Green AI

Machine learning, Energy efficiency, Ensemble learning, Model optimization, Accuracy-energy trade-off, Feature selection, Sustainable AI, Green AI

• Explore methods for optimizing machine learning models for energy efficiency.
• Investigate techniques to balance accuracy and energy consumption in feature selection and data processing.
• Study approaches to detect and handle concept drift in ML systems.
• Evaluate the impact of different embedding strategies on energy consumption.
• Apply reinforcement learning to dynamically optimize model configurations for improved performance and efficiency.

Our research group explores strategies to enhance the energy efficiency of ML-enabled systems through Green AI initiatives. We focus on ensemble learning, investigating how various configurations, including model size and fusion methods, can optimize the trade-off between prediction accuracy and energy consumption. Additionally, we examine feature selection methods to identify the most relevant features, reducing energy usage during both model training and inference.

Our work also addresses concept drift, where we evaluate different detection techniques to maintain model effectiveness over time while minimizing unnecessary retraining costs. We analyze the impact of embedding techniques in transformers, comparing static and dynamic embeddings to assess their energy implications. Moreover, we leverage reinforcement learning to dynamically optimize ML models configurations, enabling a more adaptive and sustainable approach to machine learning applications. Through these efforts, we aim to contribute to the development of environmentally responsible AI solutions.

Ensemble learning, Feature selection, Concept drift, Embedding techniques, Reinforcement learning