Main Article Content
As high dimensional data is increasing day by day, importance of fast and precise features extraction is also increasing. Huge work in deep learning leads to autoencoders that are very useful methods for feature extraction as they tend to recreate the original data from same features. In deep learning, a lot depends on loss and objective function. Normally loss functions depend on data only without any relationship between data. In this study, an extended relational model for autoencoders has been proposed that maps original data on the basis of data with combination to relation between data. This relation is based on ratio between data variance and magnitude. Convolutional autoencoder has been evaluated with proposed relational model. Benchmark datasets of Mnist and Cifar10 have been used for experimental results. Comparison of proposed model has been made with different available loss functions and experimentally it has been proved that proposed relational model achieve lower construction loss with better accuracy and visual results.
Copyright (c) 2020 Technical Journal
The author transfers all copyright ownership of the manuscript entitled (title of article) to the Technical Journal in the event the work is published.