learning from relational datasets, processing node and edge features in a way that preserves the symmetries of the graph domain. However, many complex systems--such as biological or social networks--involve multiway complex interactions that are more naturally represented by higher-order topological spaces.
The emerging field of Topological Deep Learning (TDL) aims to accommodate and leverage these higher-order structures. In this three-part workshop, we will 1) present a self-contained introduction to the motivation and theory behind TDL, 2) introduce TopoBench, an open source Python library for easily developing and training TDL models, and 3) provide participants with hands-on experience in running state-of-the-art TDL models. Specifically, participants will use TopoBench to build, train, and test graph-based TDL models as introduced in our recent work, TopoTune.
10:30 Mathilde Papillon - Topological Deep Learning Tutorial 2: Train Your Own Topological Deep Learning Model
11:30 ☕
12:00 Nelly Sélem Mojica - Topological Data Analysis for Pangenomics: Insights into Horizontal Gene Transfer in Streptococcus agalactiae and Viral Genomes
Topological Data Analysis (TDA) offers a powerful framework for studying genomic evolution. In this work, we present a lesson on applying TDA to pangenomics, focusing on the resistome of Streptococcus agalactiae. A key theoretical result states that purely vertical inheritance does not produce 1-holes in persistence barcodes. We extend this analysis to cases of non-vertical inheritance, demonstrating how horizontal gene transfer leaves topological signatures in the genomes of Influenza virus and SARS-CoV-2. Our findings illustrate how TDA can distinguish between evolutionary processes, offering a novel perspective on genome evolution and resistance gene mobility.
13:00 lunch
15:00 Sergio Rajsbaum - An introduction to distributed computing and the combinatorial topology approach
Modern computer systems are becoming more and more concurrent. Nearly every activity in our societydepends on the Internet, where distributed databases communicate with one another and with humanbeings. When a customer asks to withdraw money from an automatic teller machine, the banking system must either both provide the money and debit that account or do neither, all in the presence of failures and unpredictablecommunication delays. Concurrency is not limited to wide-area networks. As transistor sizes shrink, manufacturers have focused on making processors more parallel.
Shared-memory multiprocessors are systems that concurrently execute multiple threads of computation which communicate and synchronize through data structures in shared memory. The efficiency of these concurrent data structures is crucial to performance, yet they are far more difficult to design than sequential ones because threads executing concurrently may interleave their steps in many ways, each with a different and potentially unexpected outcome.
This talk presents an introduction to the use of combinatorial topology in helping to provide a theoretical foundations of concurrency. It illustrates how simplicial complexes can help all the way from specifying distributed tasks and data structures, to designing fault-tolerant, efficient algorithms, and also to bound the limitations of the possible solutions.
16:00 Break
16:15 Haydee Peruyero - How Topology Reveals Patterns in Biological Data
Topology allows us to go beyond visible shapes to uncover hidden structures in data. Through Topological Data Analysis (TDA), we can identify patterns and relationships in complex biological systems, providing a new perspective on their organization. This tool has proven useful in studying horizontal gene transfer and the evolution of microbial communities. In this talk, we will explore applications related to non-vertical gene inheritance and the classification of multicellular patterns.
This is a joint work with S. Guerrero-Flores, N. Sélem-Mojica, L.R. Figuero-Martínes and E. Azpeitia.
9:30 Adele Myers - Geomstats Tutorial 1: Learning from Shapes Using Geomstats
Shapes are everywhere in nature. Therefore, quantifying differences between shapes is relevant for nearly every field. For simple shapes, this task is easy. For example, one can compare two circles by computing the difference between their radii. This question becomes much more difficult when the shape in question is more complicated, like a cell outline or brain surface. Here, we delve into the field of shape analysis and shape spaces, starting from the basic question of "what is a shape" and ending with a taste of Geomstats: a software package that makes computations in shape spaces accessible for people in all fields.
10:30 Luís Pereira - Geomstats Tutorial 2: Geomstats: a Python package for Riemannian Geometry and Geometric Statistics
Geomstats is an open-source Python package for computations, and statistics on nonlinear manifolds. We provide object-oriented and extensively unit-tested implementations. Manifolds can be equipped with Riemannian metrics with associated exponential and logarithmic maps, geodesics, and parallel transport. Some manifolds can also be endowed with additional mathematical structures, such as Lie group, or fiber bundle structures. Statistics and learning algorithms provide methods for estimation, clustering, and dimensionality reduction on manifolds. All associated operations are supported in different backends, namely NumPy, Autograd, and PyTorch. In this talk, we briefly introduce the main concepts in Riemannian geometry and discuss the package design. We show that Geomstats provides reliable building blocks to both foster research in differential geometry and statistics and democratize the use of Riemannian geometry in statistics and machine learning.
The source code is freely available under the MIT license at https://github.com/geomstats/geomstats.
