Brain-Inspired AI Framework Could Complete Tasks With Less Energy

Researchers have developed a new artificial intelligence framework inspired by how the human brain organizes knowledge and solves unfamiliar problems. The approach could help create AI systems that require less computing power, operate more efficiently and offer clearer explanations for their decisions.

The research, conducted by scientists at Tsinghua University, Graz University of Technology and Italy’s National Research Council, was published in Nature Machine Intelligence. The framework uses concepts from cognitive science and neuroscience rather than relying exclusively on the computational strategies behind today’s deep neural networks and large language models.

Researchers Look to the Brain for More Efficient AI

Modern AI systems can recognize patterns, generate content and solve increasingly complex problems. However, many deep neural networks and large language models require substantial computing infrastructure and electricity, particularly during training and large-scale deployment.

Researchers have increasingly explored alternative computing architectures as concerns grow about the cost and energy requirements of advanced AI.

The human brain provides a striking comparison.

“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,” Wolfgang Maass, senior author of the study, told Tech Xplore.

“What do they look like? Can we reproduce them in artificial devices? The latter is of particular importance because the energy consumption of current AI tools has become a major economic, political and scientific impediment, whereas the brain needs only 20W for producing intelligence.”

Cognitive Maps Could Change How AI Organizes Knowledge

The researchers focused on cognitive maps, internal representations that cognitive scientists believe the brain uses to organize learned information.

Instead of storing knowledge simply as tables or large collections of parameters, the brain can use these representations to navigate relationships between experiences, possible actions and goals.

A Different Approach to Problem-Solving

“It is very tempting to use these hints for designing corresponding algorithms in artificial systems,” Maass explained. “We were especially interested in reconstructing algorithmic approaches of the brain for solving difficult problems, which require some form of intuition or fantasy. We wondered whether we could find ways to reproduce this in a simple chip, without using DNNs or LLMs.”

The team developed a framework that uses cognitive maps to organize information and guide decisions.

Initial testing indicated that a neural network based on the approach could adapt its planning and successfully address problems it had not encountered previously.

Researchers also found that its decision-making processes were comparatively easy to interpret, potentially addressing another major challenge facing advanced AI systems: understanding why a model produced a particular answer or action.

Brain-Inspired AI Could Run on Neuromorphic Chips

The framework could be particularly well suited to emerging forms of energy-efficient computing hardware.

“There already exists a very large number of studies that investigate brain algorithms at the micro level by considering computational properties of biological neurons,” Maass said. “Our results show that experimental data from the brain also provide novel insight at a higher level of computational organization and data structures.”

The researchers said the approach could be implemented on neuromorphic and in-memory computing chips.

Reducing Energy Used to Move Data

Traditional computer architectures frequently move information between memory and processors. That movement consumes energy and can become a bottleneck for intensive AI workloads.

In-memory computing attempts to address the problem by performing calculations directly where information is stored. Neuromorphic hardware, meanwhile, is designed to mimic certain characteristics of biological nervous systems.

“In contrast to current AI methods, these brain-like methods can easily be implemented in novel types of chips that aim to make AI implementations more energy efficient,” Maass said. “These are currently being developed under the names in-memory computing chips and neuromorphic chips by several large companies (IBM, Intel) and numerous startups.”

Another advantage is that the proposed model learns locally and does not depend on the same type of computationally intensive training commonly associated with large AI models.

Researchers Plan Hardware Tests in the United States

The research team is now moving toward practical hardware implementations.

Maass and his colleagues are working with engineers at Intel and another U.S. startup to implement their brain-inspired algorithms on chips.

“We are continuing our work on porting brain-like algorithms into more energy-efficient hardware,” Maass added. “In our forthcoming work, we will show that we can also capture brain-like low latency and our astounding flexibility of decision-making in the face of new goals or contingencies.”

The researchers are also studying whether their approach can produce explanations for decisions based on specific experiences.

“We are also able to capture that our brains usually automatically provide explanations for chosen actions in the form of concrete experiences that support a decision. Notably, making AI decisions explainable is a major goal in current AI research.”

A Potential Alternative Path for Artificial Intelligence

The research remains at an early stage, and the framework will need to be tested across a broader range of tasks and hardware platforms before its practical advantages can be fully established.

Still, the findings point toward an alternative direction for AI development. Rather than relying primarily on increasingly large models and computing clusters, future systems could borrow more directly from the brain’s methods for organizing knowledge, adapting to unfamiliar situations and making decisions efficiently.

If those capabilities translate successfully to neuromorphic and in-memory hardware, brain-inspired computing could eventually help reduce AI’s energy demands while producing systems that are faster, more adaptable and easier to understand.

William Faulkner

William Faulkner is a contributor to Columbuspost.com, covering a broad range of topics including news, politics, business, technology, sports, entertainment, and lifestyle. He focuses on delivering clear, accurate reporting and useful information that helps readers stay informed about current events and developments. His work highlights stories that matter to local and wider communities, with an emphasis on relevance, clarity, and providing readers with a balanced understanding of the issues shaping everyday life.