Neurosymbolic AI could be leaner and smarter than today’s LLMs

Could AI that thinks more like a human be more sustainable than today’s LLMs? The AI industry is dominated by large companies with deep pockets and a gargantuan appetite for energy to power their models’ mammoth computing needs. Data centers supporting AI already account for up to 3.7% of global greenhouse emissions.

In a Perspective published in PNAS Nexus, Alvaro Velasquez and colleagues propose an alternative model: neurosymbolic AI, which would require far less computing power, creating opportunities for smaller players to enter the field and allowing society to enjoy the benefits of AI without the environmental costs.

Neurosymbolic AI is built on data-driven neural methods and classical symbolic approaches and is partly inspired by the efficiencies of the human brain, which operates on about 20 watts of power and exhibits similar bimodal fast and slow thinking akin to neural learning and symbolic reasoning. Symbolic approaches are those that rely on semantically meaningful symbols to structure knowledge, which include logic and differential equations.

The authors show how neuro-symbolic AI models could use these abilities to reduce the volume of data and parameters otherwise required to produce reliable outputs. Rather than requiring a statistically robust correlation to emerge out of a vast dataset, models can learn some basic axioms or facts from data (e.g., “All men are mortal” and “Socrates is a man”) and infer the validity of related facts by composing such axioms using symbolic logic (e.g., “Socrates is mortal.”). Such models could be 100 times smaller than today’s leading LLMs.

According to the authors, neurosymbolic AI could enable efficient and trustworthy AI systems without unsustainable energy use or gatekeeping by companies with large financial resources.

Provided by
PNAS Nexus