Combining Active Learning and Mathematical Programming: a hybrid approach for Transprecision Computing

This paper explores the possibility of applying a hybrid approach between Active Learning and Mathematical Programming to Transprecision Computing. This would entail embedding a machine learning model trained by means of an Active Learning approach into an optimization model to automatically and intelligently tweak the representation of floating-point numerical data. This project aims to lower the energetic expenditure of every single intermediate computation in a given program, while also avoiding errors that are systematically introduced when manipulating variables using this technique, and ensure that they do not exceed a maximum acceptable error rate decided prior.