Co-PI: Massimo Esposito (CNR)
AI research challenges have all one crucial aspect in common: the use of machine learning algorithms trained with real-world data (in-the-wild data) that are unstructured, noisy, often incomplete, limited in number, and partially inconsistent. The quest for constant improvement in the performance of these systems cannot be separated from the study of specific methodologies aimed at processing data in-the-wild, making AI performance resilient and robust in these challenging contexts.
The research activities that Spoke 3 aims at carrying out will include:
- the definition of adequate data augmentation techniques, when the data are incomplete or not adequately representative;
- making the algorithms resilient and robust against possible external attacks (also resulting from training with “malicious” data);
- the investigation of the implications relating to the design, validation and verification, evolution and operation of software that implements machine or deep learning algorithms, when it must work in-the-wild.
