Deep neural networks and large language models have conquered many tasks, but they consume enormous amounts of electricity — where the human brain manages intelligence on about 20 watts. A new framework published in Nature Machine Intelligence takes a different route: it borrows the brain's own data structure, the cognitive map, to build a network that solves unfamiliar problems with a fraction of the energy.
The approach, developed by researchers at Tsinghua University, Graz University of Technology and Italy's National Research Council, rests on an insight from cognitive science: brains don't store learned knowledge as tables of parameters, but as internal maps — spatial-like representations that enable rapid, adaptive decisions in real time. Most AI ignores this organization entirely.
The team's network, described in "Neural sampling from cognitive maps enables goal-directed imagination and planning," learns locally rather than through massive offline training. When given a new goal, it samples from a 2D cognitive map to imagine plausible trajectories before acting — an algorithmic echo of the mental simulation humans perform when planning. In initial tests, the network successfully planned and solved problems it had never encountered, and its decisions were far easier to interpret than those of typical deep models.
"We were intrigued by the fact that evolution had invented algorithms and data structures that produce intelligence in brains, and that these solutions differ strongly from those used in current AI," says senior author Wolfgang Maass. The framework is designed for neuromorphic and in-memory computing chips — the emerging hardware class, pursued by IBM, Intel and numerous startups, that aims to slash AI's power bill by processing data where it's stored.
The researchers are already working with engineers at Intel and a U.S. startup to port the algorithm to silicon. Maass' team plans to extend the approach to capture the brain's low-latency decision-making and its ability to explain its own choices — a key goal in AI research.
