Machine Learning Enhanced Model Selection Methodologies and Optimization of Deep Learning (DL) Inference for accuracy, inference time and energy on Mobile and Embedded Systems

Faculty Development Competitive Research Grant Program 2021-2023

In this project, the investigators will characterize mobile gaming and Deep Learning (esp. Convolutional Neural Networks) workloads (i.e., CPU or GPU dominant workloads in gaming; and loading and inference—convolution, pooling and fully connection layers-- workloads in DL) using collected data such as utilization, frequency, performance metrics and power consumption on the mobile platforms. As a further phase, we build models from a simple yet effective heuristic model based on the characterization to more sophisticated ML enhanced models considering both accuracy and interpretability; and then, we propose Dynamic Power Management (DPM) strategies that orchestrate energy-efficient policies among heterogeneous processors and evaluate the strategies using gaming and DL test datasets. As the last phase, we introduce energy-efficient DL inference optimization through adaptive model selection for mobile and embedded systems based on all the previous observations and analyses.