Text

GreenDL: Green Deep Learning for Edge Devices

GreenDL aims to provide novel theoretical foundations and practical algorithms to automatically design scalable energy-efficient DL models with low energy footprint and facilitate fast deployment of complicated DL models for various Edge devices satisfying given hardware constraints This project will remarkably investigate performance analysis and modeling, optimization, and learning algorithms followed by extensive experiments.

Start

2022-01-01

Planned completion

2026-01-01

Main financing

Project manager at MDU

No partial template found

Deep Learning (DL) has been permeated into different aspects of our daily life due to providing extremely promising results for various learning tasks. Neural Architecture Search (NAS) showed impressive results in automatically designing efficient DL models. On the other hand, compilation tool-flows are state-of-the-art methods for deploying DL models on hardware. However, both the design and deployment of DL models suffer from significant energy consumption and enormous optimization time for data at scale, leading to substantial environmental costs. The problem is pronounced by exponentially increasing the amount of data produced by billions of Edge devices connected worldwide. With such an inefficient processing paradigm, we inevitably need novel and more efficient solutions to design and deploy DL models on Edge devices.

The proposed solutions should enable

  • exploiting both design and development stages
  • energy-efficient optimization methods scaling to extensive benchmarks
  • meeting the resource constraints of Edge devices.

Project objectives

GreenDL aims to provide novel theoretical foundations and practical algorithms to automatically design scalable energy-efficient DL models with low energy footprint and facilitate fast deployment of complicated DL models for various Edge devices satisfying given hardware constraints This project will remarkably investigate performance analysis and modeling, optimization, and learning algorithms followed by extensive experiments.