DeepLET in short

Energy Efficient and Trustworthy Deep Learning (acronym: DeepLET) is a research project which was launched on 4th December 2023 and is funded by the Hellenic Foundation for Research & Innovation (H.F.R.I) and European Eunion under the call “Basic Research Financing (Horizontal support for all Sciences), National Recovery and Resilience Plan (Greece 2.0)”. The project which will be running for two years, is coordinated by the Computational Intelligence and Deep Learning (CIDL) research group of the Aristotle University of Thessaloniki, while one more partner, Aarhus University of Denmark, will contribute to the successful implementation of the project. 

Deep Learning (DL) has achieved tremendous performance jumps in the last decade in several computer vision and machine learning tasks. However, DL models are becoming more and more complex, requiring vast amounts of computational power and energy both for training and inference purposes. These requirements are becoming especially limiting in many applications, where significant energy and computational power constraints exist, restricting the speed and the accuracy of the deployed models, while the large-scale deployment of DL can have a significant environmental impact and comes with a rising energy cost. At the same time, DL typically leads to black-box models that cannot provide any explanation for the reasons for which they took a specific decision. This limitation is especially important for many performance critical applications, in which the wrong decisions made by the model can cause immediate harm, e.g., in autonomous driving. 

DeepLET aims to overcome these limitations by developing novel methods along with the appropriate theory for designing, training and deploying lightweight and energy efficient DL models. DeepLET also aims to develop trustworthy and explainable deep learning models that will allow for increasing the confidence in the way DL models work, e.g., interpreting the way DL models behave, increasing the trust in DL models, as well as identifying situations where DL models must not be trusted. Finally, DeepLET aims to demonstrate the actual efficiency of the developed energy efficient, lightweight and trustworthy DL models by providing a collection of open source tools (in the form of an open source library) to the research community and industry, enabling them to use, adapt and extend the developed methods. DeepLET is a high-risk high-gain project that goes beyond the current state-of-the-art, aiming to solve an important problem that currently prohibits DL from providing effective solutions for various applications.