Researchers at Tsinghua University, Graz University of Technology, and the National Research Council in Italy have built an AI framework that mimics how the brain organizes knowledge, producing a neural network that solves unfamiliar problems and could run on energy-efficient neuromorphic chips. The work targets a central problem in AI: the massive energy consumption of deep neural networks and large language models.
The brain runs on about 20 watts, said Wolfgang Maass, senior author of the paper. Current AI systems require data centers with dedicated energy and cooling. The research, published in Nature Machine Intelligence, grew out of that contrast.
"We were intrigued by the fact that evolution had invented algorithms and data structures that produce intelligence in brains, and that these solutions that nature had found differ strongly from those that are used to produce intelligence in current AI," Maass said.
How the framework works
Cognitive science research suggests the brain encodes learned knowledge as cognitive maps - representations that help it respond quickly and adaptively to new problems. Most AI systems don't organize information this way. Maass and his colleagues built a framework that does, drawing on neural recordings collected while people solved different types of problems.
In initial tests, a neural network based on the framework planned adaptively and solved problems it had never encountered before. Its predictions and actions were also easier to interpret than those of many existing AI models.
Built for energy-efficient chips
The model learns locally and doesn't require the computationally intensive training that deep neural networks and large language models need. That makes it a candidate for neuromorphic chips and in-memory computing systems, which process data directly in storage rather than shuffling it between memory and processors. Maass said several large companies, including IBM and Intel, and numerous startups are developing such chips.
The team is now working with engineers at Intel and a US startup to implement the brain-like algorithm on chips. "We are continuing our work on porting brain-like algorithms into more energy-efficient hardware," Maass said.
Why this matters for science and research professionals
For research groups, the framework offers a path to AI that's both interpretable and energy-efficient. The model's decisions come with explanations - Maass said the brain automatically provides concrete experiences that support a chosen action, and making AI decisions explainable is a major goal in current AI research. Running on neuromorphic hardware could also make this kind of AI accessible to labs without large compute budgets. Researchers who want to build on this work can find relevant training in the latest AI courses.
Your membership also unlocks:
