Deep Learning-Driven Structured Energy Efficient Affordable Ecosystem for Computational Learning Theory

Abstract

Advancement in machine learning (ML) field is rising computational demands during deep learning (DL) applications development process. Emphasis on statistical, theoretical, and computational aspects of learning and graphical models, networked with sparse analysis, tensor, and topological methods, numerical simulation at various abstraction levels holds paramount importance while analyzing performance of a learning algorithm. The demand for high-end deep learning oriented, compute intensive, specialized accelerated hardware along with performance libraries is essential to train DNN models. Availability of such cost-effective computation ecosystem for research and development is challenging because of limiting factors like the financial viability, lack of appropriate technical hardware, and software knowledge while procuring, less expertise while determining and installing required computing environment, performance libraries, frameworks, tools, etc. With the availability of big data, innovation in chip design technology, significant improvement in neural network design, improvement in speed and accuracy of deep learning model has made it possible to avail necessary DL development ecosystem to common masses under affordable cost. The aim of this paper is to present the study of ready-to-use, end-to-end deep learning solution, suitable for simulation, modeling, and analysis, having latest generation of processor, accelerated computing software development environment for DNN applications development, video analytics, linear algebra, sparse matrix operations, multi-GPU communication, training-inference engine development. This enables users at grass root level to do R&D irrespective of their geographical location and environmental conditions while building technical capabilities to solve India specific problem statements.