Veronica Centorrino

veronica.centorrino[at]unina.it

PhD in: MERC
Ciclo: XXXVI

Project title: Towards trustable neural networks bridging the gap between machine learning and control

Supervisor(s): Francesco Bullo and Giovanni Russo

  • Scarica il CV

  • I am currently a Ph.D. student in Modelling and Engineering Risk and Complexity (MERC) at the Scuola Superiore Meridionale.

    I have graduated cum Laude in Mathematics from the University of Catania, with a thesis entitled “Variational Equilibrium, Lagrangian Theory, and Applications to Network Models”.
    In 2020 I held an annual scholarship at the Italian National Institute for Geophysics and Volcanology (INGV) in Catania on the topic “Mathematical models for volcanic hazard monitoring and decision-making methods for risk mitigation and uncertainty quantification”.

    Towards trustable neural networks bridging the gap between machine learning and control

    Supervisors: Francesco Bullo and Giovanni Russo

    Neural networks and their applications are profoundly changing our society. While it is undeniable that neural networks are succeeding in learning many classification tasks, they are still black-box models of reality and, as such, besides often being hard to train, are prone to errors/biases which cannot be explained/quantified a-priori.
    Yet, it has been recently shown that neural networks implementations can be highly unstable, in the sense that even small perturbations to the input of a trained neural network can cause substantial changes in its output.

    It is therefore evident that new methodological breakthroughs are needed to rigorously design neural networks that are ready for real-world, physical applications.
    In this context, we want to investigate by leveraging tools from dynamical systems, control, and optimization if we can design networks that are bias-free and sufficiently expressive to represent reality. In particular, we aim to focus on the development of neural mechanisms making the network able to guarantee robustness and resilience against disturbances, adapt to the environment, analyzing also neural networks that can evolve over time.