Summary: Researchers have developed an energy-efficient artificial intelligence model inspired by hippocampal brain mechanisms. The system mimics the way the human brain constructs cognitive maps, performs stochastic calculations, and uses coding to solve complex programming tasks without complete calculations.
Published as a proof of concept for other AI architectures, the adaptive model responds to new environments without requiring retraining, using a fraction of the power required by multi-layered neural networks or large linguistic models. The technology provides an efficient framework for spatial distribution of autonomous robots, edge computing devices, and vehicles that operate under tight power constraints.
Important Facts
- Biological Energy Efficiency: The human brain runs complex programming and cognitive tasks on about 20 watts of power, which compares sharply with the megawatt-scale power requirements of modern large-scale language models and deep neural networks.
- Hippocampal Architectural Principles: The model translates three specific neural mechanisms into algorithmic code: geometric cognitive maps, stochastic state generation, and creative reuse of action components.
- Incomplete Search Strategy: Guided by cognitive maps, the algorithm evaluates randomly generated intermediate steps toward the goal instead of combining full computational solution paths, greatly reducing computational overhead.
- Zero-Shot Practice: The system adapts to environmental structural changes or novel tasks dynamically without requiring retraining or parameter tuning.
- Validation functions: Brain-inspired structures were successfully verified in all three test benchmarks: 2D spatial navigation, abstract multidimensional positioning, and block-block silhouette integration and disassembly.
Source: Graz University of Technology
The capabilities of large AI systems are constantly improving, but they consume a lot of energy during training and operation. The human brain, in contrast, is very energy efficient: it only needs about 20 watts.
Researchers at the Graz University of Technology, in collaboration with international partners, have developed a novel, brain-inspired AI model that can dynamically program and solve complex problems. In doing so, it consumes much less power than multi-layer neural networks or large language models.
“The brain works in a completely different way in today's AI systems,” says Wolfgang Maass of the Institute of Machine Learning and Neural Computation at Graz University of Technology. “We're trying to translate the way it works into algorithms and use it in AI applications.”
The three ways of the brain
Inspired by neuroscientific research on the hippocampus, Wolfgang Maass and his colleague Yukun Yang identified three mechanisms that the human brain uses in planning and solving problems:
- the creation of cognitive maps, i.e. the transformation of relationships between abstract objects into geometric relationships between neural codes in the brain, which provide a “sense of direction” as a spatial map;
- stochastic neural statistics, i.e. continuous generation of hypothetical scenarios and predictions;
- and assembly code, i.e. the division of information and programs into reusable components.
The translation of these methods into algorithms allows Graz's AI model – similar to humans or animals – to think and evaluate possible ways to solve complex problems without completely calculating until the solution.
If a randomly selected intermediate step points to the intended goal – here, the mind map directs the plan – this method is followed. In its new state, the system re-evaluates various options for action, thus gradually moving closer to the goal.
“In this way, our AI model can also dynamically respond to changed or new situations without needing to be retrained,” Yukun Yang said.
Assembling and disassembling the silhouette
The researchers successfully tested the capabilities of their brain-inspired AI using three tasks: navigating a two-dimensional environment, guiding itself in an invisible, multi-dimensional environment, and assembling and disassembling a silhouette made up of various building blocks.
The researchers emphasize that their approach is not intended to compete with today's major language models, but provides the basis for a different approach for specific applications. “We are in the early stages of development,” says Wolfgang Maass. “But our work shows that powerful AI doesn't need huge data centers and massive power.”
Suitable for robots and edge devices
In time, such brain-inspired systems could be used in robots, autonomous vehicles or other edge devices – in other words, wherever AI needs to operate in an environment with a limited power supply.
Wolfgang Maass works as a key researcher in the Bilateral AI Cluster of Excellence. The present study was carried out in collaboration with Tsinghua University and the National Research Council of Italy.
Important Questions Answered:
A: Conventional AI systems tend to enumerate all possible action sequences to completion in order to find the optimal solution, which requires a lot of memory and processing power. Mind maps transform abstract relationships into geometric coordinates, allowing the model to quickly check that a randomly generated step points to a goal, discarding unpromising paths before using the full computational power.
A: Stochastic computing introduces controlled randomness to generate hypothetical situations and predictions on the fly. Rather than using a deterministic, dynamic search, the system samples possible paths, mimicking how biological neurons use noise and variation to quickly sample possible solutions.
A: Although not designed to replace large language models in centralized data centers, this light-footprint architecture has been optimized for edge devices, autonomous vehicles, and mobile robots. These real-world applications require real-time adaptation, complex routing, and low latency under tight battery and thermal constraints.
Editor's Notes:
- This article was edited by a Neuroscience News editor.
- The journal paper is fully revised.
- More content is added by our staff.
About this AI and cognitive research issues
Author: Philip Jarke
Source: Graz University of Technology
Contact person: Philipp Jarke – Graz University of Technology
Image: Image posted in Neuroscience News
Actual research: Open access.
“Neural sampling from cognitive maps enables goal-directed thinking and planning” by Hui Lin, Yukun Yang, Rong Zhao, Giovanni Pezzulo and Wolfgang Maass. Nature Machine Intelligence
DOI:10.1038/s42256-026-01254-4
Abstract
Neural sampling from cognitive maps enables goal-directed thought and planning
Artificial intelligence systems are becoming more intelligent, but at a much higher cost in terms of energy consumption and training requirements. In contrast, our brains only need 20 W of power, they learn online and can quickly adapt to emergencies.
This begs the question of what data structures, algorithms and learning methods allow the brain to achieve that, and whether these can be ported to artificial devices. We approach this question with a core aspect of intelligence: the ability to plan and solve problems, including new problems involving states that have never encountered them before.
Here we explore three tools the brain might use to achieve that: mind maps, stochastic computing and convolutional coding. We integrate these tools into a transparent neural network model, and demonstrate its power for adaptive programming and problem solving.
Importantly, this approach is suitable for implementation using in-memory computing and other energy-efficient neuromorphic hardware. In particular, it only requires local self-monitoring synaptic plasticity ready for on-chip learning.
Therefore, the main feature of the brain's intelligence—the ability to generate solutions to problems it has never encountered before—does not require deep neural networks or large language models, and can be implemented in energy-efficient edge devices.
