Towards layer-wise training of deep Kernel networks

n recent times, there has been a growing interest in integrating kernel methods with neural networks, capitalizing on the expressive capabilities of the former and the generalization strengths of the latter. On of the aims of the study was contribute to the understanding and development of the kchai...

Full description

Bibliographic Details
Author: Schoolkate, Pim
Format: master thesis
Publication Date:2023
Country:España
Institution:Universitat Politècnica de Catalunya (UPC)
Repository:UPCommons. Portal del coneixement obert de la UPC
Language:English
OAI Identifier:oai:upcommons.upc.edu:2117/403814
Online Access:https://hdl.handle.net/2117/403814
Access Level:Open access
Keyword:Kernel functions
Neural networks (Computer science)
deep kernel learning
xarxa neuronal
mètode del nucli
Nucli de funció de base radial
neural network
kernel methods
RBF kernel
Kernel, Funcions de
Xarxes neuronals (Informàtica)
Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial
Description
Summary:n recent times, there has been a growing interest in integrating kernel methods with neural networks, capitalizing on the expressive capabilities of the former and the generalization strengths of the latter. On of the aims of the study was contribute to the understanding and development of the kchain, a recent proposal in this line. Most prominently, this thesis introduces the Kernel Activation Layer (KAL), a novel activation function designed to process input from three linear layers using a kernel function. This research offers a comprehensive examination of the gradients associ- ated with the Radial Basis Function (RBF) kernel and presents an analytical approach to selecting the optimal value for its hyperparameter, denoted as ¿. Experimental findings illustrate that models incorporating the KAL can match the performance of conventional neural networks, particularly excelling in defining clear decision bound- aries for non-linear datasets with sharp class boundaries.