In today's society, the use of Machine Learning technologies is increasingly supporting human activities, from research contexts to work, up to aspects of daily life. If from one hand this phenomenon has allowed the resolution of individual problems at a speed never seen before, on the other hand, it is necessary to consider that in several fields of application the reliability of these tools is still far from safe use. The introduction of learning techniques that are less and less subject to human intervention, such as the review and data labeling, also focuses attention on the implementation of these new algorithms, to improve their general performance. In this thesis work, we intend to deepen the aspects introduced with the implementation of an algorithm of Deep Learning based on the Self-Supervised technique, to perform an analysis of the crucial components in model learning and reinforcement process. This algorithm is designed with an Adversarial Training module, to evaluate the model and to investigate its costs and benefits in terms of robustness.
Nella società odierna l'impiego delle tecnologie di Apprendimento Automatico affiancano sempre di più le attività di impiego umane, dai contesti di ricerca all'utilizzo lavorativo, fino ad aspetti di vita quotidiana. Se da una parte questo fenomeno ha permesso la risoluzione di singoli problemi ad una velocità mai vista prima, dall'altra è doveroso considerare che in diversi campi di applicazione l'affidabilità di questi strumenti è ancora lontana al fine di un utilizzo in sicurezza. L'introduzione di tecniche di apprendimento sempre meno soggette ad un intervento umano, come la revisione e l'etichettamento dei dati, focalizza inoltre l'attenzione sul funzionamento dei nuovi algoritmi, al fine di migliorarne le prestazioni generali.
Analisi delle Robustezza di Modelli Self-Supervised
SERGIO, FEDERICO
2021/2022
Abstract
In today's society, the use of Machine Learning technologies is increasingly supporting human activities, from research contexts to work, up to aspects of daily life. If from one hand this phenomenon has allowed the resolution of individual problems at a speed never seen before, on the other hand, it is necessary to consider that in several fields of application the reliability of these tools is still far from safe use. The introduction of learning techniques that are less and less subject to human intervention, such as the review and data labeling, also focuses attention on the implementation of these new algorithms, to improve their general performance. In this thesis work, we intend to deepen the aspects introduced with the implementation of an algorithm of Deep Learning based on the Self-Supervised technique, to perform an analysis of the crucial components in model learning and reinforcement process. This algorithm is designed with an Adversarial Training module, to evaluate the model and to investigate its costs and benefits in terms of robustness.File | Dimensione | Formato | |
---|---|---|---|
Sergio_Federico.pdf
accesso aperto
Dimensione
8.14 MB
Formato
Adobe PDF
|
8.14 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/10067