Large artificial intelligence (AI) systems keep improving but consume a great deal of energy during training and everyday use. The human brain, by contrast, needs only about twenty watts. Researchers at Graz University of Technology have developed a new AI model inspired by the brain. This model can plan in flexible ways and solve complex problems while using far less energy than the multi-layer neural networks or large language models common today.
The researchers note that the brain works in a completely different manner from present computer systems. The goal is to turn those brain methods into algorithms that AI can use.
Drawing on studies of the hippocampus, the researchers identified three key mechanisms the brain employs for planning and problem solving. Cognitive maps convert relationships among abstract objects into geometric patterns among neural activity, giving the system a sense of direction similar to a spatial map. Stochastic neural computations continuously generate possible scenarios and predictions. Compositional coding breaks information and action plans into reusable pieces that can be combined as needed.
These mechanisms allow the model to invent and test possible steps toward a solution without calculating every path completely. When a randomly chosen intermediate step points toward the goal, the cognitive map guides the system along that route. At the new position it again weighs options and moves gradually closer. In this way the model can adapt to changed or unexpected situations without needing to be retrained.
Testing and possible uses
The model was checked on three tasks: navigating a two-dimensional space, finding direction in an abstract multi-dimensional space, and assembling or taking apart a silhouette made of building blocks. The approach is not intended to replace today’s large language models. It instead offers a foundation for an alternative style of artificial intelligence suited to certain applications. The work remains at an early stage, yet it shows that capable artificial intelligence does not have to depend on huge data centres and massive energy supplies.
In the longer term such brain-inspired systems could be useful in robots, autonomous vehicles or other edge devices—machines that must operate locally with only limited power available.
This research is published in Nature Machine Intelligence.