Researchers at Tohoku University and Future University Hakodate in Japan have successfully trained living rat brain cells to autonomously perform machine learning computations, generating complex waveforms without external input. The team integrated cultured rat cortical neurons with high-density microelectrode arrays and microfluidic devices, creating a closed-loop system that learned to produce periodic and chaotic signals . This foundational research demonstrates that biological neural networks can serve as computational substrates, though the work remains in early stages with significant practical constraints.

How Did Researchers Get Rat Neurons to Learn Machine Learning Tasks?

The experiment relied on a carefully designed architectural approach. The researchers confined neuronal cell bodies to 128 square wells, each roughly 100 by 100 micrometers, with an average of 14.6 neurons per well . These wells were linked by microchannels in two configurations: a lattice design with uniform nearest-neighbor connections, and a hierarchical design with sparser, multi-scale connections. Without these physical constraints, cultured neurons form dense, highly synchronized networks that fire in lockstep, preventing them from learning any target signals.

The system recorded spike trains from the neurons across a 26,400-electrode array with a 17.5-micrometer pitch, filtered them into continuous signals, and decoded an output through a linear readout layer . That output was then fed back to the neurons as electrical stimulation, completing a feedback loop that cycled roughly every 333 milliseconds. The readout weights were optimized in real time using an algorithm called FORCE (First-Order Reduced and Controlled Error) learning, which continuously adjusted the decoder to minimize the error between the network's output and a target waveform.

What Computational Tasks Could These Biological Networks Perform?

The results demonstrated genuine learning capability. Using the lattice and hierarchical networks, the system learned to generate sine waves with periods of 4, 10, and 30 seconds, as well as triangle and square waves . The same culture preparation could be retrained to oscillate at different frequencies. The researchers also demonstrated that the system could approximate a Lorenz attractor, a three-dimensional chaotic trajectory, with pairwise correlations above 0.8 between predicted and target signals across all dimensions during the learning phase.

The patterned configurations dramatically reduced pairwise neural correlations compared to unpatterned cultures, increasing the dimensionality of the network's dynamics . Lattice networks consistently outperformed hierarchical ones across all target waveforms, likely because their denser intermodular connections produced higher firing rates that gave the linear decoder more signal to work with.

How to Understand the Key Limitations of Biological Computing

• Latency Constraints: The feedback loop's roughly 330-millisecond latency limited the system's ability to track fast-changing or sharp-edged waveforms. Reducing this delay through specialized hardware or alternative filtering could expand the range of learnable targets, but this remains a significant engineering challenge .
• Autonomous Performance Degradation: Performance degraded after training was halted and the system ran autonomously, with mean squared error increasing in 99% of trials. This suggests biological systems require continuous maintenance and retraining, unlike static silicon chips .
• Scalability Questions: While this proof-of-concept used rat neurons in a controlled laboratory setting, the researchers did not demonstrate scaling to the billions of neurons required for large language models. The current system represents a foundational research step, not a practical replacement for existing AI infrastructure.

"This work shows that living neuronal networks are not only biologically meaningful systems but may also serve as novel computational resources," said Hideaki Yamamoto.

Hideaki Yamamoto, Professor at Tohoku University's Research Institute of Electrical Communication

The enabling technology was the use of PDMS (polydimethylsiloxane) microfluidic films to constrain how neurons connected. This architectural innovation proved critical; without it, the neurons simply could not learn the target signals. The lattice configuration proved superior, suggesting that network topology matters as much in biological systems as it does in artificial neural networks.

The study, published March 12 in the journal Proceedings of the National Academy of Sciences, represents a fundamental exploration of computation itself . Rather than viewing AI as exclusively a silicon problem, this work suggests that biological systems possess untapped computational potential. The researchers noted that future applications could potentially extend to brain-machine interfaces and neuroprosthetic devices, though these remain speculative at this stage .

This research opens a new line of inquiry into alternative computational substrates. While significant hurdles remain before biological computing could scale to practical applications, the work demonstrates that living neurons can be trained to perform machine learning tasks autonomously. The findings may inspire further investigation into hybrid biological-silicon systems or entirely novel approaches to computation, though the practical viability of such systems remains an open question.