Energy Efficiency in Machine Learning

This DC aims at applying reversible computation techniques (in collaboration with DC4 and DC6) as well as other energy saving approaches to enhance energy efficiency in ML, and more specifically in Deep Neural Networks (DNN). Energy consumption will be estimated using approaches from the literature. We will re-design algorithms and re-develop software components used in the training phase of DNN layers by applying reversibility and other energy saving approaches. Concurrently, we will create modelling frameworks for estimation of energy consumption that are appropriate for experimenting with software at instruction and application levels. We will then carry out measurements and analysis work to empirically evaluate suitability of (a) the energy modelling frameworks, and of (b) the proposed energy reducing techniques.

1) Development of reversible versions of DNN layer learning algorithms, also enhanced with other energy saving features; and their implementations; 2) Formulation of suitable models for estimation of energy consumption of software, called the energy models; 3) Measurement of energy consumption of the developed software, and evaluation of the energy models; 4) Analysis of the effect of reversible computation techniques and other approaches on energy efficiency, as well as on time and space complexity.

ULEIC, M16, R. Raman, reversing algorithms for DNNs, developing their implementations; M21, UoM, M. Lujan, models to estimate energy consumption of software, experiments to measure it; M32, VAIRE, H. Earley, language validation.