[00851] Mathematics for Big Data and Artificial Intelligence: models and challenges
Session Time & Room : 4C (Aug.24, 13:20-15:00) @E504
Type : Proposal of Minisymposium
Abstract : The availability of huge amounts of data and the application of AI techniques are considered as the fourth industrial revolution, but extracting meaningful knowledge and transparent decisions from the data is not a trivial task.
Mathematics is the ‘language’ on which are based the existing algorithms for data processing and for AI. This minisymposium is organized within the ECMI SIG “Mathematics for Big Data and AI”. The talks in this minisymposium will present and discuss how Mathematics can play a leading role in improving the reliability, computational efficiency, and transparency of the existing techniques for big data analysis and AI.
[01685] Equivariant non-expansive operators as a bridge between TDA and geometric deep learning
Format : Talk at Waseda University
Author(s) :
Patrizio Frosini (University of Bologna)
Abstract : Group equivariant non-expansive operators $(\mathrm{GENEOs})$ have been recently introduced as mathematical tools for approximating data observers when data are represented by real-valued or vector-valued functions $(\mathtt{{https://rdcu.be/bP6HV}})$. The use of these operators is based on the assumption that data interpretation depends on the observers' geometric properties. In this talk we will illustrate some recent results, showing how GENEOs can make available an interesting link between topological data analysis and geometric deep learning.
[01623] a new machine learning paradigm for protein pocket detection based on Group Equivariant Non Expansive Operators
Format : Talk at Waseda University
Author(s) :
Alessandra Micheletti (Università degli Studi di Milano)
Giovanni Bocchi (Università degli Studi di Milano)
Patrizio Frosini (Università degli Studi di Bologna)
Carmine Talarico (Dompè Farmaceutici)
Filippo Lunghini (Dompè Farmaceutici)
Andrea Beccari (Dompè Farmaceutici)
Carmen Gratteri (University Magna Grecia Catanzaro)
Alessandro Pedretti (Università degli Studi di Milano)
Abstract : Protein pockets detection is a key problem in the context of drug development, since the ability to identify a small number of potential binding sites, allows to speed up drug discovery procedures. In this talk we will show how Group Equivariant Non Expansive Operators (GENEOs) can be used to build a geometrical machine learning method, able to detect protein pockets better than ML techniques already in use, but being based only on 17 unknown parameters.
[01528] SLiSeS: Subsampled Line Search Spectral Gradient Method for Finite Sums
Format : Online Talk on Zoom
Author(s) :
Stefania Bellavia (University of Florence)
Natasa Krejic (University of Novi Sad)
Natasa Krklec Jerinkic (University of Novi Sad)
Marcos Alejandro Raydan (NOVA University Lisbon)
Abstract : In this paper, we aim to exploit advantages of spectral method in stochastic optimization framework, especially in mini-batch subsampling case which is often used in Big Data setup. In order to let the spectral coefficient explore the spectrum of the approximate Hessian, we keep the same sample for several iterations before we subsample again. We analyze conditions for almost sure convergence and present initial numerical results that show the advantages of the proposed method.
[01698] Interpretable models for large-scale tabular datasets
Format : Online Talk on Zoom
Author(s) :
Claudia Soares (NOVA School of Science and Technology)
Abstract :
The purpose of this work is two-fold: on the one hand, to demonstrate that machine learning models can be considered a powerful alternative to predicting real-world variables in high-stakes scenarios, and, on the other hand, to propose a new method that is empirically the state-of-the-art rule-based method for large datasets. We accompany our method with tailored algorithms for fast learning in large datasets.
