A Design-Space Exploration Framework for Application-Specific Machine Learning Targeting Reconfigurable Computing

Abstract

Machine learning has progressed from inaccessible for embedded systems to readily deployable, thanks to efficient training on modern computers. Regrettably, requirements for each specific application which relies on machine learning varies on a case-by-case basis. In each application context, there exists multiple conditions and specifications which call for different design implementations for optimal performance. In addition to that, targeting reconfigurable computing involves further considerations and workarounds such as quantization, pruning, accelerator design, memory usage and energy-efficiency for power-constrained systems. The aim of this Phd Project is to undertake an analysis and investigation of the limitations inherent in application-specific machine learning within the context of reconfigurable computing. Our objective is to investigate in this new dimension and propose a hardware/software framework to facilitate a meticulous design-space exploration, enabling the identification of optimal strategies for achieving an effective and efficient design process by exploiting dynamic reconfiguration.