Studies on optimal deep learning Methodologies for data Classification

Abstract

The research reported in this thesis deals with the Deep Clustering Network (DCN) algorithm for efficient data classification. The primary objective of this research is to make deep learning algorithms to perform unsupervised learning for the clasification of data. By investigating various deep clustering algorithms, the problem of enhancing the Deep Neural Network (DNN) with unsupervised clustering algorithms is studied. This research explores the crucial ingredients required for the deep learning algorithm to perform unsupervised learning optimally. This study provides a detailed understanding of the significance and challenges that created the research opportunity. Further, an investigation of the various clustering algorithms, along with its insights, is studied and discussed. Furthermore, the algorithm Malleable Fuzzy Local Median C-Means (M-FLMCM) is proposed to enhance the quality of clustering towards the research objective to tune the DCN. Again the algorithms such as Fraternal K-Median (FK-Med), Meticulous Fuzzy Convolution C-Means (MFCCM), and Fuzzy Local Information K-Means (FLIKM) is proposed and derived from overcoming the limitation of the M-FLMCM. Finally, the proposed algorithms are enhanced with the DNN, where the AutoEncoders and the Convolutional Neural Network (CNN) are used. The results of the proposed works and its DNN enhancements are studied via the performance metric such as Unsupervised Clustering Accuracy (ACC), Normalised Mutual Information (NMI) and the Adjusted Rand Index (ARI). The investigation towards these deep clustering algorithms is analysed with the standard datasets. Based on the results, sufficient inferences are acquired to make further studies. newline