Towards a topological–geometrical theory of group equivariant non-expansive operators for data analysis and machine learning

Abstract

The aim of this paper is to provide a general mathematical framework for group equivariance in the machine learning context. The framework builds on a synergy between persistent homology and the theory of group actions. We define group equivariant non-expansive operators (GENEOs), which are maps between function spaces associated with groups of transformations. We study the topological and metric properties of the space of GENEOs to evaluate their approximating power and set the basis for general strategies to initialise and compose operators. We begin by defining suitable pseudo-metrics for the function spaces, the equivariance groups, and the set of non-expansive operators. Basing on these pseudo-metrics, we prove that the space of GENEOs is compact and convex, under the assumption that the function spaces are compact and convex. These results provide fundamental guarantees in a machine learning perspective. We show examples on the MNIST and fashion-MNIST datasets. By considering isometry-equivariant non-expansive operators, we describe a simple strategy to select and sample operators, and show how the selected and sampled operators can be used to perform both classical metric learning and an effective initialisation of the kernels of a convolutional neural network.

Digital Twin: Values, Challenges and Enablers From a Modeling Perspective

“…Therefore, the concept of Physics/Knowledge/Science-informed ML might offer a powerful transformation to combine practice and theory (e.g., please see [356]- [360] for promoting theory-guided ML in various disciplines). For example, a symmetry enforcing approach to learn salient features from a sparse data set has been introduced using a knowledge injection concept [361], that might be a key enabler in developing digital twin technologies.…”

Learning Topology: Bridging Computational Topology and Machine Learning

“…Also, the advances and results presented in this paper shows that deepening the interactions between computational topology and machine learning is beneficial for both: first, topological layers improve and extend the functionalities, the performance of neural networks; on the other side, pushing for mapping diagrams in a reasonable input for a neural network has increased the effort both in the definition of a general framework able to encompass a large number of PH-based descriptors, and in an extensive analysis of the statistical properties of such descriptors. Remarkably, such a crossfertilization is leading both to promising insights into accountability, interpretability, and explainability of machine learning, and to innovative perspective in the field, as in [6].…”

“…Another interesting approach described by Carlsson and Gabrielsson in [13], where topology is used to clarify the roles played by the different layers in a very deep pre-trained convolutional network; starting from this example, the authors introduced a novel formalism in order to build networks with the set of neurons in each layer represented as a graphical model; hence, models produced this way are by their very nature more explainable than before. A groundbreaking change in how to investigate learning through topology is given by [6]. In this paper authors does not look at input data, or at activation maps, but introduce a mathematical general framework to manage the learning process looking at the space of observers (or lenses).…”

Why we should use topological data analysis in ageing: Towards defining the “topological shape of ageing”

“…Finally, TDA can be used in combination with other Big Data methods, should further processing be required. That is, the topological invariants can be further processed for example with machine learning and statistical methods (Bergomi et al, 2019).…”