Unsupervised Feature Learning on Big Data Based on Deep Learning with Weighted Softmax Regression
Sangram Keshari Swain1, Srinivas Prasad2, M. Vamsi Krishna3
1Sangram Keshari Swain, Research Scholar, Department of Computer Science and Engineering, Centurion University of Technology, and Management, Odisha, India.
2Srinivas Prasad, Professor, Department of Computer Science & Engineering, K L University, (Andhra Pradesh), India.
3M.Vamsi Krishna, Associate Professor, Department of Computer Science & Engineering, & Technology, Parlakhemundi Campus, Centurion University of Technology, (Odisha), India
Manuscript received on 07 March 2019 | Revised Manuscript received on 20 March 2019 | Manuscript published on 30 March 2019 | PP: 752-761 | Volume-8 Issue-5, March 2019 | Retrieval Number: D2717028419/19©BEIESP
Open Access | Ethics and Policies | Cite | Mendeley | Indexing and Abstracting
© The Authors. Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP). This is an open access article under the CC-BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Abstract: Deep learning is the recent technology which is effectively applied to feature learning in image classification, feature learning and language processing. However, recent deep learning models process with vector space which produces failure when learning features from non-linear distribution of heterogeneous data. Large amount of labeled data can be required for supervised learning of recurrent neural networks (RNN). These large volume of training samples are time consuming and high cost. One direction of addressing this problem is to extract features from unlabeled data. This paper proposes a deep computation model for feature learning on big data to learn underlying data distribution using Deep Recurrent Neural Network based weighted Softmax regression (DRNNWSR) with no need of labeled instances. The proposed approach is moderately simple, however achieves accuracy comparable to that of more advanced techniques. The proposed strategy is significantly easier to train, contrasted with existing neural system strategies, making less prerequisites on manually labeled training data. It is additionally appeared to be impervious to over fitting. We give comes about on some outstanding datasets, specifically STL-10, Caltech-256, and Caltech 101 and CIFAR-10. The results show that the proposed system obtains really high order accuracy and is superior to the present techniques for the broad dataset. Because of learning features adaptively, the proposed system diminishes the need of tedious and makes data classification more efficient. Our numerical results demonstrated good convergence when compared to the different datasets for different classifiers
Keyword: Unsupervised Learning, Recurrent Neural Network, Big Data, Softmax Regression, Deep Learning.
Scope of the Article: Deep Learning