An increasing number of applications take advantage of Convolutional Neural Networks due to their effectiveness in image vision tasks. However, these models can be affected by critical faults that change their prediction. Previous works have studied what happens when critical faults impact a LeNet5 network. In particular, they have proposed an architecture for a fault detector composed of two networks, where the auxiliary model acts as a controller for the main network. This Master Thesis aims to expand the same analysis to a larger network such as VGG-16. Moreover, we will propose an effective and inexpensive fault detector that analyzes the vector score of VGG-16 to detect critical faults, without the need for an auxiliary network.
Un numero crescente di applicazioni impiega Reti Neurali Convoluzionali grazie alla loro efficacia in ambito di visione artificiale. Tuttavia, queste reti possono essere affette da guasti critici che modificano la loro previsione. Lavori precedenti si sono concentrati nello studio di cosa succede quando dei guasti critici colpiscono una rete LeNet5. In particolare, hanno proposto un’architettura per un rilevatore di guasti composto da due reti, dove la rete ausiliaria funge da controllore per la rete principale. Questa tesi magistrale cerca di espandere la stessa analisi ad una rete più grossa: VGG-16. Inoltre, andremo a proporre un rilevatore di guasti efficace ed economico che analizza unicamente il ‘vector score’ di VGG-16 per individuare guasti critici, senza utilizzare una rete ausiliaria.
Effective and inexpensive fault detection in VGG-16 inference
GAVARINI, GABRIELE
2020/2021
Abstract
An increasing number of applications take advantage of Convolutional Neural Networks due to their effectiveness in image vision tasks. However, these models can be affected by critical faults that change their prediction. Previous works have studied what happens when critical faults impact a LeNet5 network. In particular, they have proposed an architecture for a fault detector composed of two networks, where the auxiliary model acts as a controller for the main network. This Master Thesis aims to expand the same analysis to a larger network such as VGG-16. Moreover, we will propose an effective and inexpensive fault detector that analyzes the vector score of VGG-16 to detect critical faults, without the need for an auxiliary network.File | Dimensione | Formato | |
---|---|---|---|
Tesi - Gavarini - 08-07-2021.pdf
accessibile in internet per tutti
Dimensione
4 MB
Formato
Adobe PDF
|
4 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/177908