The research activity has mainly dealt with supervised Machine Learning algorithms, specifically within the context of kernel methods. A kernel function is a positive definite function mapping data from the original input space into a higher dimensional Hilbert space. Differently from classical linear methods, where problems are solved seeking for a linear function separating points in the input space, kernel methods all have in common the same basic focus: original input data is mapped onto a higher dimensional feature set where new coordinates are not computed, but only the inner product of input points. In this way, kernel methods make possible to deal with non-linearly separable set of data, making use of linear models in the feature space: all the Machine Learning methods using a linear function to determine the best fitting for a set of given data. Instead of employing one single kernel function, Multiple Kernel Learning algorithms tackle the problem of selecting kernel functions by using a combination of preset base kernels. Infinite Kernel Learning further extends such idea by exploiting a combination of possibly infinite base kernels. The research activity core idea is utilize a novel complex combination of kernel functions in already existing or modified supervised Machine Learning frameworks. Specifically, we considered two frameworks: Extreme Learning Machine, having the structure of classical feedforward Neural Networks but being characterized by hidden nodes variables randomly assigned at the beginning of the algorithm; Support Vector Machine, a class of linear algorithms based on the idea of separating data with a hyperplane having as wide a margin as possible. The first proposed model extends the classical Extreme Learning Machine formulation using a combination of possibly infinitely many base kernel, presenting a two-step algorithm. The second result uses a preexisting multi-task kernel function in a novel Support Vector Machine framework. Multi-task learning defines the Machine Learning problem of solving more than one task at the same time, with the main goal of taking into account the existing multi-task relationships. To be able to use the existing multi-task kernel function, we had to construct a new framework based on the classical Support Vector Machine one, taking care of every multi-task correlation factor.

Combining Kernel Functions in Supervised Learning Models

MARCELLI, ELISA
2021

Abstract

The research activity has mainly dealt with supervised Machine Learning algorithms, specifically within the context of kernel methods. A kernel function is a positive definite function mapping data from the original input space into a higher dimensional Hilbert space. Differently from classical linear methods, where problems are solved seeking for a linear function separating points in the input space, kernel methods all have in common the same basic focus: original input data is mapped onto a higher dimensional feature set where new coordinates are not computed, but only the inner product of input points. In this way, kernel methods make possible to deal with non-linearly separable set of data, making use of linear models in the feature space: all the Machine Learning methods using a linear function to determine the best fitting for a set of given data. Instead of employing one single kernel function, Multiple Kernel Learning algorithms tackle the problem of selecting kernel functions by using a combination of preset base kernels. Infinite Kernel Learning further extends such idea by exploiting a combination of possibly infinite base kernels. The research activity core idea is utilize a novel complex combination of kernel functions in already existing or modified supervised Machine Learning frameworks. Specifically, we considered two frameworks: Extreme Learning Machine, having the structure of classical feedforward Neural Networks but being characterized by hidden nodes variables randomly assigned at the beginning of the algorithm; Support Vector Machine, a class of linear algorithms based on the idea of separating data with a hyperplane having as wide a margin as possible. The first proposed model extends the classical Extreme Learning Machine formulation using a combination of possibly infinitely many base kernel, presenting a two-step algorithm. The second result uses a preexisting multi-task kernel function in a novel Support Vector Machine framework. Multi-task learning defines the Machine Learning problem of solving more than one task at the same time, with the main goal of taking into account the existing multi-task relationships. To be able to use the existing multi-task kernel function, we had to construct a new framework based on the classical Support Vector Machine one, taking care of every multi-task correlation factor.
23-giu-2021
Inglese
Inglese
DE LEONE, Renato
Università degli Studi di Camerino
File in questo prodotto:
File Dimensione Formato  
TESI_PhD_MARCELLI ELISA_ 23.06.pdf

accesso aperto

Dimensione 1.53 MB
Formato Adobe PDF
1.53 MB Adobe PDF Visualizza/Apri

I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.14242/122056
Il codice NBN di questa tesi è URN:NBN:IT:UNICAM-122056