A1672
Title: Displaying the performance-consumption tradeoff for aware and sustainable AI
Authors: Enrico Capuano - Politecnico di Torino (Italy) [presenting]
Claudia Berloco - Intesa Sanpaolo (Italy)
Abstract: The rapidly evolving field of generative AI has witnessed a proliferation of AI models, often prioritizing performance metrics at the expense of energy consumption. As the academic community and civil society increasingly focus on environmental, social, and governance (ESG) issues, concerns about the environmental impact of resource-intensive computations in natural language processing (NLP) are growing. Large language models (LLMs) require substantial computational power during both training and deployment stages. In contrast, machine learning and statistical learning models are less computationally intensive, albeit capable of accomplishing different tasks. The aim is to raise awareness among developers and users about the energy consumption associated with different models by comparing their performance on a common text classification task. The trade-offs between prediction accuracy and power consumption are highlighted by juxtaposing performance metrics with energy consumption. The proposed methodology promotes responsible and sustainable AI development practices, enabling researchers and practitioners to make informed decisions that balance accuracy, speed, and energy efficiency. By promoting transparency and accountability in AI model development, the contribution is to the broader discourse on sustainability and responsible innovation in artificial intelligence.
