Publicação
Sustainable Machine Learning: A predictive model for reflective resource consumption
| Resumo: | The increasing deployment of Machine Learning (ML) has brought significant environmental concerns, particularly related to the energy-intensive nature of training deep Neural Networks (NN). This research addresses the urgent need for sustainable ML by proposing a predictive framework for estimating the carbon footprint of neural models. Rather than relying on posthoc measurements, the study explores the feasibility of predicting energy consumption before training begins. To build the predictive model, a dataset was created by generating over 2.000 neural network configurations, which were run across different datasets, hardware types (CPU and GPU), and training setups. For each configuration, key metadata was collected, such as model architecture, dataset size, and hardware specs, alongside energy consumption estimates obtained with CodeCarbon. Several regression algorithms were trained to predict emissions in kgCO₂eq, with Random Forest and XGBoost delivering the highest accuracy (R² ≈ 0.82). The results show that model architecture has a stronger influence on accuracy, while hardware type more strongly affects energy consumption. CNNs on GPU reached a strong balance, with a mean accuracy of 0.734 and an average carbon footprint of 0.00386 kgCO₂eq nearly half the emissions of their CPU counterparts (0.00609 kgCO₂eq). These findings support the idea that small sacrifices in performance can lead to substantial energy savings, highlighting the value of carbon-aware choices during model design. Overall, these insights can support more informed design decisions in ML, balancing predictive performance with environmental responsibility. |
|---|---|
| Autores principais: | Borbon, Margarida Guimarães |
| Assunto: | Sustainable Machine Learning Energy consumption Neural Networks Carbon footprint Energy estimation SDG 9 - Industry, innovation and infrastructure SDG 12 - Responsible production and consumption SDG 13 - Climate action |
| Ano: | 2025 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso embargado |
| Instituição associada: | Universidade Nova de Lisboa |
| Idioma: | inglês |
| Origem: | Repositório Institucional da UNL |
| Resumo: | The increasing deployment of Machine Learning (ML) has brought significant environmental concerns, particularly related to the energy-intensive nature of training deep Neural Networks (NN). This research addresses the urgent need for sustainable ML by proposing a predictive framework for estimating the carbon footprint of neural models. Rather than relying on posthoc measurements, the study explores the feasibility of predicting energy consumption before training begins. To build the predictive model, a dataset was created by generating over 2.000 neural network configurations, which were run across different datasets, hardware types (CPU and GPU), and training setups. For each configuration, key metadata was collected, such as model architecture, dataset size, and hardware specs, alongside energy consumption estimates obtained with CodeCarbon. Several regression algorithms were trained to predict emissions in kgCO₂eq, with Random Forest and XGBoost delivering the highest accuracy (R² ≈ 0.82). The results show that model architecture has a stronger influence on accuracy, while hardware type more strongly affects energy consumption. CNNs on GPU reached a strong balance, with a mean accuracy of 0.734 and an average carbon footprint of 0.00386 kgCO₂eq nearly half the emissions of their CPU counterparts (0.00609 kgCO₂eq). These findings support the idea that small sacrifices in performance can lead to substantial energy savings, highlighting the value of carbon-aware choices during model design. Overall, these insights can support more informed design decisions in ML, balancing predictive performance with environmental responsibility. |
|---|