Green AI em visão computacional: uma análise de eficiência energética e pegada de carbono no treinamento de modelos de deep learning

Abstract

The growth of deep learning has driven significant advances in computer vision but has also intensified concerns regarding the energy consumption and CO2 emissions associated with training complex neural networks. The common practice of extending processing time through the simple scaling of training epochs often results in diminishing returns in predictive accuracy and a waste of resources. This work investigates the feasibility of achieving accuracy gains in an energy-sustainable manner by applying the principles of Green AI. A controlled experiment was conducted using the Fashion-MNIST dataset to train three distinct architectures: a custom convolutional neural network (CNN), a You Only Look Once (YOLO26) object detector adapted for classification, and a DeiT-Tiny Vision Transformer (ViT DeiT-Tiny). Total energy consumption and the corresponding carbon footprint were monitored in real time using the CodeCarbon library. The experimental protocol compared simple epoch scaling (10, 20, and 30 epochs) using standard parameters against runs with optimized hyperparameters—including dropout regularization, label smoothing, the AdamW adaptive optimizer, and data augmentation. The results demonstrated that increasing the number of epochs under standard configurations is highly inefficient. For the convolutional network trained for 30 epochs with the standard configuration, carbon emissions rose by over one thousand percent (1.041%), increasing from 0, 398 g to 4, 541 g, while test accuracy declined from 93, 44% to 93, 24% due to severe overfitting. Conversely, hyperparameter optimization enabled the models to generalize efficiently. Within the same 30-epoch budget, the optimized version of YOLO26 achieved an absolute increase of 1, 61% in top-1 accuracy (reaching 91, 74% versus 90, 13% for the standard version), while simultaneously reducing carbon emissions by 48, 0% (from 7, 411 g to 3, 855 g) and training time by 46.7% (from 206, 06 minutes to 109, 79 minutes) compared to its standard counterpart. Additionally, it was found that applying calibrated hyperparameters within short training budgets acts as a key driver of efficiency, allowing the Vision Transformer to achieve predictive performance gains (reaching 95, 08% accuracy versus 94, 97% for the standard model) alongside a 16, 0% reduction in total emissions (from 10, 157 g to 8, 527 g). It is concluded that rigorous hyperparameter calibration and the introduction of regularization techniques are essential for ensuring the environmental scalability of machine learning, demonstrating that accuracy and sustainability can be optimized concurrently.