Improving deep RVFL neural networks on large datasets

• Main Author: Li, Bing
• Other Authors: Ponnuthurai Nagaratnam Suganthan
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Electrical and electronic engineering
• Online Access: https://hdl.handle.net/10356/152334
• Institution: Nanyang Technological University

Recently, neural networks algorithm is becoming popular among researchers for classification problems such as Handwritten Character Recognition. Gradient descend is a kind of popular method which usually used in neural networks. However, such iteration methods usually lead to train the network slowly. Besides, overfitting the training data is also a problem of this approach. As such, researchers studied the randomised neural networks such as Random Vector Functional Link (RVFL) neural networks and Extreme Learning Machine (ELM) which significantly reduce training time and yet produce good classification performance. ELM neural network is a kind of simplified RVFL neural network. Compared with RVFL, it does not contain the direct link and bias between the input layer and output layer. In this project, improving deep RVFL neural networks are applied to classification problem. This work contains mainly deep RVFL neural networks and Convolution RVFL neural network. This latter combines RVFL neural network and CNN. The parameters in the convolution kernel are generated randomly in a certain range and keep fixed. The input of fully-connected layer includes the original input data and the data after convolved. From the results of testing, it can be seen that deep RVFL neural network is more suitable for tabular data, while the performance of CRVFL is better than traditional deep RVFL on image input data.