The unprecedented explosion of Artificial Intelligence (AI) and its applications in everyday personal and professional life is posing a major concern for the environment. Among the various reasons that make AI development extremely energy-demanding is the fact that there are generally no all-purpose guidelines when searching for the best model for a given task. Usually, many different model configurations are tested, and only the one which yields the best results is retained. Not only is this trial-and-error approach a significant waste of energy and resources, but it also leads to a centralization of innovation capabilities in a handful of big players, who can speed up the process by relying on unmatched computational firepower. To tackle the issue, we are proud to introduce NebulOS. Unlike traditional methods that search for the best model only considering the performance on the downstream task, our novel approach also takes into the energy consumption when training a model on a specific hardware, producing tailored designs on a variety of different devices. This ensures that the final model aligns with the end-user’s unique needs, resources, and preferences.
Automating the procedure of architecture design and tailoring it to be as energy-efficient as possible holds the promise of enabling a new generation of genuinely green AI models, and to stimulate a much broader involvement of underprivileged players – such as public research institutes – in the AI scene.