A recursive least squares training approach for convolutional neural networks
Date
2022
Authors
Yang, Yifan, author
Azimi-Sadjadi, Mahmood, advisor
Pezeshki, Ali, committee member
Oprea, Iuliana, committee member
Journal Title
Journal ISSN
Volume Title
Abstract
This thesis aims to come up with a fast method to train convolutional neural networks (CNNs) using the application of the recursive least squares (RLS) algorithm in conjunction with the back-propagation learning. In the training phase, the mean squared error (MSE) between the actual and desired outputs is iteratively minimized. The recursive updating equations for CNNs are derived via the back-propagation method and normal equations. This method does not need the choice of a learning rate and hence does not suffer from speed-accuracy trade-off. Additionally, it is much faster than the conventional gradient-based methods in a sense that it needs less epochs to converge. The learning curves of the proposed method together with those of the standard gradient-based methods using the same CNN structure are generated and compared on the MNIST handwritten digits and Fashion-MNIST clothes databases. The simulation results show that the proposed RLS-based training method requires only one epoch to meet the error goal during the training phase while offering comparable accuracy on the testing data sets.
Description
Rights Access
Subject
recursive least squares
convolutional neural networks