Higher Level Abstraction of Deep Architecture

Abstract

Two topics of data science are big data analytics and deep learning. Big data has become important as many private and public organizations collect massive quantities of domain-specific information that may include useful data on problems such as national intelligence, cyber security, detection of fraud, marketing and medical computing. Corporations such as Google and Microsoft analyze large volumes and impact existing and future technology for business analysis and decision-making. High-level, complex abstractions are extracted from deep learning algorithms through the hierarchical process. As data representations, complex abstractions are learned at a particular level on the basis of comparatively simple abstractions formulated at the previous level. The analysis and learning of large volumes of unattended data is a key advantage of deep learning and makes this an invaluable tool for big data analytics, with raw data largely unlabeled and unclassified. We are investigating how deep learning can be used in this research to deal with some major problems in big data analytics, such as extracting complex patterns from massive data volumes, semantic indexing, data tagging, quick recovery of information and simplifying discriminative tasks. We are examining several aspects of deep learning research, including streaming data, high-dimensional data, model scalability and distributed computers, which requires further exploration to incorporate specific challenges that have been introduced by large data analytics. In conclusion, we provide insights into relevant future works with questions such as the definition of data sampling criteria, domain adaptation modeling, the definition of criteria for useful data abstractions, better semantic indexing, semi-supervised study and active learning.