Generalization covers an important topic on neural networks training, that represents its ability to abstract significant information for classification, giving a training set ; hence to the capability to classify properly unseen data, that is the set of patterns with which the network has not been trained. The development of these technologies plays a crucial role, as it is useful to reach the state of the art models. This is inspired by the MaxUp technique, which offers an improvement of the generalization performance through the use of Augmented Data and Adversarial Training. Several tests including various datasets were carried out, especially on pictures and histopathology recognition. The aim of this work is to demonstrate how Data Augmentation and Adversaria Training could increase performances on generalization and improve robustness on methods.
La generalizzazione ricopre un ruolo importante inerente nell’apprendimento delle reti neurali. Essa rappresenta la capacità di una rete neurale di astrarre informazioni importanti per la classificazione, dato un training set; pertanto, corrispondente anche alla capacità di classificare correttamente gli unseen data, ovvero l’insieme dei pattern con i quali la rete non è stata allenata. Lo sviluppo di queste tecnologie, quindi, gioca un ruolo fondamentale, perché utile ad ottenere risultati migliori in termini di apprendimento dei machine learning model. Questo lavoro prende enorme spunto dalla tecnica del MaxUp, la quale offre un miglioramento delle prestazioni di generalizzazione tramite l’utilizzo di Augmented Data ed adversarial training. Sono stati effettuati vari test comprendenti vari dataset, in particolare il riconoscimento di immagini e istopatologie. Lo scopo di questo lavoro è quello di dimostrare come la Data Augumentation e l’adversarial training possano migliorare le performance nella generalizzazione e nell’accurancy, offrendo pure robustezza al metodo.
sviluppo modelli di adversarial training
RUSSO, MICHELE
2021/2022
Abstract
Generalization covers an important topic on neural networks training, that represents its ability to abstract significant information for classification, giving a training set ; hence to the capability to classify properly unseen data, that is the set of patterns with which the network has not been trained. The development of these technologies plays a crucial role, as it is useful to reach the state of the art models. This is inspired by the MaxUp technique, which offers an improvement of the generalization performance through the use of Augmented Data and Adversarial Training. Several tests including various datasets were carried out, especially on pictures and histopathology recognition. The aim of this work is to demonstrate how Data Augmentation and Adversaria Training could increase performances on generalization and improve robustness on methods.File | Dimensione | Formato | |
---|---|---|---|
Russo_Michele.pdf
Open Access dal 24/03/2024
Dimensione
1.16 MB
Formato
Adobe PDF
|
1.16 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/20.500.12608/34553