© 2018, IJCSE All Rights Reserved 701 International Journal of Computer Sciences and Engineering Open Access Review Paper Vol.-6, Issue-12, Dec 2018 E-ISSN: 2347-2693 A Comprehensive Study of Deep Learning Architectures, Applications and Tools Nilay Ganatra 1* , Atul Patel 2 1,2 Dept. of Computer Science and Applications, CHARUSAT, Changa, Gujarat, India * Corresponding Author: nilayganatra.mca@charusat.ac.in Tel.: +91 – 99240 48124 Available online at: www.ijcseonline.org Accepted: 25/Dec/2018, Published: 31/Dec/2018 Abstract— The Deep learning architectures fall into the widespread family of machine learning algorithms that are based on the model of artificial neural network. Rapid advancements in the technological field during the last decade have provided many new possibilities to collect and maintain large amount of data. Deep Learning is considered as prominent field to process, analyze and generate patterns from such large amount of data that used in various domains including medical diagnosis, precision agriculture, education, market analysis, natural language processing, recommendation systems and several others. Without any human intervention, deep learning models are capable to produce appropriate results, which are equivalent, sometime even more superior even than human. This paper discusses the background of deep learning and its architectures, deep learning applications developed or proposed by various researchers pertaining to different domains and various deep learning tools. Keywords— deep learning, architectures, applications, tools I. INTRODUCTION Deep Learning is considered as the subset of machine learning which is intern subset of Artificial Intelligence, a prominent field of computer science over the past decade. Artificial Intelligence makes machines to think intelligently without minimal human intervention. Machine learning comprises with various algorithms that are capable to model high level abstractions from input data. Deep Learning provides a more adaptive way using deep neural network that learns feature itself from the given input data and make machine capable for taking decision. Unlike task specific algorithms of machine learning, deep learning is a method based on learning data representations. Learning can be supervised, semi-supervised or unsupervised. Deep learning provides set of algorithms and approaches that learns features and tasks directly from data. Data can be of any type, structured or unstructured, including images, text or sound. Deep learning is often referred as end-to-end learning because it learns directly from data. Moreover, Deep learning techniques works without human mediation and sometime capable of producing more accurate result than human being itself. Nowadays, deep learning is widely used in the areas like computer vision, natural language processing, pattern recognition, object detection. Representative learning methods of deep learning provides multiple level of representation, generated by linking simple but non-linear modules that transmute the representation at one level into a representation at next higher layer, slightly in more abstract way. II. DEEP LEARNING: OVERVIEW AND BACKGROUND Machine Learning provides the vast collection of algorithms including singular algorithms, together with statistical approach like Bayesian networks, function approximation such as linear and logistic regression, or decision trees. These algorithms are indeed powerful, but there are few limitations of these algorithms when to learn for enormously complex data representations. Deep learning is emerged from cognitive and information theories and human neurons learning process along with strong interconnection structure between neurons is looking for to imitate. One of the key feature of computing neurons and the neural network model is, it can be able to apply generic neurons to any type of data and learn comprehensively [1]. Deep learning is considered as a promising avenue of research as it is capable of automatically identifying the complex features at a high level of abstraction. It is about learning multilevel representation and abstraction, which is useful for the data, such image, text and sound. One of the exceptional Deep learning characteristics is its capability of using unlabeled data during the training process [2]. According to the definition of Deep learning, it is considered as the application of multi-layer neural network with multiple neurons at each layer to perform the desired tasks like classification, regression, clustering and others. Fundamentally, each neuron with activation function is considered as the single logistic node, which connected to the