Researchers at Graz University of Technology have built an artificial intelligence system that solves complex planning problems using the same energy-saving strategies the human hippocampus uses — consuming a fraction of the power of conventional deep neural networks.
The research
Wolfgang Maass and Yukun Yang at the Institute of Machine Learning and Neural Computation, working with international partners, drew on neuroscientific studies of the hippocampus to identify three brain mechanisms worth translating into code:
- Cognitive maps — the brain converts relationships between abstract objects into geometric relationships between neural codes, creating a "sense of direction" similar to a spatial map.
- Stochastic neural computation — the brain constantly generates hypothetical scenarios and predictions rather than computing every possibility exhaustively.
- Compositional coding — information and action plans break down into reusable components, like LEGO bricks that can be reassembled for new tasks.
When the AI faces a problem, it doesn't calculate every possible solution path. Instead, guided by its cognitive map, it generates a random intermediate step and checks whether that step moves it toward the goal. If it does, the system pursues that path, then repeats the process from its new position. This step-by-step approach gradually closes in on a solution without exhaustive computation.
The team validated the architecture across three benchmarks: 2D spatial navigation, abstract multidimensional orientation, and assembling and disassembling a silhouette made of building blocks. Critically, the system adapted to novel environments and structural changes without retraining or parameter optimization — a capability called zero-shot adaptability. The human brain runs on roughly 20 watts of power, while large language models can demand megawatt-scale energy.
Why it matters
This research isn't about beating large language models at their own game. As Maass puts it, "We are still at a relatively early stage of development. But our work shows that powerful AI does not necessarily require huge data centres and enormous amounts of energy." The implications extend beyond AI efficiency. The model suggests that the brain's ability to plan flexibly — without massive computation — relies on specific, identifiable strategies: mapping abstract relationships geometrically, testing random intermediate steps instead of exhaustive search, and reusing action components.
For anyone interested in their own cognition, this research highlights that human problem-solving is remarkably efficient. Your brain doesn't brute-force solutions. It uses spatial metaphors for abstract problems, samples possibilities rather than enumerating them, and recombines familiar actions into new plans. These are trainable skills.
What you can do
Try applying these principles to your own thinking. When facing a complex problem, sketch a rough "map" of how the key elements relate to each other. Instead of planning every step, pick one plausible starting move and evaluate whether it gets you closer to your goal. Break big tasks into reusable components you can rearrange for future challenges. This mirrors the brain-inspired strategy and may help you solve problems with less mental fatigue.
Source: Neuroscience News
Curious about your own brain? Take our free adaptive IQ test or try 306 brain training levels.