Sustainable AI: Sparse Backpropagation for Low-Carbon On-Device Training

Deep learning’s growing computational demands have led to considerable energy consumption and carbon emissions. In this paper, we explore on-device learning with sparse backpropagation as a means to improve training energy efficiency. We apply this approach to several CNN architectures (ResNet-18, DenseNet-121, GoogleNet, and MobileNet-V2) trained on popular vision datasets (CIFAR-10/100 and Flowers-102) using an NVIDIA Jetson AGX Orin (64GB) edge device. We present a sparse training algorithm that updates only a fraction of model parameters per iteration, reducing computation in the backward pass. We integrate the CodeCarbon library to measure energy use and associated CO2 emissions. Experimental results show that sparse backpropagation can maintain model accuracy within 1–2% of standard training while reducing energy usage and carbon emissions by up to 30-40%. We analyze trade-offs between accuracy, speed, and sustainability and discuss deployment strategies for energy-efficient on-device learning. These findings demonstrate a practical step toward "Green AI" by cutting the carbon footprint of deep learning without significantly compromising performance.