I will be posting the papers around Green AI and Sustainable Machine Learning that I read and found interesting, with some notes on their contributions and some thoughts from myself.

This bibliography is part of my PhD program on Green AI with Luís Cruz at TU Delft, where I’ll be working in the SELF Lab for medical IoT edge devices.

Papers

📅 2024

Heli Järvenpää, Patricia Lago, Justus Bogner, Grace Lewis, Henry Muccini, Ipek Ozkaya (2024). A Synthesis of Green Architectural Tactics for ML-Enabled Systems. ICSE-SEIS.
Architectural Tactics ML systems

In this paper, the authors present a compilation of 30 green architectural tactics to improve the energy efficiency of Machine Learning systems. The tactics are organized by categories based on the different aspects of the development cycle of ML software. They compiled this list by analyzing 51 … Read more.

Timur Babakol, Yu David Liu (2024). Tensor-Aware Energy Accounting. ICSE.
Green Deep Learning Energy monitoring

The paper presents SMARAGDINE, an energy accounting system for Deep Learning models built with TensorFlow. Opposite to other similar systems, which are black or grey box systems (only report total energy consumption or consumption by hardware device), SMARAGDINE can collect the energy and pow… Read more.

📅 2023

Samarth Sikand, Vibhu Saujanya Sharma, Vikrant Kaulgud, Sanjay Podder (2023). Green AI Quotient: Assessing Greenness of AI-based software and the way forward. ASE.
Green AI in Industry Green AI Practices

The paper proposes an approach to characterize the sustainability of AI projects in industry. While there has been an increase in research on Green AI in academia, this increase has barely translated into industry application, with most industry AI projects not adopting sustainability practices p… Read more.

Alessandro Tundo, Marco Mobilio, Shashikant Ilager, Ivona Brandić, Ezio Bartocci, Leonardo Mariani (2023). An Energy-Aware Approach to Design Self-Adaptive AI-based Applications on the Edge. ASE.
Deep Learning Adaptative configuration Operation modes

The paper proposes a new framework for developing Energy-Aware Self-Adaptative AI applications. These applications can dynamically switch between different sets of configurations based on perceived conditions to prioritize different objectives, like switching from a low-energy, low-accuracy mode … Read more.

Roberto Verdecchia, June Sallou, Luís Cruz (2023). A systematic review of Green AI. WIDM.
Green AI Literature Review

This study performs a systematic literature review of the field of Green AI. They analyze 98 studies in the field of Green AI, and report a number of relevant statistics about the field. Most papers published to date propose solutions to tackle energy efficiency in AI, and focus more on **mo… Read more.

Tim Yarally, Luís Cruz, Daniel Feitosa, June Sallou, and Arie van Deursen (2023). Uncovering Energy-Efficient Practices in Deep Learning Training: Preliminary Steps Towards Green AI. CAIN.
Deep Learning Training optimization

The paper goes back to the drawing board and analyzes the optimization of Neural Networks not only from the point of view of getting maximum accuracy at any cost but also considering the energy demand of training. To do this, the authors choose the popular FashionMNIST and CIFAR-10 datasets, as w… Read more.

📅 2022

Stefanos Georgiou, Maria Kechagia, Tushar Sharma, Federica Sarro, Ying Zou (2022). Green AI: do deep learning frameworks have different costs?. ICSE.
DL frameworks Energy monitoring

The main contribution of the paper is to compare the energy efficiency of TensorFlow and PyTorch, two popular Python frameworks for Deep Learning. To do so, the authors run multiple benchmarks on both frameworks using the DeepLearningExamples collection from NVIDIA, which includes multiple mode… Read more.

📅 2021

Qingqing Cao, Yash Kumar Lal, Harsh Trivedi, Aruna Balasubramanian, Niranjan Balasubramanian (2021). IrEne: Interpretable Energy Prediction for Transformers. IJCNLP.
Transformer Models Energy prediction

The paper proposes a model to estimate the energy consumption of Transformer-based NLP models. They achieve this by abstracting the model into a tree, dividing its components into modules (e.g. BertSelfAttention), which can be composed by other submodules, Machine Learning primitives (e.g. LayerN… Read more.

📅 2020

Lasse F. Wolff Anthony, Benjamin Kanding, Raghavendra Selvan (2020). Carbontracker: Tracking and Predicting the Carbon Footprint of Training Deep Learning Models. Arxiv.
Deep Learning Carbon emissions

The paper introduces an open-source tool to measure and predict the energy usage and carbon emissions of training a neural network. To do so, the tool measures energy consumption for a small number of epochs using Intel RAPL to measure CPU and DRAM power usage and NVIDIA Management Library to mea… Read more.

Mohit Kumar, Xingzhou Zhang, Liangkai Liu, Yifan Wang, Weisong Shi (2020). Energy-Efficient Machine Learning on the Edges. IPDPSW.
Energy efficiency Code suggestions

An Eclipse plugin that offers code suggestions to improve the energy efficiency of machine learning code in Java. Although the paper presents the tool and suggestions in the context of machine learning, the rules target built-in Java functionality, like the best primitive types or best methods to… Read more.

📅 2019

Yue Wang, Ziyu Jiang, Xiaohan Chen, Pengfei Xu, Yang Zhao, Yingyan Lin, Zhangyang Wang (2019). E2-Train: Training State-of-the-art CNNs with Over 80% Energy Savings. NIPS.
Training Efficiency Edge AI

This paper presents three techniques to improve the energy efficiency of training Convolutional Neural Networks in edge devices by dropping and reducing unnecessary computations during the training phase. They propose 3 techniques at different levels: data, model, and algorithm, and manage to obt… Read more.

Eva García-Martín, Crefeda Faviola Rodrigues, Graham Riley, Håkan Grahn (2019). Estimation of energy consumption in machine learning. Journal of Parallel and Distributed Computing.
ML Energy Estimation Literature Review

The paper does an extensive literature review of the different techniques for estimating energy consumption in machine learning and deep learning software. It starts by reviewing existing energy usage estimation methods for general software: from software-level estimation with performance counter… Read more.