Title:  Self Adaptive Reconfigurable Arrays: ML-assisted HW Architectures for the Ubiquitous and Evolving AI Landscape

Committee: 

Dr. Krishna, Advisor      

Dr. Mukhopadhyay, Chair

Dr. Kim

Abstract: The objective of the proposed research is to create architectures, tool, models, and methodologies for building efficient accelerators, that are adaptable to the changing landscape of machine learning workloads and applications. The advent of new deep learning algorithms have transformed the world of computing as we know it. While in one hand these algorithms have engendered solutions to previously unsolvable problems in machine comprehension and data analysis, the huge compute demand for running these algorithms have created new exciting problems in systems and hardware design. Computer architects have taken bold strides in democratizing the power of these algorithms by creating high performance and energy efficient accelerators. However, as the landscape of deep neural networks continue to evolve so does the requirements of acceleration. Unfortunately designing accelerators is hard and expensive. This thesis proposes to address some of these problems by exploring architecture design, developing tools and analytical models, and finally by channeling the power of machine learning into the design process of machine learning accelerators.