Applying continual learning strategy on random vector functional link neural network based ensemble deep learning

Over the past 20 years, the rapidly developing computer hardware made our computer own more powerful abilities to compute and store crucial information, where more and more researchers could be able to practically research on Machine Learning and Deep Learning. Thus, because of the satisfactory resu...

Full description

Saved in:
Bibliographic Details
Main Author: Peng, Liuchang
Other Authors: Ponnuthurai Nagaratnam Suganthan
Format: Thesis-Master by Coursework
Language:English
Published: Nanyang Technological University 2021
Subjects:
Online Access:https://hdl.handle.net/10356/154367
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
Description
Summary:Over the past 20 years, the rapidly developing computer hardware made our computer own more powerful abilities to compute and store crucial information, where more and more researchers could be able to practically research on Machine Learning and Deep Learning. Thus, because of the satisfactory results yielded by Machine Learning and Deep Learning, increasing attention is paid to these areas and correspondingly, various networks are presented. In this dissertation, Random Vector Functional Link Neural Network (RVFLn) and its variant Random Vector Functional Link Neural Network based on Ensemble Deep Learning(ed-RVFLn) are the main research targets. These networks could gain robust results in classifying tasks if the network is given all the datasets at once in the training process. However, in the real world, the environment is changing all the time and the datasets could become a continuous stream and never end. Under this situation, the network needs to do Continual Learning on the streaming datasets and continually increase its ability for classifying. Therefore, in this project, the team tried to do the combination between the Continual Learning strategy and the proposed network, RVFL, ed-RVFL. After doing the comparisons and analysis, I found that the Continual Learning strategy has the possibility and indeed can be practically applied to the RVFL network and ed-RVFL network which use the Ridge Regression method to compute the weights. The accuracy after applying a continual learning strategy can reach 84% of the original accuracy and the results generated on edRVFL network are 1.36% better than the results generated on RVFL network.