Image classification is a well-studied problem in Machine Learning and can be tackled by training an ANN (Artificial Neural Network). The quality of an ANN depends on a collection of architectural choices that vary depending on the problem at hand. One of the most important choices is the design of an adequate loss function, as a tailored loss function can have a tremendous effect on the performance of the given model. The purpose of this work is to explore the various losses, propose some modifications, and compare them. The ANN used for implementing and testing these losses is a Siamese Neural Network (SNN) combined with a dissimilarity space and an SVM classifier.
L'image classification è un problema molto conosciuto e studiato in Machine Learning che può essere affrontato allenando una ANN (Artificial Neural Network). La qualità di una rete neurale artificiale dipende da diverse scelte architetturali che variano a seconda del problema da affrontare. Una delle scelte più importanti è quella della costruzione di una loss function adeguata, in quanto una loss function personalizzata può avere effetti significativi sulla performance di un dato modello. Lo scopo di questo lavoro è quello di esplorare varie loss functions, proporre delle modifiche e compararle. La ANN usata per l'implementazione e i test di queste funzioni è una rete neurale siamese, combinata con uno spazio di dissimilarità e un classificatore SVM.
Exploring Novel Loss Functions for Siamese Neural Network for Dissimilarity Image Classification
PETRUCCI, RICCARDO
2021/2022
Abstract
Image classification is a well-studied problem in Machine Learning and can be tackled by training an ANN (Artificial Neural Network). The quality of an ANN depends on a collection of architectural choices that vary depending on the problem at hand. One of the most important choices is the design of an adequate loss function, as a tailored loss function can have a tremendous effect on the performance of the given model. The purpose of this work is to explore the various losses, propose some modifications, and compare them. The ANN used for implementing and testing these losses is a Siamese Neural Network (SNN) combined with a dissimilarity space and an SVM classifier.File | Dimensione | Formato | |
---|---|---|---|
Petrucci_Riccardo.pdf
accesso riservato
Dimensione
4.89 MB
Formato
Adobe PDF
|
4.89 MB | Adobe PDF |
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/20.500.12608/34546