11:30 ☕
12:00 Nina Miolane - Geometric Intelligence in Minds and Machines (Coloquio IMATE)
Neural representations in both biological and artificial neural networks reveal rich and intricate geometric structures. From the toroidal neural manifolds that shape navigation in animals and recurrent neural networks to the complex topological interactions among neurons in brains and machines, intelligence emerges within a landscape of geometries we are only beginning to explore. Even the higher layers of neural networks in computer vision—and the visual cortex itself—exhibit striking symmetries that challenge our understanding of learning and perception.
In this talk, we delve into the mathematical foundations of neural representations and showcase ongoing research at the Geometric Intelligence Lab. By bridging geometry, neuroscience, and machine learning, we aim to expand the frontiers of how we understand and model intelligence—both artificial and biological—advancing fundamental science and transforming brain research and care.
13:00 lunch
15:00 Yalbi Balderas Martínez - Building a Multiscale Computational Framework to Address the Complexity of Respiratory Diseases
In this talk, we will discuss the process of developing a multiscale computational approach to study respiratory diseases by combining molecular, cellular, and tissue-level data. At the molecular level, we can use AlphaFold to explore protein structures—particularly transcription factors and membrane proteins—to better understand mechanisms involved in antibody response and disease progression. At the cellular level, single-cell transcriptomics allows us to investigate aging, cellular diversity, and interactions within lung organoids. Complementing this, our imaging team is working to generate high-resolution tomographic datasets that will support predictive modeling for early detection and patient care. Integrating such diverse data presents several challenges—from handling high-dimensional information and aligning outputs across methods, to ensuring that mathematical tools are meaningfully applied in a biomedical context.
We will reflect on these ongoing challenges and how collaboration across disciplines is central to making this integration possible. Ultimately, our goal is to show how these approaches can contribute to better understanding disease complexity and supporting more accurate, personalized strategies in public health and clinical care.
16:00 Break
16:15 Louisa Cornelis - Graphing the Proteome: A GNN Approach to Predict Disease Severity in Frontotemporal Dementia
Recent advances in large-scale proteomics enable capture of thousands of proteins in human biospecimen samples, supporting discovery of disease-related biomarkers and drug targets. Traditional proteomic workflows model proteins as individual predictors of disease outcomes, failing to account for how relationships between proteins influence predictive performance. Here we present a novel analytical framework applying a biologically-inspired graph attention neural networks to cerebrospinal fluid proteomic data for predicting disease severity in 252 participants with frontotemporal dementia, a progressive neurodegenerative disease that lacks robust prognostic biomarkers. Protein data is formatted into disease specific co-expression graphs using weighted correlation network analysis. Graph attention layers are utilized to extract hierarchical protein module-level graph features, which are then used alongside patient level demographic data for downstream predictions. The framework successfully forecasts cognitive decline and gold-standard markers of neurodegeneration, surpassing traditional analytical methods including differential regression and LASSO regression. Through systematic network interpretation using feature attribution techniques, we identify proteins with high importance for disease severity prediction that are not detected by other approaches, including targets linked to neuroimmune signaling and sugar metabolism.
Our study demonstrates how graph-based deep learning approaches can extract meaningful insights from high-dimensional proteomic data, advancing our understanding of the molecular underpinnings of frontotemporal dementia with potential applications for all neurodegenerative diseases.
16:40 Sarah Kushner - Exploring volume changes in the brain during pregnancy
There is an extreme lack of medical research on women. Most studies include male participants, leading to a deficit of important healthcare information for half of world’s the population. Pregnancy is an experience many women go through in their lifetime, yet there is little known about the effects of pregnancy on a woman’s brain. In our work, we aim to explore the changes in the brain on the first MRI dataset of its kind. We are trying to apply statistics on manifolds to predict whether an expecting mother will have post partum depression and design a digital twin of the brain over the course of a pregnancy.
9:30 Anayanzi Martinez - Knots and Grids
A knot can be represented as a closed path within a square grid. In this talk, I will explain various computational and combinatorial aspects of this representation and discuss its significance in the construction of databases that are essential for research in knot theory and applications in machine learning.
10:00 Rodrigo Fritz - Singular random Lissajous knots as souls of controlled genus surfaces.
Geometric and Topological Deep Learning have explored a wide range of structured data domains, from sets and graphs to manifolds, including simplicial and combinatorial complexes, while leveraging diverse mathematical techniques for representation and analysis. As part of a larger project in collaboration with Pablo Suárez-Serrato, Eduardo Velázquez Richards, Víctor Mijangos, and Anayanzi Martínez, I present a framework for constructing a controlled genus-diverse database of surfaces, derived from singular Lissajous knots as a generative model. By carefully tuning the frequency and phase parameters of Lissajous representations, we developed a systematic approach for generating surfaces with prescribed topological genus.
