Modular learning of convolutional neural networks

In recent years, a wide variety of network structures and training methods have been proposed for deep learning. However, the underlying mechanism of a deep network is not fully understood, which is considered as a black box system. The layer-wise learning method was proposed a decade ago, but now i...

Full description

Saved in:
Bibliographic Details
Main Author: Wang, Jinhua
Other Authors: Cheah Chien Chern
Format: Thesis-Master by Coursework
Language:English
Published: Nanyang Technological University 2023
Subjects:
Online Access:https://hdl.handle.net/10356/163999
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
Description
Summary:In recent years, a wide variety of network structures and training methods have been proposed for deep learning. However, the underlying mechanism of a deep network is not fully understood, which is considered as a black box system. The layer-wise learning method was proposed a decade ago, but now it is rarely used due to the trade-off in performance as compared to the standard end-to-end learning. Recently, layer-wise learning has been considered for application in interpretable or analytical neural networks. Therefore, a key target is to improve the performance of the layer-wise learning. In this dissertation, a modular deep learning method is developed on the basis of classical layer-wise learning. In addition, the network performance is further improved by proposing epoch-wise learning on the basis of modular deep learning. Through the case studies using several common datasets, the proposed approaches are compared with the traditional layer-wise learning method in terms of performance.