ABSTRACT

This thesis focuses on approximate computing, a strategy to balance performance and energy efficiency by trading a slight loss of accuracy for significant speed and power gains. However, ensuring acceptable result quality remains a challenge. The study introduces a novel approach that utilizes machine learning to predict the quality of approximate computations. The goal is to create a robust framework that helps designers make informed decisions between accuracy and efficiency. A comprehensive dataset is compiled, containing diverse approximate computing instances with quality metrics. Various features from the dataset, including computation traits and error rates, are used as inputs for machine learning models. After rigorous exploration of machine learning algorithms, a refined model that demonstrates strong predictive performance across various inaccurate computing scenarios is selected.

In order to enhance the capacity to forecast outcomes in approximate computing more precisely, this research applies transfer learning techniques. Pretrained models are more accurate than the originally trained model when they are adjusted using specific, imperfect computational input. The framework consistently produces correct quality estimates for a variety of computations and has undergone extensive validation through a battery of tests. It aids in the decision-making process for highly effective and energy-efficient system designers. The framework's flexibility and generalizability are validated by empirical evaluation of real-world applications. In order to assess approximation accuracy, a unique machine-learning process is proposed in this study. It illustrates how effective transfer learning is in this situation. This approach helps designers to effectively control the trade-off between accuracy and efficiency in the rapidly developing field of approximation computing, which is becoming more and more crucial as the need for efficient computing systems grows.

xii