Detection and location of contrast-aware and border-aware neurons in Convolutional Neural Networks
Deep leaning and the use of Convolutional Neural Networks is following an upward trend in popularity in the recent years. As a lot of the research in this area is done empirically and based on experiments against a collection of datasets, research has been pushed to use techniques that alter the tra...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis de maestría |
| Fecha de publicación: | 2022 |
| País: | España |
| Institución: | Universitat Politècnica de Catalunya (UPC) |
| Repositorio: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglés |
| OAI Identifier: | oai:upcommons.upc.edu:2117/375197 |
| Acceso en línea: | https://hdl.handle.net/2117/375197 |
| Access Level: | acceso abierto |
| Palabra clave: | Deep learning Neural networks (Computer science) Kernel functions deep learning kernels convolutional neural networks padding border aware neurons contrast aware neurons Aprenentatge profund Xarxes neuronals (Informàtica) Kernel, Funcions de Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial |
| Sumario: | Deep leaning and the use of Convolutional Neural Networks is following an upward trend in popularity in the recent years. As a lot of the research in this area is done empirically and based on experiments against a collection of datasets, research has been pushed to use techniques that alter the training behavior whose side effects have not been fully explored. In this thesis we aim to explore the effects of padding when used to allow convolution operations inside a neural network to be computed exhaustively though padding (artificial values added to complete a given shape) on the edges of the image. Previous research has been done studying the effects of padding for batching images as well as the effects of padding in the type of neurons generated in the network. This thesis aims to follow up on the second, where the concept of border aware neurons (BAN) was proposed. We will propose a method of detecting contrast aware neurons (CAN) and BAN in CNNs. We will also explore the effects of padding in the existence of CAN in the convolutional layers of a CNN by using the first proposed methods in a collection of models trained on Caltech101. Finally we will explore the existence of BAN as a subset of CAN and study the effects of padding in their appearance by clustering the sets of CAN found in the previous experiments. We found that CAN neurons can appear in a network with or without adding padding to the convolution but they have dissimilar characteristics. When padding is added, the contrast changes of the patterns is much greater. We also found that the existence of padding is the cause of BAN to appear in the network. These findings open new doors to future work: model pruning removing BAN neurons and study its implications, and optimization in the training methodology. |
|---|