AI networks build brain-like modules when juggling multiple tasks

Researchers at the Brain Intelligence Lab published a study in Nature Machine Intelligence showing how artificial neural networks organize themselves. They used recurrent neural networks, which are AI models that process information step-by-step, to simulate learning. The goal was to see if the way these networks connect mirrors how the human brain works.

The team trained the AI on different sets of cognitive tasks. Some networks learned one task at a time, while others had to handle multiple tasks simultaneously. The results showed a clear difference in how the internal connections formed. Networks that juggled multiple tasks developed distinct, separate groups of connections.

These groups are called modules. In the brain, modules are clusters of neurons that work together on specific jobs, like memory or attention. The study found that multitasking AI created these modules more strongly than single-task AI. This suggests that the pressure of learning complex things drives the structure, not just the physical layout.

Imagine a kitchen where one chef handles everything. Now imagine a team where each person specializes in a station. The study suggests that when an AI is pushed to its limits, it organizes itself like the specialized team. This organization helps it process information more efficiently. It also makes the AI’s internal map look more like a real human brain.

The researchers compared these AI networks to data from the Human Connectome Project. This is a large database of human brain scans. The AI networks trained on multiple tasks matched the human brain’s structure more closely than models that only considered physical wiring costs. This is a significant step in understanding why brains are built the way they are.

The study highlights that functional demands shape the network. In other words, what the AI needs to do determines how it is built. This challenges older ideas that brain structure is purely about saving energy or space. It points to a more active process where learning creates the architecture.

This finding helps explain how brains might develop flexibility. By forming modules, the system can adapt to new challenges without breaking old skills. It offers a computational model for how intelligence might emerge from simple rules. The work bridges the gap between computer science and neuroscience.

The researchers used Python-based simulations to generate the data. They made their code and data publicly available on GitHub and Zenodo. This allows other scientists to check their work and build on it. Transparency is key in this field, and this study provides a solid foundation for future research.

The study focuses on recurrent networks, which are different from the large language models people use daily. It does not claim that current chatbots work this way. Instead, it uses these specific AI models as a tool to understand biological brains. The insights are theoretical and structural, not about improving chatbot performance.

What does this mean for the future of AI? It suggests that to build smarter systems, we might need to look at how they organize their internal connections. The brain’s design is not an accident; it is a solution to complex problems. Understanding this could lead to more efficient and adaptable AI in the long run.

The limits of this study are clear. It is a computational simulation, not a direct observation of a human brain. The AI models used are simplified versions of real brains. While the structural match is strong, it does not prove that AI thinks like humans. It shows that similar pressures can lead to similar shapes.