TopModel


Multi-parameter Topology for Machine Learning Models

General Presentation

TopModel is an ANR JCJC grant of four years (2023-2027) that aims at studying the incorporation of multi-parameter topology to machine learning models for both single-cell and cosmological data.

Abstract

The central tenet of this project is the use of multiparameter topological data analysis for machine learning models, for both regularizing and monitoring these models, and for the automatic generation of new features and descriptors to feed these models with. On the theoretical front, a lot of efforts will be devoted to the development, implementation and generalization of standard topological data analysis techniques, who (for the most part) can only study the topological variations of at most one parameter (such as the data scale), so as to make them suitable for the study of the topological variations of several parameters jointly (such as density and scale, marker genes). Then, the focus will be on specific applications, for which topological data analysis is known to be relevant and efficient, of these new multiparameter topological data analysis methods for machine learning models. More precisely, we will emphasize the usefulness of our new tools on data sets from cosmology (large scale structures of the Universe) and biology (single-cell sequencing, mass cytometry).

WP1: Multi-Parameter Persistent Homology

  • New Invariants for Multi-parameter Persistence Modules
  • Integration in Data Science and Machine Learning Pipelines

WP2: Multi-Parameter Mapper Complexes

  • New Stratification of Mapper Complexes
  • Application to Statistical Inference and Optimization

WP3: Generative Models for Cosmology

  • New Topology-Guided Generative Models
  • Application to the Fast Generation of Galaxy Maps

WP4: Geometric Inference for Spatial Transcriptomics

  • New Topology-Guidede Signatures for Spatial Transcriptomics Data
  • Application to Biological Inference