‘Accuracy per FLOP’: A Green AI Metric for Fair and Efficient AI Development
Recent advancements in Artificial Intelligence (AI), particularly through Deep Learning and large-scale datasets, have led to the creation of powerful models like OpenAI’s GPT-$n$. These models have significantly enhanced Natural Language Processing and driven innovation in the field. However, concerns are growing about the monopolization of AI by entities with the resources to develop such large models.
The current emphasis on accuracy favors well-funded projects, potentially stifling innovation and disadvantaging researchers with limited resources. An alternative metric, “accuracy per FLOP” (floating-point operations), could provide a fairer comparison by measuring model efficiency relative to computational cost. The concept of “accuracy per FLOP” aligns with “Green AI,” a term from Schwartz et al., 2020, which advocates for considering computation cost in AI research.
To promote inclusivity and innovation, researchers should explore new techniques and models that improve efficiency rather than merely increasing model size. This shift is crucial to ensure AI research benefits everyone and avoids a potential decline in AI progress, known as an Artificial Intelligence Winter. Human-centered AI, focusing on human interaction, should also remain a primary research focus.