AI usually runs on computer chips. In this case, it runs on living brain cells.
Researchers in Japan have trained rat neurons to perform real-time machine learning tasks, moving computing into biological territory. The system uses cultured neurons connected to hardware to generate complex signals without traditional software models.
A recent study published in Proceedings of the National Academy of Sciences demonstrates that biological neural networks can act as adaptive computing systems and could eventually complement traditional computing approaches.
Living neurons perform real-time computation
Researchers at Tohoku University and Future University Hakodate trained cultured rat cortical neurons to generate structured signals using a machine learning framework known as reservoir computing.
According to Tom’s Hardware, the system recorded neural activity using a 26,400-electrode array and converted it into continuous outputs, which were then fed back to the neurons as electrical stimulation. This closed-loop process is updated roughly every 333 milliseconds.
“This work shows that living neuronal networks are not only biologically meaningful systems but may also serve as novel computational resources,” said Hideaki Yamamoto, professor at Tohoku University.
Using a method called FORCE learning, the system continuously adjusted outputs to match target signals. Researchers set a desired pattern, and the system refined its activity in real time to reduce errors. Over time, the neurons learned to reliably produce both regular and complex signals, showing that biological networks can be trained in ways similar to artificial neural networks.
Microfluidics activate usable biological networks
A key challenge in biological computing is that neurons tend to synchronize, limiting their ability to process complex information. The scientists addressed this using microfluidic devices to guide how neurons grow and connect.
Asia Research News stated that by structuring neurons into modular networks, the team reduced redundant firing and facilitated richer, high-dimensional activity needed for computation. The publication noted that this approach allowed the biological networks to generate complex time-series signals comparable to artificial systems.
Neuroscience News also reported that preventing synchronized firing was critical to enabling advanced computation in the system.
Two network designs were tested. In one setup, neurons were arranged in a grid with many consistent connections. In the other, connections were more spread out and uneven. The grid-like design performed better, likely because the neurons interacted more often and produced stronger signals.
Biological computing moves closer to reality
The system generated a range of signals, from smooth, repeating waves to more complex, chaotic patterns. Tom’s Hardware noted that one example is the Lorenz attractor, which is often used to model unpredictable systems like weather.
The same group of neurons could also be retrained to produce different signals, demonstrating flexibility similar to that of traditional AI systems.
Researchers said the platform could be used beyond computing, such as studying brain activity, testing drugs, or modeling neurological conditions. Neuroscience News emphasized that this development suggests that biological systems could support more energy-efficient computing or new forms of brain-machine interaction.
At the same time, the technology is still in its early stages. Questions remain about its reliability, scalability, and governance as computing begins to involve living cells.
Learn how AI “brains” for robots are expanding automation beyond office roles into construction, logistics, and energy.