The resulting dataset provides a structured and well-distributed resource for training neural networks, enabling models to learn how to generalize representations of surfaces with varying genera. This approach advances deep learning models with broad potential applications.
10:30 Miguel Ángel Evangelista - Symbolic regression of Hamiltonian vector fields
We approach solving to the Hamiltonization problem in two dimensions using images and neural networks. After a brief introduction and a mini-tutorial on the Python module PoissonGeometry, we show how to build a synthetic dataset of Hamiltonian vector fields from polynomial and trigonometric functions. Finally, we present a hybrid CNN–LSTM deep learning model that predicts a symbolic expression for the Hamiltonian function generating a vector field.
This is joint work with Pablo Suárez-Serrato.
11:00 Miguel Ángel Maurin - Space-Filling Curves and Positional Encodings in Transformers: A Spherical Example
Transformers excel at processing sequential data, but when applied to geometric domains like manifolds, a problem arises: how should we order our inputs? Unlike text, manifolds lack a predefined traversal order. In this talk, we present an approach based on space-filling curves to impose a structured ordering on spherical data. By constructing such a curve over the sphere, we provide a canonical way to linearize manifold data, this enables positional encodings that are both geometrically meaningful and effective for Transformer architectures. Our method enhances the ability of deep learning models to process spherical data opening new avenues for learning on geometric domains.
This is joint work with Pablo Suárez-Serrato.
11:30 ☕
12:00 Omar Antolín - Beyond Local Geometry: Vector Bundles, Topology, and the Quantum World
Imagine a mathematical object that looks locally like a simple product, but globally twists and turns in unexpected ways. These objects, called vector bundles, are essential for understanding phenomena like the Integer Quantum Hall Effect, a physical system with quantized conductivity. This talk provides an accessible introduction to vector bundles and their topological invariants, specifically Chern classes, and how they explain quantized phenomena in physics. To unlock new capabilities for modeling systems with non-trivial global structure, this talk advocates for improved support for vector bundles in computer geometry and topology packages. This enhancement could significantly impact areas such as materials design, where subtle topological features play a crucial role.
13:00 Tarde de discusión libre - Free discussion afternoon
9:30 Eduardo Velazquez Richards - Contour parametrization via anisotropic mean curvature flows
The evolution of curves by curvature-dependent flows has many applications, such as crystal growth modeling, interface dynamics, and fire propagation. In this talk, we present a parallelizable approach that couples a Poisson problem and a dynamic equation on the curve, along with the conditions ensuring an equivalence to the mean curvature flow. We also present a numerical implementation applied to the problem of contour parametrization of shapes in high-contrast pictures.
This is joint work with Pablo Suárez-Serrato
10:00 Víctor Mijangos - An adjacency-preserving sampling method for 3d surfaces processing with graph neural networks
This talk introduces a sampling method for 3D surfaces based on the EuLearn database, designed to preserve adjacency information. Maintaining adjacency is a crucial requirement for classifying topological invariants of surfaces. The proposed method samples points in an order induced by the spanning tree of the surface triangulation and subsequently generates a sphere for radial point sampling. Using this sampled data, various graph neural networks are applied to classify the genus of each surface. The results demonstrate that this sampling method effectively reduces the complexity of classification neural architectures while preserving essential topological information.
This is joint work with Rodrigo Fritz, Pablo Suárez-Serrato, Eduardo Velázquez Richards, Víctor Mijangos, and Anayanzi Martínez,
10: 30 Sergei Gukov - Math + AI = AGI
It comes as no surprise that solving challenging research-level math problems drives progress in mathematics. What is more surprising, though, is that solving such long-standing open problems also contributes to an entirely different field: the development of the next generation AI systems.
We live in an exciting time where mathematics and AI can greatly benefit each other, and the goal of the talk is to explain how and why, drawing on specific examples from topology and combinatorial group theory.
11:30 ☕
12:00 Inicia Hackathón
9:30 Hackatón
11:30 Resultados del Hackathón / Hackathon results
12:00 Fin
Hackathón para estudiantes: Examina cómo las mediciones basadas en imágenes del cerebro femenino concuerdan con el diagnóstico de enfermedades neurológicas o neuropsiquiátricas. Participa con nostros en equipo en la edición 2025 del datathon Women in Data Science, utilizando las técnicas de aprendizaje profundo geométrico y topológico que veremos en el taller.
Hackathon for students: Examine how image-based measurements of the female brain correlate with the diagnosis of neurological or neuropsychiatric diseases. Participate with us as a team in the 2025 edition of the Women in Data Science datathon, using the geometric and topological deep learning techniques that we will cover in the workshop.
Lugar / Where: Auditorio Nápoles Gándara, Instituto de Matemáticas UNAM, Ciudad Universitaria, Coyoacán CDMX
Fechas / When : 31.3.2025 - 4.4.2025
Info: GIW2025@im.unam.mx